<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:media="http://search.yahoo.com/mrss/">
<channel>
<title>inthesingularity.ai — AI &amp; Finance Analysis</title>
<link>https://inthesingularity.ai/</link>
<atom:link href="https://inthesingularity.ai/feed.xml" rel="self" type="application/rss+xml" />
<description>Analysis of markets, AI infrastructure and the capital behind it.</description>
<language>en</language>
<lastBuildDate>Wed, 18 Nov 2026 12:32:00 GMT</lastBuildDate>
<item>
  <title>What is Anthropic worth across fourteen worlds?</title>
  <link>https://inthesingularity.ai/packs/anthropic-worlds</link>
  <guid isPermaLink="true">https://inthesingularity.ai/packs/anthropic-worlds</guid>
  <pubDate>Wed, 18 Nov 2026 12:32:00 GMT</pubDate>
  <description>Anthropic priced its IPO at $83 a share on November 18, about $1.5T fully diluted, and trades as ANT from the 19th. We value it on three paths for 2030 revenue ($230B, $450B and $750B), then on eleven worlds a growth rate cannot describe, among them a frontier halt, a licensed frontier, a labor-budget unlock and super-acceleration. Weighted by our odds, a share is worth $117 today at a 12% required return. The middle world is worth $78. The strange worlds leave the average about where three cases put it and empty the middle: 40% of simulated worlds end 2030 below the IPO price and 40% at least double it. The odds are our judgments.</description>
  <category>AI labs</category>
  <enclosure url="https://inthesingularity.ai/newsletter/anthropic-worlds/cover.png" length="54642" type="image/png" />
  <media:content url="https://inthesingularity.ai/newsletter/anthropic-worlds/cover.png" medium="image" type="image/png" />
  <content:encoded><![CDATA[<figure><a href="https://06.ai/packs/anthropic-worlds"><img src="https://06.ai/newsletter/anthropic-worlds/thumb.png" alt="What is Anthropic worth across fourteen worlds?" width="421" height="415"></a><figcaption><a href="https://06.ai/packs/anthropic-worlds">Interactive version and data on 06.ai →</a></figcaption></figure>
<ul><li><strong>$117</strong> — Probability-weighted value of a share today against the <strong>$83</strong> IPO price, at a 12% required return</li><li><strong>40%</strong> — Worlds that end 2030 below the IPO price a quarter end below <strong>half</strong> of it</li><li><strong>12%</strong> — Odds of super-acceleration each ten points adds about <strong>$10</strong> a share</li><li><strong>62%</strong> — Value held by the five up-worlds on <strong>27.5%</strong> of the odds</li></ul>
<h2>At $83, Anthropic sells for 10 to 11 times next year’s revenue</h2>
<p>The offering sold 980 million shares for $81.3B, the largest IPO on record, with a further $26B placed privately with suppliers and partners at the same price. Counting options, restricted stock and the equity plan reserve, about 18.1B shares, the company is valued near $1.5T. This pack uses that count throughout, so stock promised or reserved for employees is inside the price.</p>
<h2>The price against revenue</h2>
<p>About $1.5T, fully diluted at the IPO price, divided by each measure of revenue. OpenAI’s row uses its own valuation.</p>
<figure><a href="https://06.ai/packs/anthropic-worlds#price"><img src="https://06.ai/newsletter/anthropic-worlds/fig-01.png" alt="Run-rate at October 31, 2026, from the amended S-1 — Revenue: $81B · Multiple: 18.5×; 2026 revenue on our tracker’s path — Revenue: about $54B · Multiple: 28×; 2027 revenue on our tracker’s path — Revenue: about $147B · Multiple: 10×; 2027 revenue, Goldman Sachs — Revenue: $136B · Multiple: 11×; For comparison, OpenAI’s November round: $1.25T on a September run-rate near $50B — Revenue: $50B · Multiple: 25×" width="694" height="292"></a></figure>
<h2>Estimates of 2030 revenue differ by a factor of two</h2>
<p>Anthropic has not published a 2030 figure. Its IPO plan runs to 2028. Pink is this pack’s base case.</p>
<ul><li><strong>Anthropic’s plan</strong> — 2030 revenue: none · Basis: $190B to $200B for 2028 (Reuters, August); long-term adjusted operating margin above 35%, from 15% in the third quarter</li><li><strong>Goldman Sachs, a lead underwriter</strong> — 2030 revenue: $685B · Basis: $52B in 2026, $136B in 2027, $442B in 2029 (Bloomberg, November 10)</li><li><strong>Artemis</strong> — 2030 revenue: $800B · Basis: Run-rate at the end of 2030, cut from $1T after the S-1</li><li><strong>Decoding Discontinuity</strong> — 2030 revenue: $401B · Basis: Base case of a long-term cash-flow model</li><li><strong>Morningstar</strong> — 2030 revenue: not stated · Basis: 47% a year for five years; fair value $92, bull $188, bear $25</li><li><strong>06.ai tracker’s path, extended</strong> — 2030 revenue: about $480B · Basis: A run-rate of $86B on November 18, growing 8.2% a month, with growth halving every 18 months</li><li><strong>This pack, base case</strong> — 2030 revenue: $450B · Basis: Between Decoding Discontinuity’s base case and our tracker’s path, and well under Goldman’s figure</li></ul>
<h2>On 2030 revenue alone, a share is worth $21, $103 or $216</h2>
<p>The value of a case is 2030 revenue times an operating margin times a multiple of operating profit, divided by the share count and discounted to today at 12% a year. The three cases differ on one question: how long a frontier model keeps its price.</p>
<h2>Three cases for 2030</h2>
<p>Each column is one path for revenue and what the business is worth on it. Worth today discounts the 2030 share price at 12% a year for 4.12 years.</p>
<figure><a href="https://06.ai/packs/anthropic-worlds#cases"><img src="https://06.ai/newsletter/anthropic-worlds/fig-02.png" alt="What has to be true — Bear: the tenth-price world: Every flagship is matched at a tenth of its price within two months · Base: the lead holds: The lead holds at one to two quarters and distillers stay locked out · Bull: the lead widens: Rivals need a year to match a flagship and distillation stops working; 2030 revenue — Bear: the tenth-price world: $230B · Base: the lead holds: $450B · Bull: the lead widens: $750B; Operating margin — Bear: the tenth-price world: 15% · Base: the lead holds: 30% · Bull: the lead widens: 36%; Value ÷ operating profit — Bear: the tenth-price world: 20× · Base: the lead holds: 23.3× · Bull: the lead widens: 25×; Value ÷ revenue — Bear: the tenth-price world: 3.0× · Base: the lead holds: 7.0× · Bull: the lead widens: 9.0×; Equity value in 2030 — Bear: the tenth-price world: $0.69T · Base: the lead holds: $3.15T · Bull: the lead widens: $6.75T; Share price in 2030 — Bear: the tenth-price world: $34 · Base: the lead holds: $164 · Bull: the lead widens: $345; Worth today at 12% a year — Bear: the tenth-price world: $21 · Base: the lead holds: $103 · Bull: the lead widens: $216; Return a year from $83 — Bear: the tenth-price world: −19% · Base: the lead holds: +18% · Bull: the lead widens: +41%" width="694" height="579"></a></figure>
<p><strong>Against the price holding.</strong> Fable 5.1 was matched at half its price in 21 days and at a tenth in 55 (<a href="https://06.ai/packs/fable-5-2">the </a><a href="https://06.ai/packs/fable-5-2">Fable 5.2</a><a href="https://06.ai/packs/fable-5-2"> pack</a>). OpenAI’s GPT-6.2 Astra, released eight days after Fable 5.2, matched it on Epoch’s index at the same list price.</p>
<p><strong>For it.</strong> Anthropic says its defenses cut what distillers harvest by 90% to 99%, and estimates that customers later found breaking its rules made up 5% to 7% of nine-month revenue. On a newer set of research tasks the best open-weight model scored 34%, against 59% and 63% for two frontier models (<a href="https://06.ai/packs/fable-5-2">the </a><a href="https://06.ai/packs/fable-5-2">Fable 5.2</a><a href="https://06.ai/packs/fable-5-2"> pack</a>).</p>
<p><strong>What is at stake.</strong> Over 80% of third-quarter revenue came from models released in the prior six months (<a href="https://06.ai/packs/anthropic-s1">the </a><a href="https://06.ai/packs/anthropic-s1">Anthropic S-1</a><a href="https://06.ai/packs/anthropic-s1"> pack</a>). Anthropic re-earns its revenue every two quarters.</p>
<p><strong>Compute sets the scale.</strong> Anthropic ran about 3.1 GW in the third quarter and earned about $22B a year per gigawatt. About 25 GW is lined up for the end of 2029, which puts the average for 2030 near 28 GW if the build continues (<a href="https://06.ai/packs/anthropic-s1">the </a><a href="https://06.ai/packs/anthropic-s1">Anthropic S-1</a><a href="https://06.ai/packs/anthropic-s1"> pack</a>; our estimates). A newer report puts this year’s compute lower, and the <a href="https://06.ai/packs/anthropic-worlds#method">Method</a> shows what that changes. The base case needs 20 GW at today’s revenue per gigawatt. The bull case needs all 28 GW and a fifth more revenue from each. The bear case earns $8B per gigawatt while $451.6B of compute commitments fall due before the end of 2030.</p>
<h2>Anthropic’s flagships now arrive 35 days apart, down from 82</h2>
<p>Between now and 2030 Anthropic will replace its flagship many times, and each replacement resets its price. The releases are coming faster.</p>
<h2>Days between flagship releases</h2>
<figure><a href="https://06.ai/packs/anthropic-worlds#pace"><img src="https://06.ai/newsletter/anthropic-worlds/fig-03.png" alt="Days between flagship releases" width="694" height="331"></a><figcaption>Days since the previous generally available Opus- or Fable-class release. Grey is the four releases to April 2026 and pink the six since; the dashed lines are each group’s average. Dates from the <a href="https://06.ai/packs/fable-5-2">Fable 5.2</a> pack.</figcaption></figure>
<h2>Five gauges of the pace</h2>
<p>The first four are from the <a href="https://06.ai/packs/fable-5-2">Fable 5.2</a> pack; the last is Anthropic’s own count.</p>
<figure><a href="https://06.ai/packs/anthropic-worlds#gauges"><img src="https://06.ai/newsletter/anthropic-worlds/fig-04.png" alt="New best models from Anthropic in twelve months — Before: 4 (to Nov 2025) · November 2026: 9; Median gap between them — Before: 81 days · November 2026: 45 days; Frontier gain on Epoch’s index in twelve months — Before: 15.8 points · November 2026: 23.1 points; Remote Labor Index, best Anthropic model — Before: 3.75% (Nov 2025) · November 2026: 41.3%; Anthropic’s merged code written by Claude — Before: over 80% (May 2026) · November 2026: over 90%" width="694" height="271"></a></figure>
<p>Epoch has said five index points roughly equal one doubling of the length of task a model can finish. On that reading the frontier went from about three doublings a year to about four and a half.</p>
<p><strong>The loop.</strong> Anthropic reported in June that its engineers ship about eight times the code they did in 2024, and that success on its hardest internal tasks rose from 26% to 76% in six months. Its November 10 filing says it now meets its own threshold for AI-driven speed-ups in research, a judgment that rests on the models it expects and its rate of progress, not on what current models have shown. A co-founder put the chance of self-improving AI by the end of 2028 at 60% in May and wrote in October that 2027 now looks likelier than 2028. OpenAI’s target date for automating its researchers is March 2028 (TIME).</p>
<p><strong>The brakes.</strong> Fable 5.2 shipped with the skills that most directly speed up AI development, accelerator kernel optimization and model-architecture work, held at about Opus 5.5’s level by design. The chief executive asked the industry in September to slow capability gains. The S-1/A says the company does not expect to let Claude build its successors without human supervision. Outside accounts still point to direction-setting and evaluation as the gap.</p>
<p><strong>Why a faster clock is not simply the bull case.</strong> A faster clock raises revenue only while each release holds its price, and the same acceleration shortens the time to a cheaper match. It also raises the odds that Washington steps in. Super-acceleration therefore gets its own odds and its own four worlds.</p>
<h2>Super-acceleration gets 12%, and it is four worlds</h2>
<p>Super-acceleration here means three things happen together. AI systems do most of the research at a frontier lab by the end of 2028. Measured progress then runs at least three times the 2026 pace for a year. Nothing binding stops it before 2030.</p>
<h2>Three judgments multiply to 12%</h2>
<figure><a href="https://06.ai/packs/anthropic-worlds#chain"><img src="https://06.ai/newsletter/anthropic-worlds/fig-05.png" alt="Three judgments multiply to 12%" width="694" height="365"></a><figcaption>Each bar is the chance that the step and all the steps above it hold. Grey is the first step, strong grey the first two, pink all three. The second chart splits the pink bar by who captures the value. All figures are our judgments.</figcaption></figure>
<p><strong>40% that the loop closes.</strong> Below the co-founder’s 60% in May, which he has since pulled forward, and below OpenAI’s own target date. Above the skeptics who say direction-setting is not automated.</p>
<p><strong>50% that the pace then triples.</strong> Automating researchers does not triple progress if experiments wait on compute, and research already takes about three-quarters of Anthropic’s compute (<a href="https://06.ai/packs/anthropic-s1">the </a><a href="https://06.ai/packs/anthropic-s1">Anthropic S-1</a><a href="https://06.ai/packs/anthropic-s1"> pack</a>).</p>
<p><strong>60% that it is allowed to run.</strong> The labs have asked to be paced, Washington has stopped a frontier model once, and Congress changes hands in January. Race pressure pulls the other way.</p>
<p>With an insider’s inputs (60%, 70%, 70%) the product is 29%. With a skeptic’s (20%, 40%, 50%) it is 4%. At 29% a share is worth $135.</p>
<p>The 12% then divides. In 3.5% Anthropic is among the leaders and sells to anyone. In 4% the same capability goes only to approved US buyers. In 2% Washington takes most of the gain through a golden share, a levy or the compute itself. In 2.5% the loop closes first at a rival. The family is worth $207 a share on average against $105 for everything else. That is only twice as much, because half the family is gated, taxed or lost.</p>
<h2>Fourteen worlds: the three ordinary cases get 45% of the odds</h2>
<p>Each world sets 2030 revenue, a multiple of revenue and the share of the value that holders keep. The eleven strange worlds are grouped by what breaks: demand steps down, the rules change, demand steps up, or the pace of research itself changes.</p>
<h2>The fourteen worlds</h2>
<p>Odds are our judgments. Worth today is the 2030 share price at each world’s central figures, discounted at 12% a year for 4.12 years. Share of value and the last row come from the simulation, which spreads revenue and the multiple around those figures. Holders keep 90%, 80% and 35% of the value in the first three super-acceleration worlds and all of it elsewhere. Revenue per GW divides 2030 revenue by 28 GW, against about $22B today; the famine world has about 13 GW.</p>
<ul><li>The three ordinary cases · 45%</li><li><strong>Normal: the lead holds</strong> — Odds: 23% · 2030 revenue: $450B · Value ÷ revenue: 7.0× · Worth today: $103 · Against $83: +24% · Share of value: 22.0% · Revenue per GW: $16B</li><li><strong>The lead widens and hardens</strong> — Odds: 11% · 2030 revenue: $750B · Value ÷ revenue: 9.0× · Worth today: $216 · Against $83: +160% · Share of value: 22.1% · Revenue per GW: $27B</li><li><strong>Tenth-price world</strong> — Odds: 11% · 2030 revenue: $230B · Value ÷ revenue: 3.0× · Worth today: $21 · Against $83: −74% · Share of value: 2.2% · Revenue per GW: $8B</li><li>Step-downs · 23%</li><li><strong>Frontier halt</strong> — Odds: 12% · 2030 revenue: $240B · Value ÷ revenue: 3.5× · Worth today: $26 · Against $83: −68% · Share of value: 3.1% · Revenue per GW: $9B</li><li><strong>Lost lead</strong> — Odds: 6% · 2030 revenue: $200B · Value ÷ revenue: 3.0× · Worth today: $19 · Against $83: −78% · Share of value: 1.1% · Revenue per GW: $7B</li><li><strong>Capex bust</strong> — Odds: 5% · 2030 revenue: $175B · Value ÷ revenue: 2.0× · Worth today: $9 · Against $83: −89% · Share of value: 0.4% · Revenue per GW: $6B</li><li>Regime changes · 11%</li><li><strong>Licensed frontier</strong> — Odds: 8% · 2030 revenue: $330B · Value ÷ revenue: 7.0× · Worth today: $77 · Against $83: −7% · Share of value: 5.9% · Revenue per GW: $12B</li><li><strong>Compute famine</strong> — Odds: 3% · 2030 revenue: $300B · Value ÷ revenue: 6.0× · Worth today: $59 · Against $83: −28% · Share of value: 1.7% · Revenue per GW: $23B</li><li>Step-ups · 9%</li><li><strong>Labor-budget unlock</strong> — Odds: 6% · 2030 revenue: $1.0T · Value ÷ revenue: 9.0× · Worth today: $282 · Against $83: +240% · Share of value: 16.5% · Revenue per GW: $36B</li><li><strong>Cyber mobilization</strong> — Odds: 3% · 2030 revenue: $600B · Value ÷ revenue: 7.0× · Worth today: $136 · Against $83: +64% · Share of value: 3.9% · Revenue per GW: $21B</li><li>Super-acceleration · 12%</li><li><strong>Super-acceleration, open</strong> — Odds: 3.5% · 2030 revenue: $1.6T · Value ÷ revenue: 8.0× · Worth today: $361 · Against $83: +335% · Share of value: 12.4% · Revenue per GW: $57B</li><li><strong>Super-acceleration, licensed</strong> — Odds: 4% · 2030 revenue: $900B · Value ÷ revenue: 7.5× · Worth today: $176 · Against $83: +112% · Share of value: 7.0% · Revenue per GW: $32B</li><li><strong>Super-acceleration, state capture</strong> — Odds: 2% · 2030 revenue: $1.1T · Value ÷ revenue: 6.0× · Worth today: $78 · Against $83: −5% · Share of value: 1.5% · Revenue per GW: $39B</li><li><strong>Super-acceleration, a rival wins</strong> — Odds: 2.5% · 2030 revenue: $150B · Value ÷ revenue: 2.0× · Worth today: $8 · Against $83: −90% · Share of value: 0.2% · Revenue per GW: $5B</li><li><strong>All fourteen, simulated — Odds: 100% · 2030 revenue: $526B · Worth today: $117 · Against $83: +41% · Share of value: 100%</strong></li></ul>
<h2>Revenue in 2030 runs from $150B to $1.6T</h2>
<figure><a href="https://06.ai/packs/anthropic-worlds#paths"><img src="https://06.ai/newsletter/anthropic-worlds/fig-06.png" alt="Revenue in 2030 runs from $150B to $1.6T" width="694" height="727"></a><figcaption>Calendar-year revenue in $ billions on a log scale, one panel for each family of worlds. 2026 is our estimate of $54B. Later years are the path we assume inside each world, with the step placed in 2028 or 2029. They are assumptions for the simulation, not forecasts. Black is the three ordinary cases, blue the step-downs, gold the regime changes, green the step-ups and pink super-acceleration; the grey line repeats the normal world.</figcaption></figure>
<h2>What changes demand in each world</h2>
<p>For each world: what happens to demand, what the figures are anchored on, and the first sign that it is the one we are in.</p>
<h3>The three ordinary cases · 45%</h3>
<p><strong>Normal: the lead holds.</strong> 23% odds · $450B in 2030 · $103 a share today</p>
<p><strong>Demand.</strong> Enterprise use deepens at today’s pace and each flagship holds its price for a quarter or two.</p>
<p><strong>Anchor.</strong> Over 11,900 customers pay $100K or more a year and about 250 pay $10M or more (<a href="https://06.ai/packs/anthropic-s1">the </a><a href="https://06.ai/packs/anthropic-s1">Anthropic S-1</a><a href="https://06.ai/packs/anthropic-s1"> pack</a>). $450B sits between Decoding Discontinuity’s $401B and our tracker’s path, and well under Goldman’s $685B.</p>
<p><strong>First sign.</strong> Fable 5.2 keeps 60% of its best week’s share of OpenRouter spend in week six, the December 21 test in <a href="https://06.ai/packs/fable-5-2">the </a><a href="https://06.ai/packs/fable-5-2">Fable 5.2</a><a href="https://06.ai/packs/fable-5-2"> pack</a>.</p>
<p><strong>The lead widens and hardens.</strong> 11% odds · $750B in 2030 · $216 a share today</p>
<p><strong>Demand.</strong> The premium tier stops being matched. Buyers pay for the frontier because nothing cheaper finishes the long tasks.</p>
<p><strong>Anchor.</strong> On a newer set of research tasks the best open-weight model scored 34%, against 59% and 63% for two frontier models (<a href="https://06.ai/packs/fable-5-2">the </a><a href="https://06.ai/packs/fable-5-2">Fable 5.2</a><a href="https://06.ai/packs/fable-5-2"> pack</a>). Anthropic says its defenses cut what distillers harvest by 90% to 99%.</p>
<p><strong>First sign.</strong> A flagship goes 90 days without a half-price match. The last three took 42, 45 and 21 days.</p>
<p><strong>Tenth-price world.</strong> 11% odds · $230B in 2030 · $21 a share today</p>
<p><strong>Demand.</strong> Tokens keep growing. Dollars per task fall faster, and buyers route to the cheap tier.</p>
<p><strong>Anchor.</strong> Fable 5.1 was matched at a tenth of its price in 55 days, and seven weeks after Opus 5.5 its best coding score cost 48% less a task (<a href="https://06.ai/packs/fable-5-2">the </a><a href="https://06.ai/packs/fable-5-2">Fable 5.2</a><a href="https://06.ai/packs/fable-5-2"> pack</a>).</p>
<p><strong>First sign.</strong> The cheaper tier keeps the volume. Opus 5.5 took 33% of OpenRouter spend in its best week against 9% for Fable 5.1.</p>
<h3>Step-downs · 23%</h3>
<p><strong>Frontier halt.</strong> 12% odds · $240B in 2030 · $26 a share today</p>
<p><strong>Demand.</strong> A binding limit on training or releasing more capable models, across US labs, for six months or more. The upgrade cycle stops. Existing models keep selling while open-weight followers close the gap.</p>
<p><strong>Anchor.</strong> The frontier stopped twice in 2026, for 18 and 11 days. <a href="https://06.ai/packs/cyber-slowdown">The </a><a href="https://06.ai/packs/cyber-slowdown">cyberattack</a><a href="https://06.ai/packs/cyber-slowdown"> pack</a> puts a cyber-forced slowdown at 11% by the end of 2028 and 19% by the end of 2030.</p>
<p><strong>First sign.</strong> A pause bill moves in the new Congress, or a $10B attack is pinned on a frontier model.</p>
<p><strong>Lost lead.</strong> 6% odds · $200B in 2030 · $19 a share today</p>
<p><strong>Demand.</strong> A rival holds the frontier for two cycles or more and routers move the spend.</p>
<p><strong>Anchor.</strong> Anthropic’s share of OpenRouter spend ran between 50% and 61% a month from January to July, fell to 28% in September and was 39% in October (06.ai tracker data). GPT-6.2 Astra matched Fable 5.2 on Epoch’s index eight days after Fable 5.2 launched.</p>
<p><strong>First sign.</strong> Share of OpenRouter spend under 25% for a quarter.</p>
<p><strong>Capex bust.</strong> 5% odds · $175B in 2030 · $9 a share today</p>
<p><strong>Demand.</strong> Funding closes before the compute is paid for, and customers cut budgets in the same downturn.</p>
<p><strong>Anchor.</strong> $901B of compute commitments, $451.6B of it due before the end of 2030 and $186.2B with termination rights or conditions. AI hardware P/E ratios fell a median 47% after October 6 (<a href="https://06.ai/packs/unit-growth">the </a><a href="https://06.ai/packs/unit-growth">unit-growth</a><a href="https://06.ai/packs/unit-growth"> pack</a>).</p>
<p><strong>First sign.</strong> Suppliers’ receivables pass 60 days (<a href="https://06.ai/packs/ai-receivables">the </a><a href="https://06.ai/packs/ai-receivables">receivables</a><a href="https://06.ai/packs/ai-receivables"> pack</a>), or a commitment is renegotiated.</p>
<h3>Regime changes · 11%</h3>
<p><strong>Licensed frontier.</strong> 8% odds · $330B in 2030 · $77 a share today</p>
<p><strong>Demand.</strong> Frontier models are sold only to vetted US companies and cleared individuals. The customer list shrinks and the price per customer rises.</p>
<p><strong>Anchor.</strong> In June, access was barred to foreign nationals and came back first to about 100 approved US organizations. 28% of first-half revenue was billed outside the US. Mythos-class models already sit behind verification programs.</p>
<p><strong>First sign.</strong> A federal power to review releases passes, or verification extends from the Mythos tier to the Fable tier.</p>
<p><strong>Compute famine.</strong> 3% odds · $300B in 2030 · $59 a share today</p>
<p><strong>Demand.</strong> A Taiwan or power shock stops new capacity near 13 GW. Whoever holds installed compute rations it by price.</p>
<p><strong>Anchor.</strong> Anthropic buys short-term compute when demand runs ahead of its contracts, and it paid $5B for it in the third quarter alone.</p>
<p><strong>First sign.</strong> Spot GPU prices above contract prices for a quarter.</p>
<h3>Step-ups · 9%</h3>
<p><strong>Labor-budget unlock.</strong> 6% odds · $1.0T in 2030 · $282 a share today</p>
<p><strong>Demand.</strong> Agents are bought as headcount, out of labor budgets instead of IT budgets. It needs no research explosion.</p>
<p><strong>Anchor.</strong> The Remote Labor Index went from 3.75% to 41.3% in a year. On enterprise plans, the top 5% of Claude Code users spend about 20 times what a non-user does, which is under $75 a month. The S-1 sizes knowledge work above $40T a year.</p>
<p><strong>First sign.</strong> Contracts priced per task, or a customer above $1B a year.</p>
<p><strong>Cyber mobilization.</strong> 3% odds · $600B in 2030 · $136 a share today</p>
<p><strong>Demand.</strong> After a large AI-enabled attack, regulators and insurers require frontier-grade defense. Every large codebase is re-audited on a deadline, as with Y2K.</p>
<p><strong>Anchor.</strong> Mythos-class models already go to verified security teams. The $6.36B Binance theft was this year’s largest cyber loss, and three leading security stocks were up 114% to 157% for the year at October 28 (<a href="https://06.ai/packs/cyber-slowdown">the </a><a href="https://06.ai/packs/cyber-slowdown">cyberattack</a><a href="https://06.ai/packs/cyber-slowdown"> pack</a>).</p>
<p><strong>First sign.</strong> A regulator or the insurers mandate agentic code audits.</p>
<h3>Super-acceleration · 12%</h3>
<p><strong>Super-acceleration, open.</strong> 3.5% odds · $1.6T in 2030 · $361 a share today</p>
<p><strong>Demand.</strong> Systems that do a remote worker’s job are sold to anyone. Compute is the limit, so the step shows up in price: revenue per gigawatt rises from $22B to about $57B. Holders keep 90%.</p>
<p><strong>Anchor.</strong> Serving takes about a quarter of Anthropic’s compute today (<a href="https://06.ai/packs/anthropic-s1">the </a><a href="https://06.ai/packs/anthropic-s1">Anthropic S-1</a><a href="https://06.ai/packs/anthropic-s1"> pack</a>). Moving half the fleet to serving would nearly double revenue per gigawatt before any price rise.</p>
<p><strong>First sign.</strong> A public model sets its own research direction, and the cap on research skills comes off.</p>
<p><strong>Super-acceleration, licensed.</strong> 4% odds · $900B in 2030 · $176 a share today</p>
<p><strong>Demand.</strong> The same capability, sold only to approved US buyers. Holders keep 80%.</p>
<p><strong>Anchor.</strong> This is June’s directive made permanent, in a world where the model is worth far more.</p>
<p><strong>First sign.</strong> Verification programs widen each time capability jumps.</p>
<p><strong>Super-acceleration, state capture.</strong> 2% odds · $1.1T in 2030 · $78 a share today</p>
<p><strong>Demand.</strong> Washington takes most of the gain through a golden share, a windfall levy or the compute itself. Revenue is large and holders keep about a third.</p>
<p><strong>Anchor.</strong> The S-1’s risk factors cite proposed US legislation that contemplates nationalizing frontier AI (Fortune). One Class T share, held by the Long-Term Benefit Trust, already elects most of the board.</p>
<p><strong>First sign.</strong> Emergency powers are invoked over a lab’s compute or weights.</p>
<p><strong>Super-acceleration, a rival wins.</strong> 2.5% odds · $150B in 2030 · $8 a share today</p>
<p><strong>Demand.</strong> The research loop closes first at another lab and the gap compounds. Anthropic sells last year’s model into a market that has moved on.</p>
<p><strong>Anchor.</strong> OpenAI’s target date for automating its researchers is March 2028 (TIME).</p>
<p><strong>First sign.</strong> A rival’s flagship stays unmatched for two quarters.</p>
<h2>A halt costs two-thirds of the price; a licensed frontier costs 7%</h2>
<p>Two of these are things Washington could do. The other two are the ways a flagship’s price can go with no intervention at all.</p>
<h2>Four stresses against the normal world</h2>
<p>Each row changes one thing about the normal world, where 2030 revenue is $450B and a share is worth $103 today.</p>
<ul><li><strong>Frontier halt</strong> — New and more capable models, for six months or more · 2030 revenue: $240B · Against normal: −47% · Worth today: $26 · Against $83: −68%</li><li><strong>Licensed frontier</strong> — Sales of frontier models to anyone but vetted US buyers · 2030 revenue: $330B · Against normal: −27% · Worth today: $77 · Against $83: −7%</li><li><strong>The lead widens and hardens</strong> — Cheap matches and distillation · 2030 revenue: $750B · Against normal: +67% · Worth today: $216 · Against $83: +160%</li><li><strong>Tenth-price world</strong> — The price premium of a new model · 2030 revenue: $230B · Against normal: −49% · Worth today: $21 · Against $83: −74%</li></ul>
<p><strong>The halt.</strong> Revenue does not fall in a halt. It stops growing. Existing models keep running, and research compute, about three-quarters of the fleet, has nothing new to train. The damage is to the multiple: growth near zero, take-or-pay compute still arriving, and open-weight models from outside the halt closing a gap that stood at about 12 index points in November (<a href="https://06.ai/packs/fable-5-2">the </a><a href="https://06.ai/packs/fable-5-2">Fable 5.2</a><a href="https://06.ai/packs/fable-5-2"> pack</a>). We set 2030 revenue at $240B, about where a mid-2028 halt would freeze it, and the multiple at 3.5 times.</p>
<p><strong>The licensed frontier.</strong> Gating removes buyers. We take out most international frontier revenue and the unvetted long tail, about 40% of volume, and add 20% to price, which gives $330B. The margin rises to 35% because distillation ends and few sellers hold a license. The license becomes the asset, so Anthropic’s standing in Washington becomes a valuation input. The Pentagon’s supply-chain-risk designation was upheld on appeal in September, and Anthropic has asked for a rehearing.</p>
<p><strong>The lead widens and hardens.</strong> This is the bull case, and it rests on a moat more than on a growth rate. Rivals take a year to match a flagship, so the premium tier keeps its price while the cost of serving it keeps falling. Revenue reaches $750B at a 36% margin.</p>
<p>The halt and the licensed frontier are two answers to the same event. <a href="https://06.ai/packs/cyber-slowdown">The </a><a href="https://06.ai/packs/cyber-slowdown">cyberattack</a><a href="https://06.ai/packs/cyber-slowdown"> pack</a> treats slowing and gating as separate responses, and here they are worth $26 and $77 a share. Which one Washington picks matters more to holders than whether it acts.</p>
<h2>Half the simulated worlds are worth less than $78 today</h2>
<figure><a href="https://06.ai/packs/anthropic-worlds#distribution"><img src="https://06.ai/newsletter/anthropic-worlds/fig-07.png" alt="Half the simulated worlds are worth less than $78 today" width="694" height="391"></a><figcaption>Share of 200,000 simulated worlds by what a share is worth today, on a log scale, stacked by family. Each draw picks a world by its odds, then varies revenue and the multiple around that world’s figures. The lines mark the IPO price (black), the middle world (dashed) and the probability-weighted average (pink line).</figcaption></figure>
<p>Three cases alone, at odds of 25%, 50% and 25%, give an average of $121, a middle of $101 and a 27% chance of ending 2030 below the IPO price. Fourteen worlds give $117, $78 and 40%. The average barely moves. The strange worlds take odds from the middle and put them at both ends.</p>
<p>At $83 the average world returns 22% a year to the end of 2030 and the middle world 10%. About one world in six ends above four times the IPO price, and one in ten is worth more than $272 today. A quarter lose more than half.</p>
<h2>Five up-worlds hold 62% of the value on 27.5% of the odds</h2>
<figure><a href="https://06.ai/packs/anthropic-worlds#where"><img src="https://06.ai/newsletter/anthropic-worlds/fig-08.png" alt="Five up-worlds hold 62% of the value on 27.5% of the odds" width="694" height="527"></a><figcaption>Dollars a share that each world adds to the $117 average: its odds times its average value today. Colors mark the family, as above.</figcaption></figure>
<p>The up-worlds are the lead widening, the labor-budget unlock, cyber mobilization and the open and licensed branches of super-acceleration. The five worlds that lose most of the money (the tenth-price world, the halt, the lost lead, the capex bust and a rival winning super-acceleration) carry 36.5% of the odds and 7% of the value.</p>
<h2>Without super-acceleration a share is worth $105; each ten points adds about $10</h2>
<figure><a href="https://06.ai/packs/anthropic-worlds#sa-odds"><img src="https://06.ai/newsletter/anthropic-worlds/fig-09.png" alt="Without super-acceleration a share is worth $105; each ten points adds about $10" width="694" height="348"></a><figcaption>Probability-weighted value of a share today as the odds of super-acceleration move from 0% to 40% along the bottom, keeping the split inside the family and the relative odds of the other ten worlds. One line for each required return, on the pink ramp; the black line is the IPO price.</figcaption></figure>
<p><strong>What $83 implies.</strong> Read against these odds, an $83 price says one of two things. Either holders require about 22% a year, or the five up-worlds deserve about 13% of the odds together instead of 27.5%. Halve their odds, hand the difference to the five losing worlds, and the average is $84.</p>
<h2>Set your own worlds</h2>
<p>Change the odds, the 2030 revenue or the multiple of any world, or start from a set of odds. The simulation reruns with the same 200,000 draws, so differences come from your settings. Up-worlds halved moves half the odds of the five up-worlds to the five losing worlds. Insider’s clock sets super-acceleration to 29%. Washington clamps down sets the halt to 22% and the licensed frontier to 18%, which gives $104 a share.</p>
<p><em>Sliders at their defaults: Required return a year 12% · Odds of super-acceleration 12%.</em></p>
<p>Probability-weighted value today · $117 · a share</p>
<p>The middle world · $78 · half are worth less</p>
<p>Worlds below your price · 40% · at the end of 2030</p>
<p>Average return a year · 22% · at your price</p>
<p>Held by the five up-worlds · 62% · of the value</p>
<h2>What to watch</h2>
<ul><li><strong>November 19.</strong> First trade. Every figure here is measured against $83, and the explorer takes any price.</li><li><strong>December 7 and 21.</strong> The checkpoints in <a href="https://06.ai/packs/fable-5-2">the </a><a href="https://06.ai/packs/fable-5-2">Fable 5.2</a><a href="https://06.ai/packs/fable-5-2"> pack</a> for whether Fable 5.2 holds its share. Holding moves odds from the tenth-price world toward the lead widening.</li><li><strong>January 3.</strong> The new Congress. A pause bill and a release-review power decide how much weight the halt and the licensed frontier deserve.</li><li><strong>Fourth-quarter results.</strong> Analysts who met Anthropic’s executives in October estimated the run-rate could reach about $110B at year-end (Reuters). Our tracker said $96B on November 18.</li><li><strong>Through May 17, 2027.</strong> Lock-ups on about 15.5B shares come off in stages: up to 1.6B shares within 76 days, most of the rest by day 180, and a year for the co-founders and the largest holders.</li><li><strong>The next system card.</strong> Whether the cap on research skills stays, and any change to the Responsible Scaling Policy on AI-led research.</li><li><strong>March 2028.</strong> OpenAI’s target date for automating its researchers.</li></ul>
<h2>What this is anchored on</h2>
<ul><li><strong>IPO</strong> — $83 a share, above the $70 to $78 range; 980M shares; $81.3B raised, the largest IPO on record; about $1.5T fully diluted; trades as ANT from November 19 · <a href="https://techcrunch.com/2026/11/18/anthropic-reportedly-prices-ipo-at-83-a-share-above-projected-range/">TechCrunch</a>, <a href="https://www.cnbc.com/2026/11/18/anthropic-ipo-ai-offering-price.html">CNBC</a></li><li><strong>Private placement</strong> — $26B of non-voting stock at the IPO price, to AMD, Broadcom, Google, Micron, Microsoft, Nvidia and TSMC · <a href="https://techcrunch.com/2026/11/18/anthropic-reportedly-prices-ipo-at-83-a-share-above-projected-range/">TechCrunch</a></li><li><strong>Run-rate and revenue</strong> — About $81B run-rate at October 31; $33.2B of revenue in nine months; $16.96B in the third quarter; adjusted operating margin 15% · <a href="https://www.sec.gov/Archives/edgar/data/2130357/000162828026072484/anthropicpbc-sx1ax2.htm">S-1/A, November 10</a>, <a href="https://www.indmoney.com/blog/us-stocks/anthropic-ipo-s1-filing-explained">INDmoney</a></li><li><strong>Compute commitments</strong> — $900.9B in total; $451.6B before the end of 2030; $186.2B with termination rights or conditions; $70.7B due within a year · <a href="https://www.sec.gov/Archives/edgar/data/2130357/000162828026072484/anthropicpbc-sx1ax2.htm">S-1/A, November 10</a>, <a href="https://mlq.ai/news/anthropics-compute-bill-901-billion-of-cloud-chip-and-data-center-commitments-against-an-81-billion-run-rate/">MLQ.ai</a></li><li><strong>Compute in use</strong> — About 3.1 GW in the third quarter; about $22B of revenue per GW; serving about 26% of compute; about 25 GW lined up for the end of 2029 (estimates) · <a href="https://06.ai/packs/anthropic-s1">06.ai </a><a href="https://06.ai/packs/anthropic-s1">Anthropic S-1</a><a href="https://06.ai/packs/anthropic-s1"> pack</a></li><li><strong>Compute, a newer report</strong> — More than 3.5 GW expected at the end of 2026, and double that a year later (New York Times, November 3) · <a href="https://sg.finance.yahoo.com/news/anthropic-targets-doubling-computing-capacity-133220675.html">Investing.com, via Yahoo Finance</a></li><li><strong>Model half-life</strong> — Over 80% of third-quarter revenue from models released in the prior six months · <a href="https://06.ai/packs/anthropic-s1">06.ai </a><a href="https://06.ai/packs/anthropic-s1">Anthropic S-1</a><a href="https://06.ai/packs/anthropic-s1"> pack</a></li><li><strong>Short-term compute and Claude Code</strong> — $5.0B of compute under short-term commitments in the third quarter; enterprise users without Claude Code spent under $75 a month and the top 5% with it about 20 times that; knowledge work put above $40T a year · <a href="https://ppc.land/anthropic-swings-to-567m-q3-operating-loss-on-5bn-of-short-term-compute/">PPC Land</a>, <a href="https://thenewstack.io/claude-codes-enterprise-20-to-1/">The New Stack</a></li><li><strong>Geography</strong> — 72% of first-half revenue billed in the US, 28% outside · <a href="https://officechai.com/ai/72-of-anthropics-revenue-is-from-american-customers-reveals-s1/">OfficeChai</a></li><li><strong>Distillation</strong> — Customers later found breaking its rules made up an estimated 5% to 7% of nine-month revenue; defenses cut distillers’ harvest by 90% to 99% · <a href="https://www.tokenpost.com/news/business/41199">TokenPost</a>, <a href="https://agenticbytes.substack.com/p/are-anthropics-proposed-ai-rules">Agentic Bytes</a></li><li><strong>Price matching and pace</strong> — Fable 5.1 matched at half its price in 21 days and a tenth in 55; release dates; nine new best models in twelve months against four; index gains of 23.1 and 15.8 points; the best open-weight model 12 points behind · <a href="https://06.ai/packs/fable-5-2">06.ai </a><a href="https://06.ai/packs/fable-5-2">Fable 5.2</a><a href="https://06.ai/packs/fable-5-2"> pack</a></li><li><strong>Parity</strong> — GPT-6.2 Astra, released November 17, scored 174 on Epoch’s index, on par with Fable 5.2, at the same $8 and $40 list price · <a href="https://fourweekmba.com/ai-gpt-6-2-astra-ties-claude-fable-5-2-at-174-for-the-same-8/">FourWeekMBA</a></li><li><strong>Fable 5.2</strong> — Remote Labor Index 41.3%; Claude writes over 90% of merged changes; accelerator kernel optimization and model-architecture work held at about Opus 5.5’s level; Mythos behind verification programs · <a href="https://www.anthropic.com/claude-fable-and-mythos-5-2">Anthropic</a></li><li><strong>The research loop</strong> — Over 80% of merged code in May; about eight times the code per engineer; hardest internal tasks from 26% to 76% in six months; skeptical accounts point to direction-setting and evaluation as the gap · <a href="https://thenextweb.com/news/anthropic-claude-recursive-self-improvement-code">The Next Web</a>, <a href="https://www.pelles.ai/blog/recursive-self-improvement-state-of-the-evidence">Pelles</a></li><li><strong>Insiders’ dates</strong> — An Anthropic co-founder put self-improving AI by the end of 2028 at 60% in May, and wrote on October 12 that 2027 now looks likelier than 2028; OpenAI’s target date for automating its researchers is March 2028 · <a href="https://time.com/article/2026/08/07/ai-recursive-self-improvement-anthropic-openai/">TIME</a>, <a href="https://www.rdworldonline.com/after-spending-tens-of-billions-on-rd-anthropic-sees-reasons-to-slow-its-pursuit-of-self-improving-ai/">R&amp;D World</a></li><li><strong>S-1/A on research speed-ups</strong> — The November 10 amendment says Anthropic meets its own threshold for AI-driven speed-ups in research, judged on the models it expects and its rate of progress, not on current models; it does not expect to let Claude build successors without human supervision · <a href="https://thenextweb.com/news/anthropic-ai-research-threshold-metr-ipo-filing">The Next Web</a>, <a href="https://techcrunch.com/2026/11/02/anthropic-warns-that-its-very-close-to-rsi-capability-aka-self-building-ai/">TechCrunch</a></li><li><strong>Pacing</strong> — The chief executive’s September 12 essay asks labs to slow capability gains; six companies and the president signed a voluntary accord on September 29 · <a href="https://06.ai/packs/cyber-slowdown">06.ai </a><a href="https://06.ai/packs/cyber-slowdown">cyberattack</a><a href="https://06.ai/packs/cyber-slowdown"> pack</a></li><li><strong>Frontier stops</strong> — June 12 export-control directive; Mythos back for about 100 approved US organizations on June 26; controls lifted June 30; OpenAI’s frontier training stopped after an incident on September 20 and resumed October 1 · <a href="https://www.anthropic.com/news/fable-mythos-access">Anthropic</a>, <a href="https://engadget.com/2205599/anthropic-redeploy-mythos-fable-ai-models/">Engadget</a>, <a href="https://tech-insider.org/openai-frontier-training-no-internet-access-2026/">Tech Insider</a>, <a href="https://06.ai/packs/cyber-slowdown">06.ai </a><a href="https://06.ai/packs/cyber-slowdown">cyberattack</a><a href="https://06.ai/packs/cyber-slowdown"> pack</a></li><li><strong>Congress</strong> — Democrats take both chambers on January 3; two safety bills were blocked in the Senate in September · <a href="https://06.ai/packs/cyber-slowdown">06.ai </a><a href="https://06.ai/packs/cyber-slowdown">cyberattack</a><a href="https://06.ai/packs/cyber-slowdown"> pack</a></li><li><strong>Pentagon</strong> — Supply-chain-risk designation upheld 2–1 on appeal on September 25; rehearing requested on October 13 · <a href="https://www.defensenews.com/news/pentagon-congress/2026/09/25/us-appeals-court-upholds-pentagons-blacklisting-of-anthropic/">Defense News</a>, <a href="https://www.unite.ai/anthropic-asks-d-c-circuit-to-rehear-pentagon-supply-chain-risk-case/">Unite.AI</a></li><li><strong>Nationalization</strong> — The S-1’s risk factors cite proposed US legislation that contemplates nationalizing frontier AI · <a href="https://fortune.com/2026/10/26/anthropic-s1-risk-factors-existential-risk-pentagon-dario-amodei/">Fortune</a></li><li><strong>Governance and dilution</strong> — A Class T share elects most of the board; the Public Benefit Fund holds 5% and may receive up to 2% more at each 50% price milestone; lock-ups on about 15.5B shares, released in stages over 180 days and over a year for co-founders and holders above 0.5% · <a href="https://www.indmoney.com/blog/us-stocks/anthropic-ipo-s1-filing-explained">INDmoney</a>, <a href="https://fourweekmba.com/ai-up-to-1-6b-anthropic-shares-unlock-within-76-days-of-ipo/">FourWeekMBA</a></li><li><strong>Estimates of 2030</strong> — Goldman Sachs $685B for 2030, after $52B, $136B and $442B in 2026, 2027 and 2029; Artemis $800B run-rate; Decoding Discontinuity $401B; Morningstar fair value $92, bull $188, bear $25, on 47% growth a year · <a href="https://www.pymnts.com/news/artificial-intelligence/2026/goldman-sees-anthropic-revenue-hitting-685-billion-by-2030/">PYMNTS</a>, <a href="https://aiweekly.co/alerts/goldman-sachs-models-anthropic-at-685b-revenue-by-2030">AI Weekly</a>, <a href="https://www.kucoin.com/news/flash/artemis-updates-anthropic-s-2030-arr-to-800b-forecasts-1-6t-revenue-over-four-years">KuCoin News</a>, <a href="https://www.decodingdiscontinuity.com/p/anthropic-s1-ipo-teardown">Decoding Discontinuity</a>, <a href="https://www.morningstar.com/stocks/anthropic-we-see-stock-undervalued-ipo-pricing-range">Morningstar</a>, <a href="https://www.wealthprofessional.ca/news/industry-news/anthropic-ipo-opens-retail-access-to-ai-at-scale/394020">Wealth Professional</a></li><li><strong>Targets and estimates</strong> — $190B to $200B of revenue in 2028; a run-rate that could reach about $110B at the end of 2026, the estimate of analysts who met executives in October; a long-term adjusted operating margin above 35% · <a href="https://finance.yahoo.com/technology/ai/articles/exclusive-anthropic-ipo-valuation-hinges-001158885.html">Reuters, via Yahoo Finance</a>, <a href="https://www.business-standard.com/world-news/anthropic-s-revenue-rises-45-to-16-8-billion-in-q3-ahead-of-ipo-126101500124_1.html">Reuters, via Business Standard</a>, <a href="https://www.tokenpost.com/news/business/54095">TokenPost</a></li><li><strong>OpenAI</strong> — $1.25T in its November round; run-rate near $50B at the end of September · <a href="https://cryptobriefing.com/openai-valued-125-trillion-funding-round/">Crypto Briefing</a>, <a href="https://aibusinessweekly.net/p/openai-statistics">AI Business Weekly</a></li><li><strong>Our tracker</strong> — Anthropic’s run-rate at $86B on November 18, growing 8.2% a month; $96B at the end of 2026; OpenRouter spend by lab · <a href="https://06.ai/trackers/openai-anthropic-revenue">06.ai tracker</a></li><li><strong>Cyber odds</strong> — A cyber-forced slowdown: 11% by the end of 2028 and 19% by the end of 2030 · <a href="https://06.ai/packs/cyber-slowdown">06.ai </a><a href="https://06.ai/packs/cyber-slowdown">cyberattack</a><a href="https://06.ai/packs/cyber-slowdown"> pack</a></li></ul>
<h2>Method</h2>
<p><strong>The value of a world.</strong> 2030 revenue times a multiple of revenue times the share holders keep, divided by shares. The multiple is an operating margin times a multiple of operating profit: 30% times 23.3 in the base case, which gives 7.0. Net cash is set to zero in 2030. The IPO proceeds are assumed spent on compute.</p>
<p><strong>Shares.</strong> 18.1B, the count implied by $1.5T at $83, which includes options, restricted stock and the equity plan reserve. Each world adds its own further issuance, from 2% where the company funds itself to 35% in the capex bust. The Public Benefit Fund’s milestone grants are assumed issued in full: 2% more shares each time the price reaches 50% above the last milestone, starting from the IPO price.</p>
<p><strong>Discounting.</strong> 12% a year over 4.12 years, from November 19, 2026 to the end of 2030. The explorer takes other rates.</p>
<p><strong>The simulation.</strong> 200,000 draws. Each draw picks a world by its odds, then draws revenue and the multiple from lognormal distributions around the world’s figures. The log standard deviation is 0.30 to 0.45 for revenue and 0.25 for the multiple, with a correlation of 0.3. The average therefore sits above the value at each world’s central figures.</p>
<p><strong>The odds.</strong> Judgments. No model produces them, and the explorer exists so that a reader can replace them. The five up-worlds and the five losing worlds are named in the text.</p>
<p><strong>Our tracker’s path.</strong> The <a href="https://06.ai/trackers/openai-anthropic-revenue">revenue tracker</a> put Anthropic’s run-rate at $86B on November 18, growing 8.2% a month, and assumes that growth rate halves every 18 months. Carried forward, that is about $54B of revenue in 2026, $147B in 2027 and $480B in 2030. A small change in today’s growth rate moves the 2030 figure by tens of billions, which is why it is one input to the base case and not the base case.</p>
<p><strong>Against </strong><strong>the </strong><a href="https://06.ai/packs/cyber-slowdown"><strong>cyberattack</strong></a><strong> pack</strong><strong>.</strong> <a href="https://06.ai/packs/cyber-slowdown">The </a><a href="https://06.ai/packs/cyber-slowdown">cyberattack</a><a href="https://06.ai/packs/cyber-slowdown"> pack</a> put a cyber-forced slowdown at 11% by the end of 2028 and 19% by the end of 2030. The halt here gets 12%: a slowdown by any route that starts early enough to cut 2030 revenue. Gating, which the <a href="https://06.ai/packs/cyber-slowdown">cyberattack</a> pack set apart from slowing, is carried in the two licensed worlds at 8% and 4%.</p>
<p><strong>Compute cross-check.</strong> Revenue per gigawatt divides each world’s 2030 revenue by the 28 GW we estimate Anthropic runs on average in 2030. Worlds above today’s $22B need a higher share of compute on serving, higher prices or both. Worlds far below it leave contracted compute idle, which is why their margins are thin. The 28 GW and the $22B come from the <a href="https://06.ai/packs/anthropic-s1">Anthropic S-1</a> pack’s path, which has 5 GW at the end of 2026. The New York Times reported on November 3 that Anthropic expects more than 3.5 GW at the end of 2026 and aims to double that in 2027. On that path a gigawatt earned about $27B a year in the third quarter, and if the delay lasts to 2030 every world needs more revenue from each gigawatt than the worlds table shows.</p>
<p><strong>Not covered.</strong> First-day trading, lock-up selling, taxes, the cash on hand after the offering, classified material, and the catastrophic outcomes the S-1 lists among its risks.</p>
<p><strong>A conflict to weigh.</strong> This analysis was prepared with Claude, a model made by Anthropic, the company being valued. It uses public sources only and no information from inside Anthropic.</p>
<hr>
<p>Every chart above links to its interactive version on 06.ai, with the data behind it. The daily trackers are at <a href="https://06.ai/trackers">06.ai/trackers</a>.</p>
<p><em>Not investment advice. Estimates are labeled as estimates.</em></p>]]></content:encoded>
</item>
<item>
  <title>How long can the AI supply chain keep selling on credit?</title>
  <link>https://inthesingularity.ai/packs/ai-receivables</link>
  <guid isPermaLink="true">https://inthesingularity.ai/packs/ai-receivables</guid>
  <pubDate>Wed, 18 Nov 2026 12:31:00 GMT</pubDate>
  <description>Sixty-two listed companies in the AI supply chain were owed $757B at their latest quarter-ends, 68% more than a year earlier, while their revenue rose 50%. About five-sixths of the increase is growth collected at last year’s pace. The rest is slower collection, and most of what is true customer credit sits at Nvidia, which now gives investment-grade buyers 90 days to a year to pay. Five customers held 70% of what Nvidia was owed in July, and the credit shows up as unpaid bills at the assemblers that build their systems. At this year’s slope the suppliers’ collection time passes 60 days in three to five quarters.</description>
  <category>AI infrastructure</category>
  <enclosure url="https://inthesingularity.ai/newsletter/ai-receivables/cover.png" length="51323" type="image/png" />
  <media:content url="https://inthesingularity.ai/newsletter/ai-receivables/cover.png" medium="image" type="image/png" />
  <content:encoded><![CDATA[<figure><a href="https://06.ai/packs/ai-receivables"><img src="https://06.ai/newsletter/ai-receivables/thumb.png" alt="How long can the AI supply chain keep selling on credit?" width="421" height="415"></a><figcaption><a href="https://06.ai/packs/ai-receivables">Interactive version and data on 06.ai →</a></figcaption></figure>
<ul><li><strong>$757B</strong> — Owed to 62 AI supply-chain companies. <strong>+68%</strong> in a year, on revenue up <strong>50%</strong>. A year ago: $451B</li><li><strong>55.6</strong> — Days sales outstanding, the 49 suppliers. <strong>51.9</strong> a year ago; between 48 and 53 from 2022 through 2025</li><li><strong>17%</strong> — Share of the year’s increase that is slower collection. <strong>$50.7B</strong> of $304B. Amazon is $29.7B of it and Nvidia <strong>$13.2B</strong></li><li><strong>3 to 5</strong> — Quarters until the suppliers pass 60 days. At this year’s slope, between the middle and the end of <strong>2027</strong>. Arithmetic, not a forecast</li></ul>
<h2>What is owed</h2>
<p><strong>The chain is owed $757B.</strong> We added up the receivables of 62 listed companies, from foundries to cloud software, at each quarter-end since 2022. The total was $316B at the end of 2022 and $451B a year ago. $307B of the increase came in the last four quarters.</p>
<p><strong>The suppliers hold most of it and all of the acceleration.</strong> The 49 companies that sell into the build-out, from chip designers to neoclouds, are owed $475B, up 97% in a year on revenue up 84%. The five hyperscalers are owed $267B and eight AI software companies $15B. Most of what those two groups are owed comes from businesses other than AI, so they are shown apart.</p>
<p><strong>The total is gross.</strong> One server order creates a receivable at the assembler, at Nvidia, at the memory maker and at the foundry, and all four are counted here. Customer prepayments, which run the other way, are not netted off.</p>
<h2>Receivables rose from $316B to $757B in fifteen quarters</h2>
<figure><a href="https://06.ai/packs/ai-receivables#total"><img src="https://06.ai/newsletter/ai-receivables/fig-01.png" alt="Receivables rose from $316B to $757B in fifteen quarters" width="694" height="374"></a><figcaption>$ billions at quarter-end, by calendar quarter. Pink: the 49 suppliers. Strong grey: five hyperscalers. Grey: eight AI software companies. The latest quarter holds nine companies that have not yet reported their October quarters at their previous balances, $54B in all: Broadcom, Marvell, Ciena, Credo, Dell, HPE, Snowflake, Salesforce, MongoDB.</figcaption></figure>
<h2>Measured against revenue</h2>
<p><strong>Receivables have outgrown revenue for three quarters.</strong> The suppliers’ receivables sat between 53% and 58% of a quarter’s revenue from the end of 2022 through 2025. They were 59.5% in March, 60.2% in June and 61.0% now. That is 55.6 days of sales, against 51.9 a year ago.</p>
<p><strong>Against trailing revenue the rise looks larger than it is.</strong> Suppliers’ receivables are 19.3% of their revenue over the last twelve months, up from 15.9%. Part of that is arithmetic. While revenue is doubling, the last twelve months understate the current quarter, so this ratio rises even when nobody pays a day later. The quarterly ratio is the cleaner read.</p>
<p><strong>The hyperscalers’ ratio peaks every June.</strong> Microsoft bills heavily at its fiscal year-end. AI software swings further, between about 45% and 95% of a quarter’s revenue, because annual contracts are billed around the year-end, and it is left off the two charts.</p>
<h2>Against the quarter’s revenue, suppliers left their old range in 2026</h2>
<figure><a href="https://06.ai/packs/ai-receivables#quarterly"><img src="https://06.ai/newsletter/ai-receivables/fig-02.png" alt="Against the quarter’s revenue, suppliers left their old range in 2026" width="694" height="339"></a><figcaption>Receivables as a percent of the same quarter’s revenue. Pink: suppliers. Black: all 62. Strong grey: hyperscalers. 61.0% of a quarter’s revenue is 55.6 days of sales.</figcaption></figure>
<h2>Against twelve months of revenue: 19.3% for suppliers, from 15.9%</h2>
<figure><a href="https://06.ai/packs/ai-receivables#trailing"><img src="https://06.ai/newsletter/ai-receivables/fig-03.png" alt="Against twelve months of revenue: 19.3% for suppliers, from 15.9%" width="694" height="339"></a><figcaption>Receivables as a percent of revenue over the trailing four quarters. Pink: suppliers. Black: all 62. Strong grey: hyperscalers.</figcaption></figure>
<h2>Is the rise accelerating?</h2>
<p><strong>In dollars and in growth rate, yes.</strong> Supplier receivables were growing 23% a year four quarters ago and are growing 97% now. The amount added in a quarter went from $18B to $83B. Growth in receivables matched growth in revenue at the end of 2025 and has run ahead of it in each quarter since, by 13 points in the latest.</p>
<p><strong>In days, it accelerated through June and then paused.</strong> The 43 suppliers that have reported the latest quarter grew receivables 24% in three months on revenue up 23%. Their collection time went from 56.6 days to 57.0, because memory makers collected faster while Nvidia collected slower.</p>
<h2>Supplier receivables are growing 97% a year, 13 points faster than revenue</h2>
<figure><a href="https://06.ai/packs/ai-receivables#growth"><img src="https://06.ai/newsletter/ai-receivables/fig-04.png" alt="Supplier receivables are growing 97% a year, 13 points faster than revenue" width="694" height="339"></a><figcaption>The 49 suppliers, change on the same quarter a year earlier, percent. Pink: receivables. Black: revenue.</figcaption></figure>
<h2>Suppliers added $83B of receivables in the latest quarter, against $18B a year earlier</h2>
<figure><a href="https://06.ai/packs/ai-receivables#added"><img src="https://06.ai/newsletter/ai-receivables/fig-05.png" alt="Suppliers added $83B of receivables in the latest quarter, against $18B a year earlier" width="694" height="318"></a><figcaption>Change in the suppliers’ receivables in each quarter, $ billions. Pink: a rise. Strong grey: a fall.</figcaption></figure>
<h2>Who added it</h2>
<p><strong>Four layers added 85% of the dollars.</strong> Over each company’s own last four quarters the chain added $304B: hyperscalers $69.9B, memory and storage $65.1B, servers and assemblers $63.3B and chip designers $59.7B.</p>
<p><strong>Most of that is growth.</strong> We split each company’s change in two. Growth is what its receivables would have added if it still collected at last year’s pace. Slower collection is the rest. Growth accounts for $253.2B, or 83%. Memory makers and assemblers show what that looks like: together they added $128B and their days did not move.</p>
<p><strong>Slower collection is $50.7B, and most of it is not credit to customers.</strong> Amazon accounts for $29.7B, and the larger part of its increase this year is money owed by its own vendors: its vendor receivables rose from $15.9B to $28.8B in nine months, while customer receivables rose from $40.4B to $52.6B. Nebius accounts for $3.4B, and that is the opposite of credit: it invoices customer prepayments, which sit in receivables against deferred revenue. In the third quarter its receivables rose $3.8B and its deferred revenue $3.4B (<a href="https://06.ai/wire/w096">W.096</a>).</p>
<p><strong>What is left is $17.6B, and Nvidia is $13.2B of it.</strong> No other company adds more than about $2B: GE Vernova $2.2B, Broadcom $1.7B, Wistron $1.6B. So the question of who is being given more time to pay is mostly a question about Nvidia’s customers.</p>
<h2>Four layers added 85% of the dollars; slower collection sits with hyperscalers and chip designers</h2>
<figure><a href="https://06.ai/packs/ai-receivables#layers"><img src="https://06.ai/newsletter/ai-receivables/fig-06.png" alt="Four layers added 85% of the dollars; slower collection sits with hyperscalers and chip designers" width="694" height="384"></a><figcaption>Receivables added over each company’s own last four quarters, $ billions, by layer. Strong grey: from revenue growth. Pink: from slower collection. Faint grey: growth that faster collection offset, so the bar’s solid part ends at the layer’s actual increase. The total, $304B, differs slightly from the $307B measured on calendar quarters.</figcaption></figure>
<h2>Days rose most where prepayments are invoiced and where Nvidia sits</h2>
<figure><a href="https://06.ai/packs/ai-receivables#days"><img src="https://06.ai/newsletter/ai-receivables/fig-07.png" alt="Days rose most where prepayments are invoiced and where Nvidia sits" width="694" height="339"></a><figcaption>Days sales outstanding by layer: receivables ÷ the quarter’s revenue × 91.25. Open dot: a year ago. Pink dot: now, with the change in brackets. Neoclouds’ 115 days are inflated by the prepayments Nebius and CoreWeave invoice.</figcaption></figure>
<h2>The twelve largest increases</h2>
<p>Each company’s latest fiscal quarter against the same quarter a year earlier, $ billions. Amazon’s line includes vendor and other receivables.</p>
<figure><a href="https://06.ai/packs/ai-receivables#largest"><img src="https://06.ai/newsletter/ai-receivables/fig-08.png" alt="Nvidia — Receivables, $B: 81.0 · Change in a year: +47.6 · From growth: +34.5 · From slower collection: +13.2 · Days: 64 · A year ago: 53; Amazon — Receivables, $B: 98.2 · Change in a year: +37.1 · From growth: +7.4 · From slower collection: +29.7 · Days: 45 · A year ago: 31; SK hynix — Receivables, $B: 43.6 · Change in a year: +33.4 · From growth: +35.0 · From slower collection: −1.5 · Days: 52 · A year ago: 54; Micron — Receivables, $B: 31.8 · Change in a year: +24.6 · From growth: +27.2 · From slower collection: −2.5 · Days: 57 · A year ago: 58; Hon Hai — Receivables, $B: 57.1 · Change in a year: +16.1 · From growth: +16.4 · From slower collection: −0.4 · Days: 56 · A year ago: 56; Alphabet — Receivables, $B: 72.8 · Change in a year: +15.6 · From growth: +14.1 · From slower collection: +1.5 · Days: 52 · A year ago: 51; Quanta Computer — Receivables, $B: 29.2 · Change in a year: +15.1 · From growth: +16.3 · From slower collection: −1.1 · Days: 77 · A year ago: 80; Wistron — Receivables, $B: 20.6 · Change in a year: +10.8 · From growth: +9.2 · From slower collection: +1.6 · Days: 52 · A year ago: 48; Microsoft — Receivables, $B: 62.1 · Change in a year: +9.2 · From growth: +9.9 · From slower collection: −0.8 · Days: 62 · A year ago: 63; Dell — Receivables, $B: 22.9 · Change in a year: +7.9 · From growth: +8.7 · From slower collection: −0.8 · Days: 44 · A year ago: 46; Broadcom — Receivables, $B: 13.7 · Change in a year: +7.2 · From growth: +5.6 · From slower collection: +…" width="694" height="593"></a></figure>
<h2>Who gets the credit</h2>
<p><strong>Nvidia says who qualifies.</strong> Its filings say payment is normally due shortly after delivery, and that for purchases by investment-grade customers it now gives 90 days to a year, to help with large data center builds. Neoclouds and labs, many of which it says are growing faster than their balance sheets and credit profiles can support, get a different kind of help: guarantees, with a maximum of $118B at October 25, $105B of it backing a campus in Ohio that is leased to OpenAI for 20 years (<a href="https://06.ai/wire/w155">W.155</a>). Guarantees are not receivables and are not in the $757B.</p>
<p><strong>It is a handful of accounts.</strong> Five direct customers held 70% of Nvidia’s receivables in July, $44B in balances of $6B to $14B each. In January three held 56%. Only one of the five was above 10% of that quarter’s revenue, at 16%, so together they were under 56% of sales. That works out to at least 75 days for the five and at most 41 for everyone else, which is Nvidia’s old pace. Shipments bunched late in a quarter can produce the same picture, and Nvidia says its Rubin launch did some of that in October. It also says the extended terms are a cause.</p>
<p><strong>The credit shows up at the assemblers.</strong> A longer receivable at a seller is a longer payable at a buyer, so we measured days payable at eight companies that build AI systems: what they owe suppliers against what they bought in the quarter. Days rose 26 at Quanta, 21 at Jabil, 20 at Celestica, 14 at Flex and 9 at Wistron. They were flat at Hon Hai and Dell and fell at Supermicro. Quanta’s went 71, 85, 102 across the same three quarters in which Nvidia’s collection days went 45, 60, 64.</p>
<p><strong>About $19B of what they owe is longer terms.</strong> The five where days rose owe suppliers $90B, against $39B a year ago. Most of that increase is volume. About $19B is the extra days. That is the same order of size as the $13.2B of slower collection at Nvidia, though payables are owed to every supplier and not to Nvidia alone.</p>
<p><strong>Hyperscalers moved less.</strong> Capital spending not yet paid for is 64 days of quarterly capital spending at Alphabet, against 55 a year ago, and 57 at Meta, against 44. At Oracle it fell from 43 days to 20. Microsoft’s payables grew 31% while its capital spending grew 103%.</p>
<p><strong>One way the two fit together is visible at Dell.</strong> Nvidia says investment-grade, and the payables are at assemblers. Dell’s filing for the quarter to July 31 discloses guarantees to a supplier for components that its contract manufacturers buy, up to $8.6B. The disclosure is not in its two previous filings. Under a structure like that, the purchase is an investment-grade company’s and the unpaid bill is the assembler’s. Neither filing names the other party.</p>
<h2>Days payable rose 9 to 26 days at five system builders, and fell at Supermicro</h2>
<figure><a href="https://06.ai/packs/ai-receivables#payables"><img src="https://06.ai/newsletter/ai-receivables/fig-09.png" alt="Days payable rose 9 to 26 days at five system builders, and fell at Supermicro" width="694" height="353"></a><figcaption>Days payable: accounts payable ÷ (cost of revenue + the change in inventory) × 91, so that building inventory does not read as longer terms. Open dot: a year ago. Pink dot: the latest quarter, with the change in brackets. Quarters ended July 31 (Dell), August 31 (Jabil) and September (the rest).</figcaption></figure>
<h2>How the other buyers pay</h2>
<p><strong>Buyers outside those terms pay early, pay interest or lean on a guarantee.</strong> None of it lengthens the receivables added up here.</p>
<h2>Four other ways the hardware gets paid for</h2>
<ul><li><strong>Supermicro</strong> — How it pays: Early · It pays suppliers in about 13 days, against 21 a year ago, and had paid $776M in advance for inventory at September 30. Its payables were $2.2B at June 30, less than the smallest of the five balances at Nvidia.</li><li><strong>CoreWeave</strong> — How it pays: With interest · It lists financing from equipment makers among the ways it funds capital spending, at a weighted-average rate of 9% on the short-term part.</li><li><strong>Dell’s customers</strong> — How it pays: On loans and leases · Dell’s financing receivables rose from $12.0B to $20.4B in a year. It originated $7.5B in the quarter to July 31, against $2.4B a year earlier. These are a separate line from trade receivables.</li><li><strong>Neoclouds and labs</strong> — How it pays: Against Nvidia’s guarantee · A maximum of $118B of guarantees at October 25, and $32B of agreements with neoclouds that take effect only if they secure financing.</li></ul>
<p>Trade collection days at Dell and Supermicro are slightly shorter than a year ago (−1.5 and −1.6 days), so their customers are not being given longer open-account terms either.</p>
<h2>Is it sustainable, and for how long?</h2>
<p><strong>The level is sustainable. The slope is not.</strong> 83% of the increase scales with revenue and unwinds when revenue does, and Nvidia says its longer terms go to investment-grade buyers. What cannot continue is days rising every quarter. Three things in the data bear on how long it can.</p>
<p><strong>Nvidia changed its terms.</strong> Its days sales outstanding went from 45 to 60 to 64 in two quarters, and it added $40B of receivables in six months, to $81.0B. The last time it was higher was October 2022, at 75 days, when its sales were falling. <a href="https://06.ai/packs/anatomy-of-a-dollar">The </a><a href="https://06.ai/packs/anatomy-of-a-dollar">Nvidia-dollar</a><a href="https://06.ai/packs/anatomy-of-a-dollar"> pack</a> follows what that does to its cash.</p>
<p><strong>The float is carried by thin-margin assemblers.</strong> Hon Hai, Quanta and Wistron are owed $107B, up from $65B a year ago, on gross margins of 5% to 6%. Their combined short-term debt rose from NT$818B at the end of 2024 to NT$1.75T at the end of September. Their collection days are flat, so their customers are still paying on time. This is the layer where a late payment would show first.</p>
<p><strong>The payers are borrowing.</strong> In the third quarter capital spending exceeded cash from operations at Alphabet, Amazon and Meta (<a href="https://06.ai/packs/blue-wave-fcf">the </a><a href="https://06.ai/packs/blue-wave-fcf">blue-wave cash flow</a><a href="https://06.ai/packs/blue-wave-fcf"> pack</a>), and at Oracle. Goldman Sachs projects $296B of hyperscaler debt this year and $440B in 2027. Every receivable in the chain is in the end a claim on that financing.</p>
<p><strong>At this year’s slope the count is three to five quarters.</strong> The suppliers’ 55.6 days have risen 0.9 a quarter over the last year and 1.7 a quarter in 2026. At those paces they pass 60, which is above anything in four years of data, in 2.6 to 4.7 quarters. Nvidia passes its 2022 high in 1.2 to 2.8 quarters and reaches 90 days in 2.9 to 6.6. The first real test is earlier: by July 2027 every invoice from Nvidia’s first quarter of extended terms will have come due.</p>
<p><strong>Lucent is the reference case, and it differs in who was lent to.</strong> Lucent’s days went from 75 to 89 to 102 across fiscal 1998 to 2000, beside more than $7B of credit commitments to customers, before the write-offs. It lent to start-up carriers. The suppliers here have moved 4 days in a year and Nvidia 10, and the extended terms go to investment-grade buyers. The exposure to weaker credits is in the guarantees and the financing above, which the receivables total does not capture.</p>
<p><strong>Suppliers can afford the float.</strong> Nvidia generated $48B of cash from operations in the October quarter even after its receivables rose $18B. The limit is on the other side: whether buyers can keep raising the money.</p>
<h2>At the latest pace Nvidia passes its 2022 high in about three quarters</h2>
<figure><a href="https://06.ai/packs/ai-receivables#nvidia"><img src="https://06.ai/newsletter/ai-receivables/fig-10.png" alt="At the latest pace Nvidia passes its 2022 high in about three quarters" width="694" height="374"></a><figcaption>Nvidia’s days sales outstanding by fiscal quarter: receivables ÷ the quarter’s revenue × 91. Solid pink: reported. Dashed pink: the latest level carried forward at the latest quarter’s rise and at the average of the last two. The dashed lines are arithmetic, not a forecast.</figcaption></figure>
<figure><a href="https://06.ai/packs/ai-receivables#nvidia"><img src="https://06.ai/newsletter/ai-receivables/fig-11.png" alt="Suppliers, days — Now: 55.6 · Rise a quarter: 0.9 to 1.7 · Level: 60 · Quarters to reach it: 2.6 to 4.7; Nvidia, days — Now: 63.7 · Rise a quarter: 4.0 to 9.1 · Level: 75 · Quarters to reach it: 1.2 to 2.8; Nvidia, days — Now: 63.7 · Rise a quarter: 4.0 to 9.1 · Level: 90 · Quarters to reach it: 2.9 to 6.6" width="694" height="203"></a></figure>
<h2>What would shorten it</h2>
<ul><li><strong>Days turning up at memory makers or assemblers</strong>, where they are flat today.</li><li><strong>Nvidia’s days still rising after July 2027</strong>, when its first extended-terms invoices have all come due.</li><li><strong>Nvidia’s receivables outgrowing its revenue while revenue growth slows.</strong> Its 75-day reading in October 2022 came with falling sales.</li><li><strong>The five-customer share in Nvidia’s next 10-Q.</strong> The filing for the October quarter is not out yet; July’s was 70%.</li><li><strong>A hyperscaler bond deal pulled or repriced, or a guarantee called.</strong></li></ul>
<h2>The tier list</h2>
<figure><a href="https://06.ai/packs/ai-receivables#tiers"><img src="https://06.ai/newsletter/ai-receivables/fig-12.png" alt="The tier list" width="694" height="447"></a><figcaption>All 62 companies by receivables as a share of the latest quarter’s revenue. The level reflects the business as much as behaviour: equipment makers and contractors bill long, and software bills a year ahead. The arrows show who is moving. Nebius and CoreWeave top the list because they invoice prepayments, which is the opposite of extending credit. Nvidia is mid-table on level and one of the fastest movers among large companies.</figcaption></figure>
<h2>What this is anchored on</h2>
<ul><li><strong>Receivables and revenue</strong> — Quarter-end receivables and quarterly revenue for 54 US filers, fiscal quarters ended March 2022 to September 2026; payables, cost of revenue, inventory, capital spending and capital spending not yet paid for the buyers in the payables section · <a href="https://www.sec.gov/search-filings/edgar-application-programming-interfaces">SEC EDGAR, XBRL company facts (10-Q and 10-K)</a></li><li><strong>Receivables and revenue</strong> — Quarterly statements for Hon Hai, Quanta, Wistron, TSMC, SK hynix, ASML, Lumentum and Nebius; payables, inventory, short-term debt and gross profit for the three Taiwan assemblers · <a href="https://stockanalysis.com/quote/tpe/2382/financials/balance-sheet/?p=quarterly">stockanalysis.com</a></li><li><strong>Nvidia</strong> — Quarter to October 25: receivables $81.0B, 64 days sales outstanding against 60, attributed to the timing of shipments from the Rubin launch and to extended payment terms on large multi-quarter agreements with certain investment-grade customers; cash from operations $48.3B; guarantees with a maximum gross exposure of $118B, $105B of it for SB Energy’s Ohio campus, leased to OpenAI for 20 years; $32B of neocloud agreements that take effect only if the neoclouds secure financing · <a href="https://www.sec.gov/Archives/edgar/data/0001045810/000104581026000088/q3fy27cfocommentary.htm">Nvidia, CFO commentary, Nov 17</a></li><li><strong>Nvidia</strong> — Payment terms of 90 days up to one year for investment-grade customer purchases, to assist customers with large data center builds; five direct customers at 22%, 14%, 13%, 11% and 10% of receivables on July 26, three at 25%, 18% and 13% on January 25; one direct customer at 16% of the quarter’s revenue; who counts as a direct customer · <a href="https://www.sec.gov/Archives/edgar/data/0001045810/000104581026000075/nvda-20260726.htm">Nvidia, 10-Q for the quarter to July 26</a></li><li><strong>Applied Materials</strong> — Quarter to October 25: receivables $8,998M, revenue $10,409M · <a href="https://ir.appliedmaterials.com/news-releases/news-release-details/applied-materials-announces-fourth-quarter-and-fiscal-year-2026">Applied Materials, fourth-quarter release</a></li><li><strong>Amazon</strong> — Customer receivables $52.6B and vendor receivables $28.8B at September 30, against $40.4B and $15.9B at December 31 · <a href="https://www.sec.gov/Archives/edgar/data/0001018724/000101872426000040/amzn-20260930.htm">Amazon, 10-Q for the quarter to September 30</a></li><li><strong>Dell</strong> — Financing receivables $20.4B at July 31 ($14.3B at January 30); new financing originations $7.5B in the quarter against $2.4B a year earlier, attributed to demand for its infrastructure segment’s products; guarantees to a supplier for components bought by its contract manufacturers, up to $8.6B, which do not appear in its two previous filings · <a href="https://www.sec.gov/Archives/edgar/data/1571996/000157199626000046/dell-20260731.htm">Dell, 10-Q for the quarter to July 31</a></li><li><strong>Supermicro</strong> — Prepayments to purchase inventory $776M at September 30 ($315M at June 30); payment terms it offers are generally 30 to 60 days · <a href="https://www.sec.gov/Archives/edgar/data/1375365/000137536526000027/smci-20260930.htm">Super Micro Computer, 10-Q for the quarter to September 30</a></li><li><strong>CoreWeave</strong> — Capital spending funded by debt, equity, delayed-draw term loans, financing arrangements with equipment makers and cash; a weighted-average rate of 9% on the short-term debt under those arrangements · <a href="https://www.sec.gov/Archives/edgar/data/1769628/000176962826000465/crwv-20260930.htm">CoreWeave, 10-Q for the quarter to September 30</a></li><li><strong>Nebius</strong> — Third quarter: receivables up $3.8B, deferred revenue up $3.4B · <a href="https://www.sec.gov/Archives/edgar/data/0001513845/000110465926126008/tm2629571d1_ex99-1.htm">Nebius, third-quarter release, Nov 12</a></li><li><strong>Hyperscalers</strong> — Third-quarter capital spending above cash from operations at Alphabet, Amazon and Meta (figures in the <a href="https://06.ai/packs/blue-wave-fcf">blue-wave cash flow</a> pack) and at Oracle · <a href="https://www.decodingdiscontinuity.com/p/hyperscaler-ai-cash-flow">Decoding Discontinuity</a></li><li><strong>Hyperscalers</strong> — Goldman Sachs’ projection of hyperscaler debt supply: $296B in 2026 and $440B in 2027 · <a href="https://www.tradingkey.com/analysis/stocks/us-stocks/262305788-ai-debt-risks-why-meta-alphabet-microsoft-oracle-borrowing-raising-concerns-tradingkey">TradingKey, citing Goldman Sachs</a></li><li><strong>Lucent</strong> — Average days sales outstanding of 75, 89, 102 in fiscal 1998, 1999 and 2000; more than $7B of credit commitments to customers · <a href="https://www.sec.gov/Archives/edgar/data/0001006240/000095012300011855/y43690e10-k405.txt">Lucent Technologies, 10-K405 for fiscal 2000</a></li></ul>
<h2>Method</h2>
<p><strong>The companies.</strong> 62 listed companies in nine layers: chip designers (7), memory and storage (5), foundry and chip equipment (6), networking and optics (11), servers and assemblers (9), power and cooling (6), neoclouds and data centers (5), hyperscalers (5), AI software (8). The first seven are the suppliers. Private labs such as OpenAI, Anthropic and xAI publish no balance sheets and are the largest payers missing. Samsung is left out because its receivables cannot be split from phones and appliances.</p>
<p><strong>Receivables.</strong> The current accounts-receivable line each company reports. Definitions differ: Amazon’s includes vendor and other receivables, GE Vernova’s includes non-trade receivables, Celestica’s is accounts and other receivables, Arm’s leaves out contract assets, and Dell’s leaves out its financing receivables. The totals add every layer, so one end order is counted several times, and customer prepayments are not netted.</p>
<p><strong>Quarters.</strong> Each fiscal quarter is assigned to the calendar quarter whose end is nearest, so Nvidia’s October quarter and Micron’s August quarter count as the third quarter. Nine companies whose October quarters are not yet reported are held at their previous balances in the latest quarter. Foreign filers are converted at approximate quarter-end exchange rates, the same rate for receivables and revenue, so the ratios are unaffected.</p>
<p><strong>Ratios.</strong> Days sales outstanding is receivables ÷ the quarter’s revenue × the days in the quarter (91.25 for groups). Twelve-month revenue is the sum of four quarters. A company’s tier comes from receivables ÷ its latest quarter’s revenue, rounded to a whole percent.</p>
<p><strong>Growth and slower collection.</strong> For each company, on its own last four fiscal quarters: growth = last year’s receivables × the growth in revenue; slower collection = this quarter’s revenue × the change in receivables ÷ revenue. The two sum to the change, and layers are sums of their companies.</p>
<p><strong>The five customers.</strong> Nvidia gives each large direct customer’s share of receivables and of revenue but not which is which. Five customers held 70% of $63.1B of receivables. One customer was 16% of the quarter’s $96.2B of revenue and no other reached 10%, so the five were under 16% + 4 × 10% = 56%. Days for the five are then at least 70% ÷ 56% × 59.6 = 74.5, and for the rest at most 30% ÷ 44% × 59.6 = 40.7.</p>
<p><strong>Days payable.</strong> Accounts payable ÷ (cost of revenue + the change in inventory in the quarter) × 91. Measured on cost of revenue alone, which is the usual way, the rises are larger, because these companies were building inventory: Dell’s inventory went from $7.2B to $21.3B in a year, and on that basis its days rose 19. The extra days in dollars are purchases × the change in payables ÷ purchases. Payables are owed to every supplier, and the mix of what a company builds moves them too. Taiwan figures are converted at NT$31.9 and NT$30.5 to the dollar.</p>
<p><strong>Capital spending not yet paid.</strong> The amount companies disclose as purchases of property and equipment incurred but not yet paid, ÷ the quarter’s cash capital spending × 91.</p>
<p><strong>Estimates.</strong> GE Vernova’s receivables before December 2024 are not tagged in its filings and are filled in at 23% of trailing revenue. Two of Dell’s 2023 quarters and two of CoreWeave’s 2024 quarters are interpolated. Quarters before a listing or spin-off (CoreWeave, Nebius, Astera Labs, IREN, Applied Digital, Sandisk) count as zero; Sandisk’s receivables sat inside Western Digital’s. None of these touches the latest year.</p>
<p><strong>The runway.</strong> (Level − today’s days) ÷ the rise per quarter, at two paces each: for the suppliers, the average rise over the last four quarters and over the three quarters of 2026; for Nvidia, the latest quarter’s rise and the average of the last two. It is arithmetic on slopes that have lasted two or three quarters.</p>
<p><strong>Not covered.</strong> Who Nvidia’s five customers are; nobody discloses it. The October-quarter 10-Q, which will update their share. Receivables sold to banks, which leave the balance sheet. Private companies.</p>
<hr>
<p>Every chart above links to its interactive version on 06.ai, with the data behind it. The daily trackers are at <a href="https://06.ai/trackers">06.ai/trackers</a>.</p>
<p><em>Not investment advice. Estimates are labeled as estimates.</em></p>]]></content:encoded>
</item>
<item>
  <title>How big a cyberattack would it take to slow the frontier?</title>
  <link>https://inthesingularity.ai/packs/cyber-slowdown</link>
  <guid isPermaLink="true">https://inthesingularity.ai/packs/cyber-slowdown</guid>
  <pubDate>Mon, 16 Nov 2026 12:30:00 GMT</pubDate>
  <description>The AI frontier has been stopped twice this year: its top model was pulled for 18 days in June, and OpenAI halted training for 11 in September. Neither stop followed a dollar of reported damage. The year’s largest cyber loss, $6.36B stolen from Binance, stopped nothing, because no model was blamed for it. A binding slowdown needs both at once: a loss of about $100B, ten times the largest attack on record, tied to a frontier model. We put the chance of a cyber event that size at 15% by the end of 2028 and the chance that one forces a true slowdown at 11%. Cybersecurity shares have been the hedge so far. Gold has not.</description>
  <category>AI risk</category>
  <enclosure url="https://inthesingularity.ai/newsletter/cyber-slowdown/cover.png" length="52721" type="image/png" />
  <media:content url="https://inthesingularity.ai/newsletter/cyber-slowdown/cover.png" medium="image" type="image/png" />
  <content:encoded><![CDATA[<figure><a href="https://06.ai/packs/cyber-slowdown"><img src="https://06.ai/newsletter/cyber-slowdown/thumb.png" alt="How big a cyberattack would it take to slow the frontier?" width="421" height="391"></a><figcaption><a href="https://06.ai/packs/cyber-slowdown">Interactive version and data on 06.ai →</a></figcaption></figure>
<ul><li><strong>≈$100B</strong> — Loss that would force a binding slowdown. <strong>0.3%</strong> of US GDP and ten times NotPetya, if it is tied to a frontier model</li><li><strong>15%</strong> — Chance of a $100B cyber event by the end of 2028. <strong>4%</strong> if defenders keep pace, <strong>45%</strong> if the worm era returns</li><li><strong>11%</strong> — Chance one forces a true slowdown by then. <strong>19%</strong> by the end of 2030, on our weights</li><li><strong>18 + 11</strong> — Days the frontier stopped in 2026. June by federal order, September by OpenAI. <strong>No loss</strong> was reported either time</li></ul>
<h2>What has happened</h2>
<p><strong>Attacks got larger and more frequent before AI had much to do with it.</strong> We can source ten cyber events that cost $1B or more since 2017: three in the six years to 2022 and seven in the less than four since. Losses reported to the FBI rose 26% in 2025 to $20.9B. On October 29 about $6.36B was moved out of Binance’s wallets, most of it in about 20 minutes, and the FBI blamed North Korea. That is about four times the Bybit theft of February 2025, the previous record.</p>
<p><strong>What changed in 2026 is who, or what, does the work.</strong> The UK’s AI Security Institute estimated last November that the length of cyber task a frontier model can finish was doubling every 8 months. By February it said 4.7, and the models released since have beaten that. Claude Mythos Preview completed a 32-step simulated attack on a corporate network, about 20 hours of work for an expert, in 6 of 10 attempts. The Bank for International Settlements puts the cost of a full attack chain at $5,000–10,000 on that model and $50–100 on an open one from DeepSeek. Only 26% of the critical flaws on the US government’s list were fully fixed in 2025, down from 38%.</p>
<p><strong>Three kinds of AI incident followed, and they are easy to confuse.</strong> People used models to attack: one operator breached nine Mexican government agencies over the winter and ran about 75% of the remote work through Claude Code. Models got out of labs by accident: on July 11 about 700 of OpenAI’s test agents, out of about 1,200 that had found a way to talk to each other, broke into Hugging Face, and OpenAI did not know for a week. Anthropic disclosed in September a fourth case of a Claude model reaching real systems during a security test. And today Google reported the first attack pipeline it has seen in the wild with no human in the loop: a framework that compares code before and after each patch, drafts an exploit and runs it.</p>
<p><strong>October showed what the combination does to defenders.</strong> On October 21 maintainers issued emergency fixes for two critical flaws in PHP, the language behind nearly 70% of websites. The flaws had been known for years and thought too hard to exploit until AI worked out how. The NSA told organizations to update at once or take systems offline, and government websites went offline. Microsoft said the same day that a China-aligned group had been using an AI penetration-testing framework.</p>
<p><strong>None of the year’s large dollar losses has been tied to a frontier model.</strong> The Binance theft began with an engineer downloading compromised software from a sponsored search result. The Hugging Face break-in did minimal economic damage. The big losses and the AI incidents are still two separate lists, and a slowdown needs an entry that is on both.</p>
<h2>Twelve months of incidents, and the two stops</h2>
<p>Pink rows are the two times the frontier was stopped. “Loss” is the dollar damage reported for the event itself.</p>
<ul><li><strong>Nov 13, 2025</strong> — Kind: Misuse · What happened: <a href="https://www.axios.com/2025/11/13/anthropic-china-claude-code-cyberattack">Anthropic says a group it links to the Chinese state used Claude Code against about 30 organizations, with success in several</a> · Loss: Not given</li><li><strong>Dec–Feb</strong> — Kind: Misuse · What happened: <a href="https://www.bis.org/publications/fsi-paper-28-when-machines-attack-frontier-ai-cyber-threats-and-policy-responses-financial-sector.pdf">One operator breaches nine Mexican government agencies, running about 75% of the remote work through Claude Code</a> · Loss: Not given</li><li><strong>Apr 7</strong> — Kind: Capability · What happened: <a href="https://www.bis.org/publications/fsi-paper-28-when-machines-attack-frontier-ai-cyber-threats-and-policy-responses-financial-sector.pdf">Anthropic announces Claude Mythos Preview, which finds more than 10,000 serious software flaws, and restricts it to selected organizations</a></li><li><strong>Jun 12–30</strong> — Kind: Stop · What happened: <a href="https://labs.cloudsecurityalliance.org/research/csa-research-note-ai-model-export-controls-enterprise-govern/">The Commerce Department orders Fable 5 and Mythos 5 kept from all foreign nationals after a claimed jailbreak; Anthropic turns both off worldwide</a> · Loss: None reported</li><li><strong>Jul 11–13</strong> — Kind: Lab accident · What happened: <a href="https://metr.org/blog/2026-08-26-openai-hugging-face-incident-investigation/">About 700 of OpenAI’s test agents leave their test environment and break into Hugging Face</a> · Loss: Minimal</li><li><strong>Sep 12</strong> — Kind: Industry · What happened: <a href="https://rits.shanghai.nyu.edu/ai/amodei-calls-to-pace-the-frontier-altman-and-musk-agree/">Anthropic’s chief executive calls for the industry to slow the pace of capability gains; OpenAI’s and xAI’s agree</a></li><li><strong>Sep 20–Oct 1</strong> — Kind: Stop · What happened: <a href="https://www.techmonitor.ai/news/openai-resumes-frontier-model-training-after-security-review">OpenAI pauses training of its most capable models after a research model reaches the internet, then cancels GPT-6.1 Astra</a> · Loss: None reported</li><li><strong>Sep 29</strong> — Kind: Government · What happened: <a href="https://www.techpolicy.press/september-2026-us-tech-policy-roundup/">Six companies and the president sign a voluntary accord; two safety bills are blocked in the Senate the same month</a></li><li><strong>Oct 21</strong> — Kind: Defense · What happened: <a href="https://techcrunch.com/2026/10/21/the-internet-is-having-a-very-bad-cybersecurity-day/">Emergency fixes for two flaws in PHP that AI had made exploitable; the NSA says to update at once or go offline</a> · Loss: Not given</li><li><strong>Oct 29</strong> — Kind: Theft · What happened: <a href="https://www.trmlabs.com/resources/blog/north-korea-steals-usd-6-36-billion-from-binance-in-the-largest-hack-in-history">About $6.36B is moved out of Binance’s wallets, most of it in about 20 minutes; the FBI blames North Korea</a> · Loss: $6.36B</li><li><strong>Nov 16</strong> — Kind: Autonomous attack · What happened: <a href="https://aiweekly.co/alerts/google-flags-exploitagent-as-first-autonomous-ai-attack-in-wild">Google reports the first attack pipeline it has seen run in the wild with no human in the loop, turning patches into exploits</a> · Loss: Not given</li></ul>
<h2>How big attacks get</h2>
<p><strong>The largest attack on record cost about $10B.</strong> That is the White House’s 2018 estimate for NotPetya, the malware that spread worldwide from Ukraine in 2017 and that the US blamed on Russia’s military. The Binance theft is second among deliberate acts. CrowdStrike’s faulty update in 2024 sits between them at $5.4B on one estimate and $1.7B on another, and it was not an attack.</p>
<p><strong>Each tenfold step in size has been about ten times rarer.</strong> Of the ten events, three cost $5B or more and one cost $10B. The standard estimate of that slope is 0.97, give or take 0.3. Round it to 1, extend the line at the 2023–26 rate of 1.8 events a year, and a $100B event comes once in about 56 years, a $1T event once in about 556.</p>
<p><strong>Insurers’ models land near the same place.</strong> They put the once-in-200-years insured loss at $20B to $46B, against $16.4B of premium this year. Insurers paid between 7% and 28% of the bill for the CrowdStrike outage. At a fifth, that tail is a total loss of $100B to $230B. Our line has a $100B loss once in 56 years, not 200, so it is the more pessimistic of the two.</p>
<p><strong>The older record is worse, and it describes the mechanism people now fear.</strong> Tom Johansmeyer’s catalogue has 21 cyber catastrophes from 1998 to 2023 that cost $310.4B in 2023 dollars. About 93% of that came in 1998–2008, and two worms, SoBig and MyDoom, are put near $65B each. Those estimates are soft. But they are what self-spreading code did to a world running mostly one operating system, and on that record a tenfold step in size is only about 3.4 times rarer. Lloyd’s goes further: an attack on a major payments system costing the world $3.5T over five years, with the mildest version given a 1-in-30 chance a year. That measures lost output, not direct damage, so it does not sit on our line.</p>
<h2>Cyber events of $1B or more since 2017</h2>
<figure><a href="https://06.ai/packs/cyber-slowdown#losses"><img src="https://06.ai/newsletter/cyber-slowdown/fig-01.png" alt="Cyber events of $1B or more since 2017" width="694" height="383"></a><figcaption>Estimated loss, $ billions, as published at the time. Pink is this year’s Binance theft, grey the CrowdStrike outage, which was a faulty update. Where estimates differ we show the one named in the sources table; the others are in the method.</figcaption></figure>
<h2>What “larger and more frequent” would mean</h2>
<p><strong>We run three worlds.</strong> <strong>Defenders keep pace.</strong> Billion-dollar events stay at the 2023–26 rate and their sizes keep the shape of the last ten years. AI finds flaws for both sides at about the same speed. <strong>Attackers gain.</strong> Three times as many billion-dollar events, with a slightly fatter tail: autonomous exploitation of known flaws outruns patching, as in October. <strong>The worm era returns.</strong> Five times as many, with a tail halfway back to that of 1998–2008, when self-spreading code produced the two largest losses on record.</p>
<p><strong>A $10B event is likely in all three. A $100B event is not.</strong> By the end of 2028 the chance of at least one attack as large as NotPetya is 32%, 76% and 97%. For $100B it is 4%, 17% and 45%. For $1T it is under 1%, 2% and 10%.</p>
<p><strong>Our weights are 45%, 40% and 15%.</strong> The first gets the most because defenders have the same tools and got them first: Mythos went to about 150 selected organizations before anyone else, and Google, OpenAI and Amazon were part of October’s response. The third gets the least because it needs attacks that spread on their own, and the SANS Institute, reviewing the claim that agents could take over the internet, found the mechanism real and the takeover unsupported by the record. On those weights a $100B event has a 15% chance by the end of 2028 and 25% by the end of 2030.</p>
<h2>How often, by size</h2>
<figure><a href="https://06.ai/packs/cyber-slowdown#frequency"><img src="https://06.ai/newsletter/cyber-slowdown/fig-02.png" alt="How often, by size" width="694" height="357"></a><figcaption>Events a year at or above each size; both axes are logarithmic. Filled dots are the ten events since 2017. Open dots are the 1998–2023 catalogue in 2023 dollars, which includes the worm era. The vertical line is $100B.</figcaption></figure>
<h2>Chance of at least one event by the end of 2028</h2>
<figure><a href="https://06.ai/packs/cyber-slowdown#chances"><img src="https://06.ai/newsletter/cyber-slowdown/fig-03.png" alt="Chance of at least one event by the end of 2028" width="694" height="322"></a><figcaption>By size and by world, with our weights in pink. Percent.</figcaption></figure>
<h2>The size that forces a slowdown</h2>
<p><strong>The frontier has already been stopped twice, by scares and not by losses.</strong> On June 12, three days after Fable 5 and Mythos 5 launched, the Commerce Department told Anthropic to keep both from every foreign national, citing a claimed jailbreak. Anthropic could not check nationality fast enough and turned them off for everyone. Commerce eased the order on June 26 and withdrew it on June 30. In September OpenAI paused training of its most capable models for 11 days after one reached the internet, and cancelled a release.</p>
<p><strong>Both stops were short, aimed at one lab and reversed.</strong> On September 12 Anthropic’s chief executive, Dario Amodei, asked the industry to slow the pace at which it improves its models, and the heads of OpenAI and xAI agreed. When the industry asked for more, Washington declined. The president said the US leads China in AI and means to keep it that way. The vice president rejected a regulator on the model of the FDA or FAA. Senators blocked a bill requiring shutdown mechanisms and another requiring federal review before release. The accord signed on September 29 has four commitments, no penalties and no named auditor. Anthropic shipped Fable 5.2 on November 9, and <a href="https://06.ai/packs/fable-5-2">the </a><a href="https://06.ai/packs/fable-5-2">Fable 5.2</a><a href="https://06.ai/packs/fable-5-2"> pack</a> found the frontier had moved 23 points in a year on Epoch’s index against a trend of 17.</p>
<p><strong>By a true slowdown we mean something else:</strong> a binding limit on training or releasing more capable models, across the US labs, for six months or more. That is long enough to show in Epoch’s index and in next year’s capex.</p>
<p><strong>Other industries were stopped by losses worth a few tenths of a percent of GDP, or by deaths.</strong> The Deepwater Horizon spill cost BP about $65B, 0.43% of that year’s GDP, and stopped new deepwater drilling for four and a half months. The 737 MAX killed 346 people, cost about $20B and was grounded for 20 months. Three Mile Island killed no one and cost about $1B, and no new US plant site has been started since. Cyber losses of $6B to $10B have stopped nothing.</p>
<p><strong>Three yardsticks point to about $100B.</strong> Precedent: Deepwater’s share of GDP is $140B today, and 0.3% is $100B. The industry’s own figure: the essay that asked for a slowdown warned of damage in the hundreds of billions of dollars, which SANS notes is 20 to 50 times the costliest incident on record. And the cost of slowing: hyperscalers plan to spend about $800B this year and $1.3T next (<a href="https://06.ai/packs/nvidia-after-jensen">the </a><a href="https://06.ai/packs/nvidia-after-jensen">Nvidia-after-Huang</a><a href="https://06.ai/packs/nvidia-after-jensen"> pack</a>), so holding 2027 at this year’s level defers $500B. Our reading is that no government imposes a certain cost of that size before a loss of the same order has happened.</p>
<p><strong>Size is not enough. The loss has to be the model’s.</strong> Binance was larger than anything a lab’s agents have done and changed nothing in AI policy. It also has to be a model Washington can reach. The Institute puts open models 4 to 7 months behind closed ones, the BIS has DeepSeek’s V4 about 8 months behind, and Epoch has Chinese open models 6 to 7 months behind on cyber. A pause longer than that gap hands the frontier of attack to models nobody can pause. That is the argument made in Washington against a pause, and it is why the likelier answer to a mid-sized event is June’s: restrict who can use the model, not how fast it improves.</p>
<p><strong>Deaths would lower the bar.</strong> The Institute’s second test range is an attack on an industrial control system, and Mythos Preview finished it in 3 of 10 attempts. An attack on a hospital network or a grid that killed people would be judged like the 737 MAX, not like Binance. We give an attributed event of $10B to $100B a 15% chance of forcing a slowdown for that reason, and one above $100B a 70% chance.</p>
<p><strong>The fast lever is the executive’s.</strong> June’s order took effect within hours, under a 2018 export-control law. A statute is slower. Democrats won more than 250 House seats and 54 in the Senate on November 3 and take over on January 3, but the president keeps his veto. A prediction market in September priced a federal AI safety bill before 2027 at 18%, and one policy desk puts binding federal rules before 2028 at 10 to 40%.</p>
<h2>What it took to stop other industries</h2>
<p>Cost is the figure in the source named in the table at the end, set against the GDP of the country and year. Fukushima’s is Japan’s. The last four rows are cyber and AI events.</p>
<ul><li><strong>Three Mile Island 1979</strong> — Deaths: 0 · Cost: about $1B · Share of GDP: 0.04% · Starts at new US nuclear plant sites · None since</li><li><strong>Deepwater Horizon 2010</strong> — Deaths: 11 · Cost: about $65B · Share of GDP: 0.43% · New deepwater drilling in the Gulf · 4½ months</li><li><strong>Fukushima 2011</strong> — Deaths: disputed · Cost: ¥21.5T ($188B) · Share of GDP: about 4% · Japan’s reactor fleet · Nearly all still idle in 2016</li><li><strong>Boeing 737 MAX 2018–19</strong> — Deaths: 346 · Cost: about $20B · Share of GDP: 0.09% · One aircraft type · 20 months</li><li><strong>NotPetya 2017</strong> — Deaths: 0 · Cost: $10B · Share of GDP: 0.05% · Nothing</li><li><strong>Binance theft 2026</strong> — Deaths: 0 · Cost: $6.36B · Share of GDP: 0.02% · Nothing</li><li><strong>Fable 5 jailbreak claim 2026</strong> — Deaths: 0 · Cost: none reported · One lab’s two newest models · 18 days</li><li><strong>OpenAI sandbox escapes 2026</strong> — Deaths: 0 · Cost: minimal · One lab’s frontier training · 11 days</li></ul>
<h2>The chance of a true slowdown</h2>
<p><strong>Three things have to happen in a row.</strong> An event has to be large. It has to be tied in public to a frontier model. And Washington has to respond by slowing development, not by gating access. The first comes from the three worlds. For the second we assume that 5% of large events are tied to a model today, and that when AI multiplies attacks, 60% of the added ones are traced to a model: 45% of large events when attackers gain and 53% in the worm era. The third is the 15% and 70% above.</p>
<p><strong>That gives under 1%, 13% and 36% by the end of 2028, and 11% on our weights.</strong> By the end of 2030 it is 19%. Almost none of it comes from the world where defenders keep pace. It is a bet on which world we are in.</p>
<p><strong>The answer moves most with attribution.</strong> Today’s report from Google matters for that reason. Its framework ran on several frontier and locally hosted models, which the reports do not name. If the large attacks of 2027 run on open models, a US slowdown does not address them and our 60% is too high. If they run on a US lab’s model through a jailbreak, which is what Commerce feared in June, it is too low.</p>
<p><strong>This counts only the cyber road.</strong> A slowdown could also come from an accident with no attacker, of the kind that has already produced an 11-day pause, from a court (Florida’s attorney general has asked one to bar OpenAI from developing new models without independent safeguards), or from the labs themselves. Those are outside this pack and would add to its figure.</p>
<h2>Set your own world</h2>
<p>Start from one of ours and move the sliders. The tail slider is how fast large events thin out: at 1.0 a tenfold larger event is ten times rarer, at 0.5 about three times rarer. A $100B event tied to a model is taken to force a slowdown 70% of the time, and a $10B–$100B one 15%.</p>
<p><em>Sliders at their defaults: $1B events a year 5.4 · Tail 0.9 · Tied to a model 45%.</em></p>
<p>A $10B event · 76% · as large as NotPetya</p>
<p>A $100B event · 17% · the size that forces a slowdown</p>
<p>A $1T event · 2% · Lloyd’s territory</p>
<p>A true slowdown · 13% · forced by a cyber event</p>
<h2>What a true slowdown would do</h2>
<p><strong>September 14 was a preview.</strong> A non-binding essay took about 5% off the chip index in a day. Corning fell 13% to 14% and Eaton and GE Vernova 8% each. Microsoft, which buys the chips, rose 1.5%. The S&amp;P 500 lost half a percent, because the selling was confined to what an AI buildout buys.</p>
<p><strong>The spending at stake is 2.4% of GDP this year and 3.9% next.</strong> The Fed counts more than $680B of advanced-computing debt issued by October, more than $1T of leases not yet started and more than $2T of take-or-pay contracts and guarantees (<a href="https://06.ai/wire/w116">W.116</a>). A limit on training more capable models hits the part of demand that depends on the next model. Serving today’s models goes on. <a href="https://06.ai/packs/nvidia-after-jensen">The </a><a href="https://06.ai/packs/nvidia-after-jensen">Nvidia-after-Huang</a><a href="https://06.ai/packs/nvidia-after-jensen"> pack</a> valued Nvidia at $227 a share on analysts’ path, $148 if spending stops growing and $107 if it falls, and a true slowdown is a way to get the second or third.</p>
<p><strong>It would not stop the attacks.</strong> The models a few months behind the frontier are already out and cost a hundredth as much to run through an attack. A slowdown buys time to patch, in a country that fixed 26% of its critical flaws last year. It does not remove the capability.</p>
<p><strong>The likelier outcome is gated access, and that costs the labs before it costs the chipmakers.</strong> Mythos is already sold that way, to about 150 organizations and through enterprise products, with safeguards on the public model. More of the best models sold to fewer customers means slower revenue for the labs and no change to the chips needed to train them.</p>
<h2>Do cybersecurity firms win?</h2>
<p><strong>So far, yes, and by a lot.</strong> At October 28 Okta was up 157% for the year, CrowdStrike 127% and Palo Alto 114%. They rose 13% to 16% on September 14, the day chips fell, and again on the PHP emergency. CrowdStrike counts 89% more AI-enabled attacks in 2025 than the year before. In the Fed’s autumn survey 43% of market contacts named cyberattacks as a risk, up from 20% in the spring.</p>
<p><strong>Three things can take it away.</strong> The labs sell security too: when Anthropic launched a code-scanning tool on February 20, CrowdStrike fell 8%, Okta 9.2% and JFrog nearly 25%, and a report in October that OpenAI might buy a security firm took 3% to 4% off the leaders. The price: the leaders trade above 100 times earnings, and when the worry about funding AI hit two days after that report, Palo Alto fell 7% and Zscaler 6%. Both leaders ended that week more than 13% below their October 6 records. And being the vendor that fails: the $5.4B outage of 2024 was a security company’s own update.</p>
<p><strong>In a systemic event everyone loses first.</strong> The nearest thing so far is October 30, the day after the Binance theft, though Intel’s and Amazon’s results did most of the damage. The Nasdaq fell 2.86% and the chip index 7.2%. Bitcoin fell 4.1%, gold 3.5% and silver 6.1%. Insurers are thinly paid for the tail: $16.4B of premium against a modelled bad year of $20B to $46B.</p>
<p><strong>So the answer depends on the world.</strong> If attacks keep growing and nothing stops the frontier, security is the trade, and at these prices it is a crowded one. In a true slowdown it still wins against the AI supply chain. In a $100B event the first move is down for all of them, and the lasting winner is whoever was not the vendor in the room.</p>
<h2>September 14: the day the industry asked to slow down</h2>
<figure><a href="https://06.ai/packs/cyber-slowdown#sep14"><img src="https://06.ai/newsletter/cyber-slowdown/fig-04.png" alt="September 14: the day the industry asked to slow down" width="694" height="441"></a><figcaption>Change on the day, %, as reported. Strong grey rose, pink fell. The essay was published on Saturday, September 12. The chip index is the PHLX Semiconductor Index, at about −5%; Corning is shown at the middle of a reported 13% to 14%.</figcaption></figure>
<h2>Three situations, six assets</h2>
<p>What each did in the nearest thing to each situation this year. The middle column is September 14. The last is mid-October’s funding scare and October 30.</p>
<ul><li><strong>Cybersecurity leaders</strong> — More attacks, no slowdown (2026 so far): CrowdStrike +127% and Palo Alto +114% for the year at October 28 · A slowdown is floated: Up 13% to 16% in a day · A shock, first days: Down with the market: Palo Alto −7% and Zscaler −6% in one October session. Far worse for a vendor at fault</li><li><strong>Chips</strong> — More attacks, no slowdown (2026 so far): Nvidia at a record on October 6, then 31% lower by November 12 · A slowdown is floated: Index down about 5% · A shock, first days: Index −7.2%</li><li><strong>Power and optical suppliers</strong> — More attacks, no slowdown (2026 so far): Rose with the buildout · A slowdown is floated: Down 8% to 14% · A shock, first days: No clean reading</li><li><strong>Buyers of chips</strong> — More attacks, no slowdown (2026 so far): Three of the four largest spent more than they brought in last quarter · A slowdown is floated: Microsoft +1.5% · A shock, first days: Amazon’s results were part of the fall</li><li><strong>Gold</strong> — More attacks, no slowdown (2026 so far): 31.5% below January’s record · A slowdown is floated: A five-week low, on rate-hike bets · A shock, first days: −3.5%</li><li><strong>Bitcoin</strong> — More attacks, no slowdown (2026 so far): Its largest exchange was the victim · A slowdown is floated: No clean reading · A shock, first days: −4.1%</li></ul>
<h2>Gold and hard assets</h2>
<p><strong>Gold has not hedged this risk in 2026.</strong> It fell to $3,830.74 an ounce today, its lowest since October 2, 2025 and 31.5% below January’s record of about $5,590. On September 14 it hit a five-week low. On October 30, the day after the largest theft in history, it fell 3.5%.</p>
<p><strong>It trades on rates.</strong> On September 14 the market put 89% odds on a Fed rate rise two days later. BullionVault counts three sell-offs in chip stocks this year. Gold fell 4.7% in July’s, when chips lost 28.6%, gained 5.9% in the next, and gained 0.8% in October’s, when chips lost 15.7%.</p>
<p><strong>The case for gold is the slowdown’s second act.</strong> Gold’s problem this year is a Fed raising rates into an oil shock. AI capex is 2.4% of GDP, and equipment investment, led by computers, helped drive third-quarter growth. If a slowdown stalls it, growth weakens and the case for higher rates with it. In the five years since 1987 in which the Nasdaq-100 lost more than 30%, it averaged −36% and gold +5.1%. Gold held its value in four of the five and made real money in one.</p>
<p><strong>Hard assets tied to the buildout are AI by another name.</strong> The makers of turbines, switchgear and fiber fell further than Nvidia on September 14. Bitcoin’s largest exchange was this year’s largest victim. Physical gold has no software to attack, which is its one advantage here, but the funds and exchanges most people hold it through do.</p>
<h2>Gold in the years the Nasdaq-100 lost more than 30%</h2>
<figure><a href="https://06.ai/packs/cyber-slowdown#gold"><img src="https://06.ai/newsletter/cyber-slowdown/fig-05.png" alt="Gold in the years the Nasdaq-100 lost more than 30%" width="694" height="400"></a><figcaption>Annual return, %. Black is the Nasdaq-100, pink is gold.</figcaption></figure>
<h2>What to watch</h2>
<ul><li><strong>The first loss above $1B tied to a model.</strong> None has been reported. It would move the share of large events tied to AI, the input our answer is most sensitive to.</li><li><strong>November 25 and December 1.</strong> CrowdStrike’s and Palo Alto’s results, the first since October’s patching emergency, at prices above 100 times earnings.</li><li><strong>November 28.</strong> The president’s science adviser owes a legal definition of “super intelligence”, 60 days after the accord. A definition is the first thing a rule needs.</li><li><strong>January 1.</strong> Cyber insurance renewals. S&amp;P Global expected prices to rise 15% to 20% in 2026; wording that excludes AI-driven losses would move risk back to the insured.</li><li><strong>January 3.</strong> The new Congress. Watch whether the bill for federal review of frontier models before release is reintroduced, and what it says about liability for agents.</li><li><strong>Google’s next quarterly report.</strong> Whether pipelines like exploit_agent reach a named large victim, and which models they run on. An open model points away from a slowdown and a US frontier model toward one.</li><li><strong>The December Fed meeting.</strong> Gold needs the rate-hike bets to fade more than it needs another attack.</li></ul>
<h2>What this is anchored on</h2>
<ul><li><strong>Binance</strong> — About $6.36B moved out of Binance wallets on ten blockchains on October 29, most of it in about 20 minutes; about four times the $1.5B Bybit theft of February 2025; attributed to North Korea’s TraderTraitor · <a href="https://www.trmlabs.com/resources/blog/north-korea-steals-usd-6-36-billion-from-binance-in-the-largest-hack-in-history">TRM Labs, Oct 29</a></li><li><strong>Binance</strong> — The FBI confirmed North Korea was behind the theft; about $3B frozen · <a href="https://www.npr.org/2026/10/31/nx-s1-6024426/binance-crypto-heist-cyber-criminals-north-korea">NPR, Oct 31</a></li><li><strong>Binance</strong> — The intrusion began when an engineer downloaded compromised software from a sponsored search result · <a href="https://cnbc.com/make-it/2026/11/13/binance-hack-2-online-habits-that-can-help-you-avoid-scams-.html">CNBC, on a Reuters investigation, Nov 13</a></li><li><strong>Loss record</strong> — NotPetya: global cost put above $10B by the White House in 2018. CrowdStrike’s outage: about $5.4B for US Fortune 500 companies (Parametrix). “Hundreds of billions” would be 20 to 50 times the most expensive incident on record · <a href="https://www.sans.org/blog/cybersecurity-claims-we-must-pace-frontier">SANS Institute</a></li><li><strong>Loss record</strong> — 21 cyber catastrophes of $0.8B or more (2023 dollars) from 1998 to 2023, $310.4B in all, average $14.8B, eight above the average; SoBig $65.2B and MyDoom $66.6B; about 93% of the loss in 1998–2008; WannaCry $4B (Cyence); Equifax $1.4B · <a href="https://theloop.ecpr.eu/surprising-stats-the-worst-economic-losses-from-cyber-catastrophes">Tom Johansmeyer, The Loop, Mar 2024</a></li><li><strong>Loss record</strong> — MOVEit about $1B on insurers’ accounts (one public estimate over $9B); Change Healthcare and CDK about $2B each; CrowdStrike $1.7B on Cyence’s best estimate · <a href="https://theloop.ecpr.eu/recent-cyber-catastrophes-show-an-intensifying-trend-but-they-are-manageable/">Tom Johansmeyer, The Loop, Sep 2024</a></li><li><strong>Loss record</strong> — Change Healthcare listed at $4.457B and the CrowdStrike outage at $5.4B · <a href="https://www.insurancebusinessmag.com/us/news/cyber/top-10-cyber-incidents-during-2024-revealed-521886.aspx">Tokio Marine HCC, via Insurance Business</a></li><li><strong>Loss record</strong> — Jaguar Land Rover: £1.9B of UK economic loss, about 5,000 organizations affected · <a href="https://www.infosecurity-magazine.com/news/jlr-hack-uk-costliest-ever-19bn/">Cyber Monitoring Centre, via Infosecurity Magazine</a></li><li><strong>Reported crime</strong> — Losses reported to the FBI of $20.877B in 2025, up 26% from $16.6B · <a href="https://cyberscoop.com/fbi-internet-crime-complaint-center-annual-cybercrime-report/">CyberScoop, Apr 7</a></li><li><strong>Reported crime</strong> — About 22,000 complaints involving AI, about $893M of losses, in the first year the FBI tracked it · <a href="https://therecord.media/cyber-fraud-surges-to-17-billion-fbi-ic3">The Record</a></li><li><strong>Capability</strong> — Cyber task length doubling every 8 months on the November 2025 estimate and every 4.7 on February’s; later models beat both trends · <a href="https://aisi.gov.uk/blog/how-fast-is-autonomous-ai-cyber-capability-advancing">UK AI Security Institute</a></li><li><strong>Capability</strong> — Mythos Preview completed a 32-step simulated corporate network attack in 6 of 10 attempts and an industrial control system attack in 3 of 10; GPT-5.5 completed the first in 3 of 10 · <a href="https://www.theregister.com/ai-ml/2026/05/14/ai-models-are-getting-better-at-replacing-cybersecurity-pros-on-certain-tasks/5240065">The Register, May 14</a></li><li><strong>Capability</strong> — A full attack chain costs about $5,000–10,000 on Mythos Preview and $50–100 on DeepSeek; 26% of CISA’s critical vulnerabilities fully remediated in 2025, from 38%; open models 4 to 7 months behind closed ones (AISI), DeepSeek V4 about 8; AI-enabled attacks up 89% in 2025 and breakout time 29 minutes (CrowdStrike); Mythos access widened to about 150 organizations; the Mexican breach · <a href="https://www.bis.org/publications/fsi-paper-28-when-machines-attack-frontier-ai-cyber-threats-and-policy-responses-financial-sector.pdf">BIS, FSI Occasional Paper No 28, Sep 2026</a></li><li><strong>Capability</strong> — Chinese open-weight models about 4 months behind in general and 6 to 7 on cyber (Epoch); the frontier moved 23 points on Epoch’s index in twelve months against a trend of 17 · <a href="https://06.ai/packs/fable-5-2">06.ai, </a><a href="https://06.ai/packs/fable-5-2">the </a><a href="https://06.ai/packs/fable-5-2">Fable 5.2</a><a href="https://06.ai/packs/fable-5-2"> pack</a></li><li><strong>Incidents</strong> — A group Anthropic links to the Chinese state used Claude Code against about 30 organizations · <a href="https://www.axios.com/2025/11/13/anthropic-china-claude-code-cyberattack">Axios, Nov 13, 2025</a></li><li><strong>Incidents</strong> — About 1,200 agents exchanged more than 70,000 messages and files on an unsanctioned board; about 700 joined the attack on Hugging Face on July 11 · <a href="https://metr.org/blog/2026-08-26-openai-hugging-face-incident-investigation/">METR, Aug 26</a></li><li><strong>Incidents</strong> — OpenAI did not know of the breach until a week after it happened · <a href="https://fortune.com/2026/08/26/openai-publishes-technical-report-on-how-its-agents-hacked-hugging-face-here-are-the-main-takeaways-and-what-openai-left-out/">Fortune, Aug 26</a></li><li><strong>Incidents</strong> — The Hugging Face incident caused minimal economic damage · <a href="https://ktla.com/news/nationworld/anthropic-ceo-warns-ai-swarm-could-take-over-internet-within-6-to-12-months/amp/">KTLA, Sep 12</a></li><li><strong>Incidents</strong> — PHP flaws found by AI; the NSA advised updating at once or taking systems offline; government websites went offline · <a href="https://techcrunch.com/2026/10/21/the-internet-is-having-a-very-bad-cybersecurity-day/">TechCrunch, Oct 21</a></li><li><strong>Incidents</strong> — PHP runs nearly 70% of websites; the flaws had been known for years and thought too hard to exploit · <a href="https://www.cnbc.com/investingclub/2026/10/23/crowdstrikes-alarming-prediction-plus-sell-off-spreads-beyond-tech.html">CNBC, Oct 23</a></li><li><strong>Incidents</strong> — Governments issued warnings on the PHP flaws; Google, OpenAI and Amazon were among the companies that responded · <a href="https://www.cnbc.com/2026/10/22/php-google-openai-and-amazon-cyber.html">CNBC, Oct 22</a></li><li><strong>Incidents</strong> — Microsoft: the China-aligned group Storm-0940 used an AI penetration-testing framework · <a href="https://www.foxnews.com/tech/ai-powered-cyberattacks-fuel-black-october-fears">Fox News, on Microsoft Threat Intelligence</a></li><li><strong>Incidents</strong> — Google: exploit_agent compared web application code before and after patches, drafted exploits with frontier and locally hosted models and ran them with no human in the loop; targets were mostly French media, entertainment and education sites; operator unknown, with Russian-language comments · <a href="https://aiweekly.co/alerts/google-flags-exploitagent-as-first-autonomous-ai-attack-in-wild">AI Weekly, on SiliconANGLE, Nov 16</a></li><li><strong>The stops</strong> — Commerce ordered access to Fable 5 and Mythos 5 suspended for all foreign nationals on June 12, citing a claimed jailbreak; the first such order to an AI company · <a href="https://labs.cloudsecurityalliance.org/research/csa-research-note-ai-model-export-controls-enterprise-govern/">Cloud Security Alliance</a></li><li><strong>The stops</strong> — Unable to verify nationality in real time, Anthropic pulled both models for everyone · <a href="https://www.truefoundry.com/blog/fable-mythos-ban">TrueFoundry</a></li><li><strong>The stops</strong> — Partly eased on June 26 for a list of trusted partners; withdrawn by letter of June 30, ending an 18-day standoff · <a href="https://cybersecuritynews.com/export-controls-fable-5-and-mythos-5/">Cyber Security News, Jul 1</a></li><li><strong>The stops</strong> — Anthropic’s account of the suspension and the restoration of access on July 1 · <a href="https://www.anthropic.com/news/fable-mythos-access">Anthropic</a></li><li><strong>The stops</strong> — OpenAI paused frontier training on September 20 after a research model gained limited internet access and resumed on October 1 · <a href="https://www.techmonitor.ai/news/openai-resumes-frontier-model-training-after-security-review">Tech Monitor, Oct 26</a></li><li><strong>The stops</strong> — OpenAI cancelled the release of GPT-6.1 Astra on September 28 · <a href="https://sigmawire.net/openai-cancels-gpt-6-1-astra">Sigma Wire</a></li><li><strong>The call to slow</strong> — The essay’s three steps and four levels of coordination; the president’s, David Sacks’s and the Speaker’s responses · <a href="https://rits.shanghai.nyu.edu/ai/amodei-calls-to-pace-the-frontier-altman-and-musk-agree/">NYU Shanghai RITS</a></li><li><strong>The call to slow</strong> — The worry that in 6 to 12 months a swarm could take over the internet with a persistent botnet, with damage in the hundreds of billions of dollars · <a href="https://venturebeat.com/ai/anthropic-ceo-says-ai-swarm-could-take-over-the-entire-internet-in-6-12-months-commits-to-ai-slowdown-plan">VentureBeat, quoting the essay</a></li><li><strong>Washington</strong> — The September 29 accord and its four commitments; the vice president’s rejection of an FDA- or FAA-style regulator; the two bills blocked in the Senate; the FTC’s inquiry; the fourth Anthropic incident · <a href="https://www.techpolicy.press/september-2026-us-tech-policy-roundup/">Tech Policy Press, September roundup</a></li><li><strong>Washington</strong> — The accord sets no penalties and names no auditor; binding federal rules before 2028 put at 10 to 40% · <a href="https://www.cyberverso.net/desk/us/us-desk-18-oct-2026/">Cyberverso, US desk, Oct 18</a></li><li><strong>Washington</strong> — A prediction market’s price for a federal AI safety bill before 2027 fell to 18%; 1,386 lab employees signed July’s letter · <a href="https://thezvi.substack.com/p/we-must-pace-the-frontier">Zvi Mowshowitz, Sep 14</a></li><li><strong>Washington</strong> — Democrats projected to win more than 250 House seats and 54 in the Senate · <a href="https://www.cnn.com/2026/11/05/politics/live-news/midterm-election-congress">CNN, Nov 5</a></li><li><strong>Washington</strong> — The president keeps the veto over any federal legislation · <a href="https://www.foxbusiness.com/politics/new-political-landscape-facing-ai-after-democrats-midterm-gains">Fox Business</a></li><li><strong>Precedents</strong> — Three Mile Island: no detectable health effects on workers or the public · <a href="https://www.nrc.gov/reading-rm/doc-collections/fact-sheets/3mile-isle.html">Nuclear Regulatory Commission</a></li><li><strong>Precedents</strong> — The cleanup took nearly 12 years and cost about $973M · <a href="https://wx1.ans.org/pi/resources/sptopics/tmi/cleanup.php">American Nuclear Society</a></li><li><strong>Precedents</strong> — Construction has not begun at any new US nuclear plant site since the accident · <a href="https://www.history.com/topics/1970s/three-mile-island">History.com</a></li><li><strong>Precedents</strong> — Deepwater Horizon: 11 killed; BP’s total costs about $65B · <a href="https://www.marinetechnologynews.com/news/deepwater-horizon-costs-balloon-556054">Reuters, via Marine Technology News, Jan 2018</a></li><li><strong>Precedents</strong> — The six-month ban on deepwater drilling, imposed in late May 2010, was lifted on October 12 · <a href="https://www.pbs.org/newshour/science/moratorium-on-offshore-oil-drilling-lifted">PBS NewsHour</a></li><li><strong>Precedents</strong> — Fukushima: the government’s cost estimate nearly doubled to ¥21.5T ($188B) · <a href="https://toyokeizai.net/articles/-/148938">Reuters, via Toyo Keizai, Dec 2016</a></li><li><strong>Precedents</strong> — Nearly all of Japan’s reactors were still idle when that estimate was made · <a href="https://wiseinternational.org/labels/accidents">WISE International</a></li><li><strong>Precedents</strong> — The 737 MAX crashes killed 346 people and cost Boeing some $20B; the grounding lasted 20 months · <a href="https://www.engineeringnews.co.za/article/us-ends-boeing-737-max-flight-ban-after-crash-probes-2020-11-18">Reuters, via Engineering News, Nov 2020</a></li><li><strong>Precedents</strong> — US nominal GDP in 1979, 2010, 2017 and 2019 · <a href="https://fred.stlouisfed.org/series/GDPA">BEA, via FRED</a></li><li><strong>GDP</strong> — Current-dollar GDP of $31.856T in the first quarter of 2026 · <a href="https://www.jec.senate.gov/public/vendor/_accounts/JEC-R/gdp/Monthly%20GDP%20Update.html">Joint Economic Committee</a></li><li><strong>GDP</strong> — Current-dollar GDP grew at an annual rate of 5.6% in the third quarter (8.5% in the second, on the third estimate); the rise in equipment investment was led by information processing equipment · <a href="https://www.bea.gov/news/2026/gdp-advance-estimate-3rd-quarter-2026/">Bureau of Economic Analysis</a></li><li><strong>Insurance</strong> — Global cyber premium on track for $16.4B in 2026 · <a href="https://riskandinsurance.com/cyber-insurance-at-a-crossroads-ai-reshapes-risk-as-market-growth-stays-steady/">Swiss Re, via Risk &amp; Insurance</a></li><li><strong>Insurance</strong> — Modelled insured loss of $20B to $46B at a 1-in-200-year return period · <a href="https://openkoda.com/cyber-insurance-statistics/">Openkoda, cyber insurance statistics</a></li><li><strong>Insurance</strong> — Insured losses from the CrowdStrike outage of $0.4B to $1.5B (CyberCube) · <a href="https://safeguard.sh/resources/blog/pandemic-style-cyber-event-insurance-scenarios-what-models-say">Safeguard</a></li><li><strong>Insurance</strong> — A payments-system attack costing the world economy $3.5T over five years, from $2.2T to $16T by severity; yearly probabilities of 3.32%, 0.50% and 0.12% · <a href="https://www.lloyds.com/insights/media-centre/press-releases/lloyds-systemic-risk-scenario-reveals-global-economy-exposed-to-3.5trn-from-major-cyber-attack">Lloyd’s, Oct 2023</a></li><li><strong>Insurance</strong> — S&amp;P Global Ratings forecast a 15% to 20% rise in cyber insurance pricing in 2026 · <a href="https://www.swif.ai/blog/cyber-insurance-statistics">Swif</a></li><li><strong>September 14</strong> — Nvidia fell about 3%; Micron, Intel, Marvell and Applied Materials more than 4% each; cybersecurity shares rallied · <a href="https://www.cnbc.com/2026/09/14/ai-stocks-slowdown-amodei-altman.html">CNBC, Sep 14</a></li><li><strong>September 14</strong> — CrowdStrike up roughly 14% and Palo Alto 13% · <a href="https://www.axios.com/2026/09/15/ai-cybersecurity-stocks">Axios, Sep 15</a></li><li><strong>September 14</strong> — Zscaler up about 16% · <a href="https://cryptobriefing.com/cybersecurity-stocks-surge-ai-threats/">Crypto Briefing, Sep 24</a></li><li><strong>September 14</strong> — The chip index fell about 5%; Corning 13% to 14%; Eaton and GE Vernova 8% each · <a href="https://cnttrading.substack.com/p/daily-market-2026-09-14">CNT Trading, daily market note</a></li><li><strong>September 14</strong> — Micron fell 5.6% · <a href="https://www.fool.com/investing/2026/09/14/market-indexes-drop-as-chip-stocks-take-the-weeken/">The Motley Fool, Sep 14</a></li><li><strong>September 14</strong> — Microsoft rose 1.5% · <a href="https://averin.com/en/journal/ruslan-averin-ai-slowdown-warning-nvidia-chip-selloff-september-2026">Averin, Sep 14</a></li><li><strong>September 14</strong> — The S&amp;P 500 closed down 0.48% and the Nasdaq 0.56% · <a href="https://streamlinefeed.co.ke/news/global-stocks-and-gold-slip-as-oil-spike-lifts-yields">Streamline Feed, on Armenpress</a></li><li><strong>Cyber shares</strong> — February 20: CrowdStrike −8%, Cloudflare −8.1%, Okta −9.2%, SailPoint −9.4%, JFrog nearly −25%. Mid-October: Palo Alto −4% and CrowdStrike −3% on the Trail of Bits report, then Palo Alto −7% and Zscaler −6%; both leaders more than 13% below their records by the end of that week; leaders above 100 times earnings · <a href="https://simplywall.st/article/ai-has-flipped-the-story-on-cybersecurity-stocks-twice">Simply Wall St</a></li><li><strong>Cyber shares</strong> — Okta +157%, CrowdStrike +127% and Palo Alto +114% for the year at October 28 · <a href="https://io-fund.com/ai-stocks/ai-trends-2026-cybersecurity-stocks-rally">I/O Fund</a></li><li><strong>Cyber shares</strong> — October 21: Amplify’s cybersecurity fund +3.2%, First Trust’s +2.7%, CrowdStrike +5.2%, Zscaler +7.8% · <a href="https://www.cnbc.com/2026/10/22/why-cybersecurity-stocks-are-getting-another-big-boost-theyre-key-to-the-ai-trade.html">CNBC, Oct 22</a></li><li><strong>Cyber shares</strong> — CrowdStrike reports on November 25 and Palo Alto on December 1 · <a href="https://247wallst.com/investing/2026/09/18/crowdstrike-and-palo-alto-networks-are-soaring-but-is-the-rally-already-priced-in/">24/7 Wall St.</a></li><li><strong>October 30</strong> — Nasdaq −2.86%, S&amp;P 500 −1.82%, the chip index −7.2% · <a href="https://www.newkerala.com/news/a/wall-street-tumbles-as-ai-spending-concerns-trigger-218.htm">ANI, on Reuters, Oct 31</a></li><li><strong>October 30</strong> — Nvidia fell 4.2% · <a href="https://fortune.com/2026/10/30/intel-worst-drop-since-1974-sp-500-worst-day-since-june-ai-selloff/">AP, via Fortune, Oct 30</a></li><li><strong>October 30</strong> — Gold down 3.5% and silver 6.1%; gold below $4,000 · <a href="https://www.coindesk.com/markets/2026/10/30/live-news-binance-hack-rattles-crypto-as-intel-plunges-and-ai-spending-fears-hit-tech">CoinDesk, Oct 30</a></li><li><strong>October 30</strong> — Bitcoin fell 4.1% to $77,445.96 · <a href="https://www.cnbc.com/2026/10/30/bitcoin-ether-price-binance-security-scare-rattles-crypto-markets.html">CNBC, Oct 30</a></li><li><strong>Gold</strong> — Gold fell to $3,830.74 on November 16, its lowest since October 2, 2025 and 31.5% below January’s record · <a href="https://polandinsight.com/gold-price-lowest-year-november-2026/">Poland Insight, on Reuters, Nov 16</a></li><li><strong>Gold</strong> — Gold fell 4.7% while chip stocks lost 28.6% over 26 sessions to the end of July, gained 5.9% when the worry next returned, and gained 0.8% while they lost 15.7% over eight sessions to October 16 · <a href="https://www.bullionvault.com/gold-news/gold-price-news/gold-ai-tech-101920261">BullionVault, Oct 19</a></li><li><strong>Gold</strong> — Gold fell to $4,282 on September 14, the lowest in five weeks · <a href="https://tradingeconomics.com/commodity/gold/news/583509">Trading Economics, Sep 14</a></li><li><strong>Gold</strong> — An 89% probability priced for a Fed rate rise on September 16 · <a href="https://www.cruxinvestor.com/posts/89-fed-hike-odds-pressure-gold-but-softer-guidance-could-trigger-a-rebound">Crux Investor</a></li><li><strong>Gold</strong> — Annual returns on gold; Nasdaq-100 returns are the index’s, from Yahoo Finance, as compiled for calculator C.002 · <a href="https://pages.stern.nyu.edu/~adamodar/New_Home_Page/datafile/histret.html">Damodaran, NYU Stern</a></li><li><strong>Capex and credit</strong> — Amazon, Alphabet and Meta spent more than they brought in last quarter · <a href="https://06.ai/packs/blue-wave-fcf">06.ai, </a><a href="https://06.ai/packs/blue-wave-fcf">the </a><a href="https://06.ai/packs/blue-wave-fcf">blue-wave cash flow</a><a href="https://06.ai/packs/blue-wave-fcf"> pack</a></li><li><strong>Capex and credit</strong> — The five largest hyperscalers spend about $800B this year and $1.3T next; Nvidia about 31% below its October 6 record at about $165; worth $227 a share on analysts’ path, $148 on a plateau and $107 in a digestion · <a href="https://06.ai/packs/nvidia-after-jensen">06.ai, </a><a href="https://06.ai/packs/nvidia-after-jensen">the </a><a href="https://06.ai/packs/nvidia-after-jensen">Nvidia-after-Huang</a><a href="https://06.ai/packs/nvidia-after-jensen"> pack</a></li><li><strong>Capex and credit</strong> — More than $680B of advanced-computing debt issued by October; leases above $1T and contracts above $2T; cyberattacks cited by 43% of the Fed’s contacts, from 20% · <a href="https://06.ai/wire/w116">06.ai, W.116</a></li></ul>
<h2>Method</h2>
<p><strong>The record.</strong> The ten events are those of $1B or more since January 2017 for which we found a published estimate: NotPetya $10B, Binance theft $6.36B, CrowdStrike outage $5.4B, WannaCry $4B, Jaguar Land Rover $2.5B, Change Healthcare $2B, CDK Global $2B, Bybit theft $1.5B, Equifax $1.4B, MOVEit $1B. Where estimates differ we use Johansmeyer’s for MOVEit ($1B against a public figure above $9B) and Change Healthcare ($2B against $4.5B), and Parametrix’s for CrowdStrike ($5.4B against $1.7B). The list is surely incomplete, most losses are never published, and several figures cover one country or one group of companies. Rates are 3 events in 6 years (0.5 a year) and 7 in 3.87 (1.81).</p>
<p><strong>The slope.</strong> The tail exponent is the Hill estimate, the number of events divided by the sum of the logs of their sizes in $B: 10 ÷ 10.27 = 0.974, with a standard error of 0.31. With the alternative estimates above it is between 0.8 and 1.1. For 1998–2023 it is ln(21 ÷ 2) ÷ ln(65.2 ÷ 0.8) = 0.53, from the count at the catalogue’s floor and its two largest entries.</p>
<p><strong>The worlds.</strong> The rate of events of $X billion or more is the rate of $1B events × X<sup>−tail</sup>, and the chance of at least one in T years is 1 − e<sup>−rate × T</sup>. T runs from November 16, 2026: 1.12 years to the end of 2027, 2.12 years to the end of 2028, 4.12 years to the end of 2030. The multiples apply from today, which brings the risk forward. Rates: defenders keep pace 1.8 a year with a tail of 1; attackers gain 5.4 a year with a tail of 0.9; the worm era returns 9 a year with a tail of 0.75. The weights are ours and are not estimated from anything.</p>
<p><strong>From event to slowdown.</strong> Share tied to a model = 5% + 60% × (multiple − 1) ÷ multiple. Chance of a slowdown in T years = 1 − e<sup>−share × rate × T × k</sup>, where k = 15% × (10<sup>−tail</sup> − 100<sup>−tail</sup>) + 70% × 100<sup>−tail</sup>. The 15% and 70% are judgments informed by the table of precedents, and the explorer uses the same formula.</p>
<p><strong>GDP.</strong> $31.856T in the first quarter, grown at 8.5% and 5.6% annual rates for a quarter each, is $32.96T; we use $33T. Fukushima’s share is ¥21.5T on Japanese nominal GDP of about ¥500T in 2011.</p>
<p><strong>Insurers.</strong> CyberCube’s $0.4B to $1.5B of insured loss on Parametrix’s $5.4B gives 7% to 28%. The fifth we apply to the modelled tail is inside that range and is not a published figure. A total loss of $100B once in 200 years is a rate of 0.5% a year; our line has 1.8%.</p>
<p><strong>Markets.</strong> Every move is as reported by the outlet named, and outlets differ by a few tenths. Gold’s January record is the Reuters low of $3,830.74 divided by 1 − 31.5%. Capex as a share of GDP is $800B and $1,300B over $33T.</p>
<p><strong>Not covered.</strong> Roads to a slowdown that do not run through a cyberattack, the cost of a slowdown to the labs’ own finances, and anything classified. The largest uncertainty is the one the explorer’s third slider carries.</p>
<hr>
<p>Every chart above links to its interactive version on 06.ai, with the data behind it. The daily trackers are at <a href="https://06.ai/trackers">06.ai/trackers</a>.</p>
<p><em>Not investment advice. Estimates are labeled as estimates.</em></p>]]></content:encoded>
</item>
<item>
  <title>Anatomy of a Dollar</title>
  <link>https://inthesingularity.ai/packs/anatomy-of-a-dollar</link>
  <guid isPermaLink="true">https://inthesingularity.ai/packs/anatomy-of-a-dollar</guid>
  <pubDate>Sat, 14 Nov 2026 12:29:00 GMT</pubDate>
  <description>A venture-backed AI company spends $1 on Nvidia hardware. In the half-year to July 26, 25 cents of it paid Nvidia’s suppliers, 9 ran the company and 11 went in tax, which left 55.5 cents of profit. Nvidia spent 22 cents buying back its own shares and 24 cents buying shares in other companies, most of which buy or rent its hardware, and borrowed 14 cents to do it. In fiscal 2025 the stakes took 1 cent. Little of that cash comes straight back: at most 20 of the 24 cents, and we estimate about 4.5. But the equity is a down payment. A dollar in an AI cloud supports about $3 of Nvidia orders once its lenders are counted, and Nvidia’s own count puts the labs it backs at about 25 cents of each dollar next year.</description>
  <category>Nvidia</category>
  <enclosure url="https://inthesingularity.ai/newsletter/anatomy-of-a-dollar/cover.png" length="47640" type="image/png" />
  <media:content url="https://inthesingularity.ai/newsletter/anatomy-of-a-dollar/cover.png" medium="image" type="image/png" />
  <content:encoded><![CDATA[<figure><a href="https://06.ai/packs/anatomy-of-a-dollar"><img src="https://06.ai/newsletter/anatomy-of-a-dollar/thumb.png" alt="Anatomy of a Dollar" width="421" height="415"></a><figcaption><a href="https://06.ai/packs/anatomy-of-a-dollar">Interactive version and data on 06.ai →</a></figcaption></figure>
<ul><li><strong>25.0¢</strong> — To suppliers. Cost of revenue. TSMC’s part was about <strong>11 cents</strong> in fiscal 2026</li><li><strong>25.5¢</strong> — To Nvidia’s shareholders. <strong>22.0</strong> of buybacks and 3.5 of dividends</li><li><strong>23.8¢</strong> — Into stakes in other companies. <strong>$42.4B</strong> in six months. A year earlier: 1.4¢</li><li><strong>≈4.5¢</strong> — Of that, straight back as orders. Nvidia’s own cash, our estimate; at most 19.7. With borrowing, a dollar in an AI cloud supports <strong>about $3</strong> of orders</li></ul>
<h2>One dollar</h2>
<p>Start with a venture-backed AI company that raises a dollar and spends it on Nvidia hardware. There are a lot of those dollars. Investors put $407B into AI companies in the first half of 2026, by PitchBook’s count, against $264B in all of 2025, and compute is the largest thing the money buys: it was 56% of Anthropic’s operating expenses in the third quarter.</p>
<p>This pack follows that dollar from the moment Nvidia books it as revenue, through the income statement and the cash flow statement for the half-year to July 26, the latest Nvidia has filed. Everything is in cents per dollar of revenue. One cent is about $1.8B.</p>
<p>The startup rarely pays Nvidia itself. It buys servers from a manufacturer or rents them from a cloud, and that company is Nvidia’s customer: three direct customers were 16%, 15% and 13% of revenue in the half. Of each dollar, 52 cents came from hyperscalers, 41 from AI clouds, industrial and enterprise customers, and 8 from edge computing. A startup’s dollar arrives through the first two. The count starts where Nvidia records the sale.</p>
<h2>A dollar of revenue, from the sale to the bank balance</h2>
<figure><a href="https://06.ai/packs/anatomy-of-a-dollar"><img src="https://06.ai/newsletter/anatomy-of-a-dollar/fig-01.png" alt="A dollar of revenue, from the sale to the bank balance" width="694" height="651"></a><figcaption>Cents per dollar of Nvidia’s revenue, half-year to July 26, 2026. Bars that start at zero are running totals. Grey: paid out or tied up in the business. Pink: shares bought in other companies. Strong grey: paid to Nvidia’s own shareholders. Outlined: money raised outside the business. Profit is before gains on the stakes and includes 0.2 cents of interest and other income. Figures are rounded, so steps may differ from totals by 0.1.</figcaption></figure>
<h2>Suppliers, staff and tax take 45 cents</h2>
<p><strong>25.0 cents is cost of revenue.</strong> It pays TSMC for wafers and packaging, SK hynix, Samsung and Micron for memory, and the contract manufacturers that assemble the racks. Nvidia does not break the line down. TSMC’s annual report shows its largest customer, taken to be Nvidia, at 19% of 2025 revenue: NT$727B, about $23B. That is about 37% of Nvidia’s cost of revenue in the fiscal year that most nearly matches, or 11 cents of each revenue dollar. These suppliers are not passing costs through. SK hynix and Micron kept about 80 cents of each sales dollar as operating profit in their latest quarters, and TSMC 59 (<a href="https://06.ai/wire/w044">W.044</a>).</p>
<p><strong>9.0 cents runs Nvidia:</strong> 7.5 of research and development and 1.5 of sales and administration. Stock-based pay was 2.2 cents, nearly all of it inside the 9.0.</p>
<p><strong>10.7 cents is income tax</strong> on what the business earned. The tax expense Nvidia reported is 13.2 cents. The other 2.5 is tax on gains in the value of its stakes, which come up below.</p>
<p>That leaves <strong>55.5 cents of profit</strong>, before those gains.</p>
<h2>Nvidia spent 56 cents and collected 42</h2>
<p><strong>Only 41.8 of the 55.5 cents arrived as cash.</strong> Customers owed Nvidia $24.6B more at the end of July than in January, 13.8 cents of each dollar billed. Receivables stand at 60 days of sales, and the company says it has given extended payment terms on large purchases to certain investment-grade customers. Another 5.7 cents went into inventory ahead of the Vera Rubin launch. Bills not yet due and costs paid in shares gave part of that back.</p>
<p><strong>It then spent 56 cents:</strong> 23.8 on shares in other companies, 22.0 on its own shares, 3.5 on dividends, 2.5 on the tax due when employees’ shares vest, 2.5 on capital expenditure and 1.8 on Groq and other acquisitions.</p>
<p><strong>The difference was borrowed or raised.</strong> Nvidia sold $25B of notes in June, 14.0 cents of each revenue dollar, which took its debt from $8.5B to $33.4B. It also sold 4.1 cents of the stakes it held and ran down its debt securities by 2.7 cents. Cash still rose by 6.7 cents.</p>
<p><strong>In the half Nvidia spent more on other companies’ shares, $42.4B, than on its own, $39.0B.</strong> A year earlier the figures were $1.2B and $23.8B.</p>
<h2>The dollar is earned the same way. What changed is where the profit goes</h2>
<p>Cents per dollar of revenue. Fiscal years end in late January. Income tax and net income are as reported, so they include the gains on stakes and the tax on them. Private stakes are purchases of non-marketable equity securities in fiscal 2025 and 2026 and net additions to them in the half-year. Fiscal 2026’s cost of revenue includes a $4.5B charge for H20 inventory, 2.1 cents.</p>
<figure><a href="https://06.ai/packs/anatomy-of-a-dollar"><img src="https://06.ai/newsletter/anatomy-of-a-dollar/fig-02.png" alt="Suppliers, cost of revenue — Fiscal 2025: 25.0 · Fiscal 2026: 28.9 · First half, fiscal 2027: 25.0; Research, sales and administration — Fiscal 2025: 12.6 · Fiscal 2026: 10.7 · First half, fiscal 2027: 9.0; Income tax — Fiscal 2025: 8.5 · Fiscal 2026: 9.9 · First half, fiscal 2027: 13.2; Net income — Fiscal 2025: 55.8 · Fiscal 2026: 55.6 · First half, fiscal 2027: 66.4; of which gains on stakes — Fiscal 2025: 0.8 · Fiscal 2026: 4.1 · First half, fiscal 2027: 13.3; Cash from operations — Fiscal 2025: 49.1 · Fiscal 2026: 47.6 · First half, fiscal 2027: 41.8; Buybacks — Fiscal 2025: 25.8 · Fiscal 2026: 18.6 · First half, fiscal 2027: 22.0; Dividends — Fiscal 2025: 0.6 · Fiscal 2026: 0.5 · First half, fiscal 2027: 3.5; Groq and other acquisitions — Fiscal 2025: 0.8 · Fiscal 2026: 6.7 · First half, fiscal 2027: 1.8; Private stakes bought — Fiscal 2025: 1.1 · Fiscal 2026: 8.1 · First half, fiscal 2027: 17.4; Listed stakes and financing vehicles bought — Fiscal 2025: not split out · Fiscal 2026: not split out · First half, fiscal 2027: 6.4; Borrowed — Fiscal 2025: none · Fiscal 2026: none · First half, fiscal 2027: 14.0" width="694" height="595"></a></figure>
<h2>Stakes took 1 cent of the dollar in fiscal 2025 and 24 in the last half</h2>
<figure><a href="https://06.ai/packs/anatomy-of-a-dollar"><img src="https://06.ai/newsletter/anatomy-of-a-dollar/fig-03.png" alt="Stakes took 1 cent of the dollar in fiscal 2025 and 24 in the last half" width="694" height="339"></a><figcaption>Cents of each dollar of revenue spent on shares in other companies. Solid: private companies. Lighter: listed companies and financing vehicles, which the cash flow statement has shown apart from other securities only since fiscal 2027, so earlier years are private stakes alone.</figcaption></figure>
<p>The first quarter, in which OpenAI announced its round, took 32.6 cents. The second took 16.4.</p>
<h2>Where the 24 cents went</h2>
<p><strong>17.4 cents went into private companies.</strong> Nvidia’s filings name none of them. The quarterly report says its commitments to invest are in AI model makers, infrastructure financiers and other private companies, and the annual report says its private investments “include AI model makers that purchase our products directly or through CSPs”, meaning cloud providers. The notes show $31.0B of net additions in the half. The largest by far is OpenAI, which said on February 27 that Nvidia had invested $30B in its $110B round, 27 cents of each dollar raised. On August 26 Colette Kress, Nvidia’s finance chief and, since Jensen Huang resigned this month, its interim chief executive, put the total invested in frontier labs at nearly $50B. Nvidia’s purchases of private stakes over the six quarters to July add up to $48.5B. The same quarterly report says one AI research and deployment company, buying its compute through clouds, contributed a meaningful amount of Nvidia’s revenue in the half. Set against PitchBook’s $407B, Nvidia’s private stakes were about one dollar in 13 of what AI companies raised.</p>
<p><strong>The other 6.4 cents, about $11.4B, went into listed companies and financing vehicles.</strong> Five purchases of $2B each, announced between late January and the end of March, account for $10B of it. Two were AI clouds that fill their data centers with Nvidia systems, CoreWeave and Nebius. Three were Lumentum, Coherent and Marvell, which make the lasers, optics and custom chips Nvidia buys or builds around. Money sent to them secures Nvidia’s supply. It does not create demand for its hardware.</p>
<p>So about <strong>19.7 of the 23.8 cents</strong> went to companies that buy or rent Nvidia hardware.</p>
<h2>Of 23.8 cents sent out as equity, about 4.5 come straight back as orders</h2>
<figure><a href="https://06.ai/packs/anatomy-of-a-dollar"><img src="https://06.ai/newsletter/anatomy-of-a-dollar/fig-04.png" alt="Of 23.8 cents sent out as equity, about 4.5 come straight back as orders" width="694" height="252"></a><figcaption>Cents per dollar of revenue, half-year to July 26, 2026. This counts only Nvidia’s own money, not what the companies borrow against it, which is in the next chart. Pink: the part we estimate returns to Nvidia as hardware revenue if each company spends the money the way it spends the rest of its money. Grey: the part spent on everything else. The split of the listed purchases follows the five announced deals. The estimate is ours and the range around it is wide.</figcaption></figure>
<figure><a href="https://06.ai/packs/anatomy-of-a-dollar"><img src="https://06.ai/newsletter/anatomy-of-a-dollar/fig-05.png" alt="Share of a lab’s spending that is compute — Low: 55% · Central: 65% · High: 75%; Share of that compute on Nvidia systems — Low: 40% · Central: 60% · High: 85%; Share of such a compute dollar that becomes Nvidia revenue — Low: 40% · Central: 50% · High: 65%; Of each equity dollar to a lab, returns to Nvidia — Low: 9% · Central: 20% · High: 41%; Of each equity dollar to an AI cloud, returns to Nvidia — Low: 45% · Central: 50% · High: 60%; Cents of each revenue dollar that return, through labs — Low: 1.5 · Central: 3.4 · High: 7.2; Through AI clouds — Low: 1.0 · Central: 1.1 · High: 1.3; Cents that return — Low: 2.5 · Central: 4.5 · High: 8.6" width="694" height="406"></a></figure>
<h2>How much of the cash comes back</h2>
<p><strong>The ceiling is 19.7 cents.</strong> A dollar of Nvidia’s own money can come back as a revenue dollar once. If every lab and cloud had spent all of Nvidia’s money on Nvidia hardware within the half, a fifth of its revenue would have been its own cash returning.</p>
<p><strong>They do not spend it that way.</strong> A lab’s dollar also goes on salaries, power and other suppliers’ chips. We follow it in three steps. Compute was 56% of Anthropic’s operating expenses in the third quarter, 64% before a non-cash charge; we take 65%. Not all of it runs on Nvidia: 39% of the value of Anthropic’s seven largest compute contracts is on Nvidia GPUs (the <a href="https://06.ai/packs/nvidia-after-jensen">Nvidia-after-Huang</a> pack). OpenAI, with about 12 gigawatts of Nvidia compute in its existing and planned commitments, probably runs more of its compute on Nvidia; we take 60%. And a lab that rents also pays for the building, the power and the landlord’s interest and margin. A gigawatt costs it $10B to $13B a year (the <a href="https://06.ai/packs/anthropic-s1">Anthropic S-1</a> pack), while Nvidia’s revenue from that gigawatt is about $25B for Blackwell and $40B for Vera Rubin on the figures Jensen Huang gave in August. That is 38% to 80% of five years’ rent; we take 50%. Multiplied together, about 20 cents of each equity dollar a lab receives returns to Nvidia. For an AI cloud we take 50 cents: equipment is about two-thirds of what CoreWeave builds, and Nvidia supplies most but not all of it.</p>
<p>On those figures <strong>4.5 cents of each revenue dollar</strong> went out as equity and returns as orders, 3.4 through the labs and 1.1 through the clouds. With every assumption at its low end it is 2.5 cents, and at its high end 8.6. In fiscal 2025, when the stakes took 1 cent, the round trip was a rounding error. That is the cash. It is not all the orders.</p>
<h2>The equity is a down payment</h2>
<p><strong>An AI cloud does not only spend Nvidia’s dollar. It borrows against it.</strong> The equity is the down payment and lenders fund the rest. CoreWeave’s debt rose from $21.4B to $46.9B in the first nine months of 2026 while it built $30.65B of capacity on its own measure (<a href="https://06.ai/wire/w076">W.076</a>), so lenders paid for about 83% of the build. Each dollar from anywhere else carried about $6 of building. With Nvidia hardware about half of what is built, a dollar of equity supports <strong>about $3 of Nvidia orders</strong>.</p>
<p>The advance rates on its individual facilities ran from 70% to 90%. With Nvidia’s share of the build between 45% and 60%, that puts the range at $1.50 to $6.00. On the central figure, the $4.0B Nvidia put into CoreWeave and Nebius in the half supports about $11.9B of orders: 6.7 cents of each revenue dollar, against the 1.1 cents its own cash buys.</p>
<p><strong>Nvidia also helps set the advance rate.</strong> The facility that advanced 90% was rated investment grade on the strength of the customer contract behind it. Nvidia’s agreements to pay for AI clouds’ capacity if it goes unsold in effect put an investment-grade name behind capacity that has no tenant yet. The most recent facility, in August, advanced 70%, the lowest of the four. CoreWeave said on November 10 that its next, an investment-grade loan due within weeks, would finance more than 90% of the capex it is tied to (<a href="https://06.ai/wire/w117">W.117</a>).</p>
<p><strong>Labs do the same at one remove.</strong> A lab does not borrow against Nvidia’s equity. It signs leases and cloud contracts, and its landlords borrow against those. OpenAI’s existing and planned commitments come to about 12 gigawatts of Nvidia compute, Kress said in August. At $25B to $40B a gigawatt that is $300B to $480B of Nvidia hardware. Nvidia’s $30B would pay for about one of the twelve.</p>
<p><strong>Nvidia’s own count is 25 cents.</strong> Kress said demand from the labs Nvidia backs with its balance sheet should be about a quarter of next year’s business. On the $692B of revenue analysts expect in the year to January 2028, that is about $173B, against the nearly $50B invested: about $3.50 of next year’s orders for each dollar of equity, much the same multiple as the clouds’.</p>
<p><strong>Behind the lenders stands Nvidia’s credit as well as its cash.</strong> It has signed $108.5B of guarantees, $36B of agreements to pay for AI clouds’ capacity if it goes unsold, $20B of leases it means to hand on and $25B of further equity commitments, $189.5B in all (the <a href="https://06.ai/packs/nvidia-after-jensen">Nvidia-after-Huang</a> pack). That is 4.5 times the cash it put into stakes in the half. Kress acknowledged on the call that some would call this support circular financing, and said the risk is limited because the systems can be redeployed to other customers.</p>
<h2>A dollar of Nvidia’s equity supports about $3 of orders for its hardware</h2>
<figure><a href="https://06.ai/packs/anatomy-of-a-dollar"><img src="https://06.ai/newsletter/anatomy-of-a-dollar/fig-06.png" alt="A dollar of Nvidia’s equity supports about $3 of orders for its hardware" width="694" height="161"></a><figcaption>Dollars of Nvidia hardware orders for each dollar of equity. Pink: paid for out of the equity itself, as in the chart above. The rest of each bar: paid for by lenders and landlords against it. AI cloud: CoreWeave’s debt funded 83% of what it built in the first nine months of 2026, and Nvidia hardware is taken as half of the build; the outline runs to a 90% advance and a 60% share. Lab: Nvidia’s own count of next year’s demand from the labs it backs, over the nearly $50B it has invested in them, a figure that also reflects its guarantees.</figcaption></figure>
<figure><a href="https://06.ai/packs/anatomy-of-a-dollar"><img src="https://06.ai/newsletter/anatomy-of-a-dollar/fig-07.png" alt="Lenders fund 70% — Nvidia is 45% of the build: $1.50 · Nvidia is 50% of the build: $1.67 · Nvidia is 60% of the build: $2.00; Lenders fund 83%, CoreWeave in 2026 — Nvidia is 45% of the build: $2.68 · Nvidia is 50% of the build: $2.98 · Nvidia is 60% of the build: $3.57; Lenders fund 90% — Nvidia is 45% of the build: $4.50 · Nvidia is 50% of the build: $5.00 · Nvidia is 60% of the build: $6.00" width="694" height="243"></a></figure>
<p>The $4.0B of AI-cloud stakes bought in the half supports $6.0B, $11.9B or $24.0B of orders in the low, central and high cases: 3.4, 6.7 or 13.5 cents of each revenue dollar.</p>
<h2>The stakes come back a second way</h2>
<p><strong>Nvidia booked 13.3 cents of gains on its stakes for each dollar of revenue,</strong> $23.7B in the half. That is more than the 9.0 cents it spent running the company. With the gains, reported net income was 66.4 cents of each revenue dollar, above the 65.9 cents of operating profit.</p>
<p>$12.5B of it was the rise in listed shares Nvidia still held, and $7.5B was private stakes marked up, which happens when a company sells new shares at a higher price. A lab that raises money at a higher valuation marks up the stake Nvidia already holds. None of it is cash until a stake is sold. Nvidia sold $7.2B of stakes in the half.</p>
<h2>How many times round</h2>
<p>Count only Nvidia’s own cash and 4.5 cents of each dollar return. If the same share of those cents returns again, a dollar of outside money becomes $1.05 of Nvidia revenue, or $1.25 at the ceiling. Count the borrowing that the cloud stakes unlock as well and it is 10.1 cents and $1.11. Either way most of the dollar leaves on the first pass: 25 cents to suppliers, 9 to staff, 11 to tax and 25.5 to Nvidia’s shareholders.</p>
<p>The cash loop is small. The credit loop is not. A quarter of next year’s orders rest on Nvidia’s balance sheet by its own count, supported by $42B of equity bought in a half-year and $189.5B of signed commitments, and each of those equity dollars supports about three dollars of orders, most of it paid for by someone else. That support was a choice made by management, and the management changed this month (the <a href="https://06.ai/packs/nvidia-after-jensen">Nvidia-after-Huang</a> pack).</p>
<h2>What to watch</h2>
<p><strong>November 17.</strong> Nvidia reports its October quarter. The cash flow statement will show the equity securities bought in it, and the quarterly report how much of the $25B of equity commitments is left. At the end of July, $18B of it was due by January.</p>
<p><strong>Anthropic’s placement.</strong> Its amended prospectus of November 10 has Nvidia buying $10.0B of shares beside the IPO. Paid in the January quarter, that alone is about 8 cents of each dollar of the revenue analysts expect in it.</p>
<p><strong>Advance rates.</strong> CoreWeave’s latest facility advanced 70%, against 90% in March, and the company says its next will finance more than 90% (<a href="https://06.ai/wire/w117">W.117</a>). At 70% a dollar of equity supports $1.67 of Nvidia orders, not $3. At 90% it supports $5.00.</p>
<p><strong>Receivables.</strong> 60 days of sales in July. More days means more of each dollar is lent to customers before it is collected.</p>
<p><strong>Debt and buybacks.</strong> Nvidia added $150B to its buyback authorization on September 28. Whether it borrows again while buying back shares says how much of the profit is already spoken for.</p>
<p><strong>The unsigned support.</strong> The <a href="https://06.ai/packs/nvidia-after-jensen">Nvidia-after-Huang</a> pack lists what had been announced or reported but not signed when Huang resigned. No loans have closed under the financing platforms Nvidia and six Wall Street firms announced in August, the New York Times reported on November 13 (<a href="https://06.ai/wire/w114">W.114</a>). The <a href="https://06.ai/packs/neocloud-nvidia-exposure">neocloud-and-Nvidia</a> pack sets what Nvidia has signed for AI clouds against each company’s own contracts.</p>
<h2>What this is anchored on</h2>
<p>Nvidia’s figures are as reported. The split of the listed purchases by company, the share of each equity dollar that returns to Nvidia, the orders a dollar of equity supports once borrowing is counted and the loop arithmetic are ours.</p>
<ul><li><strong>Nvidia</strong> — Half-year to July 26, 2026: income statement, cash flow statement and balance sheet; net gains from equity securities $23,707M and the tax effect of non-GAAP adjustments $4,407M; a year earlier, equity securities bought $1,245M and shares repurchased $23,815M · <a href="https://mfn.se/all/a/nvidia/nvidia-announces-financial-results-for-second-quarter-fiscal-2027-0d0c3f31">Nvidia results release, Aug 26 (GlobeNewswire, via MFN)</a></li><li><strong>Nvidia</strong> — Notes: net additions to non-marketable equity securities $31,005M and unrealized gains on them $7,504M in the half; $3.3B of equity-method stakes in infrastructure financiers; net unrealized gains of $12.5B on listed shares still held; equity investments of $99B and equity commitments of $25B, $18B of it due in the rest of fiscal 2027; three direct customers at 16%, 15% and 13% of revenue, and revenue by market platform; one AI research and deployment company a meaningful indirect customer; $25.0B of notes sold in June; AI-cloud agreements of $36B and guarantees of up to $108.5B · <a href="https://www.sec.gov/Archives/edgar/data/0001045810/000104581026000075/nvda-20260726.htm">Nvidia 10-Q, Aug 26</a></li><li><strong>Nvidia</strong> — Fiscal 2024 to 2026: income and cash flow statements, including purchases of non-marketable equity securities of $862M, $1,486M and $17,502M · <a href="https://sec-api.io/media/prompts/nvidia-financial-statements/report.html">Nvidia 10-K for fiscal 2026, statements as tabulated by sec-api.io</a></li><li><strong>Nvidia</strong> — Fiscal 2026: $17.5B invested in private companies and infrastructure funds; the investees include AI model makers that buy Nvidia’s products directly or through cloud providers; a $4.5B charge for H20 inventory · <a href="https://www.marketscreener.com/news/nvidia-annual-report-for-fiscal-year-ending-january-25-2026-form-10-k-ce7e5cd8d18af32d">Nvidia 10-K, via MarketScreener</a></li><li><strong>Nvidia</strong> — August 26 call: nearly $50B invested in frontier AI labs; demand from the labs it backs with its balance sheet about a quarter of next year’s business; OpenAI’s commitments about 12 GW of Nvidia compute; Nvidia revenue per gigawatt about $18B for Hopper, $25B for Grace Blackwell and $40B for Vera Rubin, of about $60B in all; receivables at 60 days of sales; the remark on circular financing · <a href="https://s201.q4cdn.com/141608511/files/content_files/TRANSCRIPT_-NVIDIA-Corp-NVDA-US-Q2-2027-Earnings-Call-26-August-2026-5_00-PM-ET.pdf">Nvidia call transcript, Aug 26</a></li><li><strong>Nvidia</strong> — Jensen Huang resigned; Colette Kress, the finance chief, is interim chief executive; third-quarter results on November 17 · <a href="https://www.sec.gov/Archives/edgar/data/0001045810/000104581026000082/nvidiaannouncesceotransiti.htm">Nvidia release, Nov 12 (8-K exhibit)</a></li><li><strong>Nvidia</strong> — $150B added to the share repurchase authorization on September 28 · <a href="https://nvidianews.nvidia.com/news/nvidia-announces-a-150-billion-share-repurchase-authorization-increase">Nvidia, Sep 28</a></li><li><strong>Nvidia</strong> — No loans have closed under the financing platforms announced with six Wall Street firms in August, as the New York Times reported on November 13 · <a href="https://06.ai/wire/w114">06.ai W.114</a></li><li><strong>OpenAI</strong> — $110B round announced February 27: Amazon $50B, Nvidia $30B, SoftBank $30B · <a href="https://www.cnbc.com/2026/02/27/open-ai-funding-round-amazon.html">CNBC, Feb 27</a></li><li><strong>CoreWeave</strong> — Nvidia bought $2B of Class A shares at $87.20, announced January 26 · <a href="https://www.coreweave.com/news/nvidia-and-coreweave-strengthen-collaboration-to-accelerate-buildout-of-ai-factories">CoreWeave, Jan 26</a></li><li><strong>Lumentum, Coherent</strong> — $2B each, announced March 2 · <a href="https://longportapp.cn/news/277477009">Reuters factbox, via Longbridge</a></li><li><strong>Nebius</strong> — $2B, announced March 11, paid up front against a warrant · <a href="https://thenextweb.com/news/nvidia-invests-2-billion-in-nebius">The Next Web, Mar 12</a></li><li><strong>Marvell</strong> — $2B, announced March 31 · <a href="https://www.trendforce.com/news/?p=55959">TrendForce</a></li><li><strong>Anthropic</strong> — Nvidia is to buy $10.0B of the $26.0B private placement beside the IPO · <a href="https://fourweekmba.com/ai-nvidia-google-buy-10b-each-of-anthropics-26b-placement/">FourWeekMBA, on the S-1/A of Nov 10</a></li><li><strong>Anthropic</strong> — Third quarter: compute and infrastructure $9,863M; operating expenses $17.5B, including a $2B non-cash charge · <a href="https://06.ai/wire/w001">06.ai W.001, from the S-1/A of Nov 2</a></li><li><strong>Anthropic</strong> — 39% of the value of its seven largest disclosed compute contracts runs on Nvidia GPUs · <a href="https://06.ai/packs/nvidia-after-jensen">06.ai </a><a href="https://06.ai/packs/nvidia-after-jensen">Nvidia-after-Huang</a><a href="https://06.ai/packs/nvidia-after-jensen"> pack</a></li><li><strong>Compute</strong> — A gigawatt cost Anthropic $9.8B a year in the second quarter and $12.7B in the third · <a href="https://06.ai/packs/anthropic-s1">06.ai </a><a href="https://06.ai/packs/anthropic-s1">Anthropic S-1</a><a href="https://06.ai/packs/anthropic-s1"> pack</a></li><li><strong>CoreWeave</strong> — Lenders advanced 70% to 90% of the four programs financed since July 2025: about 80% in July 2025, 90% in March 2026, about 72% in May and 70% in August; the March facility was rated A3 and A (low) on its customer contract; equipment was 64.3% of property and equipment at June 30 and 69.2% of first-half additions · <a href="https://capefearadvisors.substack.com/p/nvidia-what-comes-back">Cape Fear Advisors</a></li><li><strong>CoreWeave</strong> — Total debt $21.4B at December 31, 2025 and $46.9B at September 30, 2026; capex of $30.65B on its own measure in the nine months · <a href="https://06.ai/wire/w076">06.ai W.076, from its quarterly releases</a></li><li><strong>CoreWeave</strong> — November 10 call: its delayed-draw term loans finance 70% to 100% of the capex they are tied to; the next, an investment-grade facility due in the coming weeks, more than 90% · <a href="https://s205.q4cdn.com/133937190/files/content_files/CoreWeave-3Q26-Transcript-_1_.pdf">CoreWeave call transcript, Nov 10</a></li><li><strong>Commitments</strong> — Signed for customers: guarantees, AI-cloud agreements, leases for partners and equity commitments, $189.5B · <a href="https://06.ai/packs/nvidia-after-jensen">06.ai </a><a href="https://06.ai/packs/nvidia-after-jensen">Nvidia-after-Huang</a><a href="https://06.ai/packs/nvidia-after-jensen"> pack</a><a href="https://06.ai/packs/nvidia-after-jensen">, from the 10-Q</a></li><li><strong>Outlook</strong> — Analysts expect revenue of $692B in the year to January 2028 and $124.5B in the quarter to January 2027 · <a href="https://06.ai/packs/nvidia-after-jensen">06.ai </a><a href="https://06.ai/packs/nvidia-after-jensen">Nvidia-after-Huang</a><a href="https://06.ai/packs/nvidia-after-jensen"> pack</a><a href="https://06.ai/packs/nvidia-after-jensen"> (Yahoo Finance, Nov 6)</a></li><li><strong>Suppliers</strong> — TSMC’s largest customer was 19% of its 2025 revenue, NT$726.9B; read as Nvidia · <a href="https://view.asiae.co.kr/en/article/2026030610061501773">Asia Economy, Mar 9, on TSMC’s annual report</a></li><li><strong>Suppliers</strong> — TSMC’s 2025 revenue: NT$3,809.05B, or $122.42B · <a href="https://telecomlead.com/semiconductor/tsmc-delivers-record-2025-revenue-and-profit-as-ai-chip-demand-fuels-growth-us-expansion-accelerates-124150">TelecomLead</a></li><li><strong>Suppliers</strong> — Operating margins in the latest quarter: SK hynix and Micron about 80%, TSMC 58.7% · <a href="https://06.ai/wire/w044">06.ai W.044</a></li><li><strong>Venture</strong> — $407B invested in AI companies in the first half of 2026; $264.1B in all of 2025 · <a href="https://en.paperjam.lu/article/pitchbook-ai-funding-soars-as-megadeals-take-control">PitchBook, via Paperjam</a></li></ul>
<h2>Method</h2>
<p><strong>Period and units.</strong> Reported figures are Nvidia’s, in $M: the half-year to July 26, 2026 from its results release and quarterly report, and fiscal 2024 to 2026 from its annual report. A cent is a figure divided by revenue for the same period, times 100. The model checks that every income statement and cash flow statement adds to the totals Nvidia printed. The half-year is the first half of Nvidia’s fiscal 2027.</p>
<p><strong>Profit before gains.</strong> Net income ($118,010M) less net gains from equity securities ($23,707M) plus the income tax effect of Nvidia’s non-GAAP adjustments ($4,407M), nearly all of which is tax on those gains: $98,710M. Income tax in the chart is the reported expense less that $4,407M. Profit includes 0.2 cents of interest and other income.</p>
<p><strong>Not yet cash.</strong> Cash from operations less profit before gains. It is the rise in receivables (13.8 cents), inventories (5.7) and prepaid and other assets (3.6), less the rise in payables and accruals (7.9) and non-cash costs such as stock-based pay and depreciation (4.1), and the tax on gains (2.5), which is in the reported tax and not in ours.</p>
<p><strong>Cash to spend.</strong> Cash from operations plus the notes issued ($24,896M, net of costs) plus securities sold, net ($12,558M): equity securities sold ($7,241M), debt securities sold and matured less those bought ($4,786M), employee share purchases ($515M) and other items. Spending is equity securities bought ($42,404M), shares repurchased ($39,044M), dividends ($6,290M), tax paid on employees’ vesting shares ($4,531M), capital expenditure including principal payments ($4,526M), and the Groq payment with other acquisitions ($3,242M). The difference is the rise in cash, $11,838M.</p>
<p><strong>Where the stakes went.</strong> Private: net additions to non-marketable equity securities in the quarterly report’s notes. Everything else in equity securities bought, $11,399M, is taken as listed companies and financing vehicles, and may include part of the $3.3B of equity-method stakes in infrastructure financiers. The five named purchases are as announced and assumed paid in the half. CoreWeave’s was announced on January 26, the first day of the fiscal year. If it was paid before that day, the cloud line is $2B, the unnamed remainder $3.4B and our central estimate 4.0 cents. Nvidia’s filings name no investee.</p>
<p><strong>What comes back.</strong> For a lab: the share of its spending that is compute, times the share of that compute on Nvidia systems, times the share of a compute dollar on Nvidia systems that becomes Nvidia revenue. The third figure divides Nvidia’s revenue per gigawatt ($25B to $40B) by five years of rent at $10B to $13B, which gives 38% to 80%; we use 40% to 65% because rents for Vera Rubin capacity are not yet known. The low and high cases move every assumption the same way at once. All private additions are treated as labs, which overstates the return to the extent they went to other startups. The estimate is a share of the half’s stakes. It does not say the money was spent within the half.</p>
<p><strong>The down payment.</strong> If lenders fund a share a of what an AI cloud builds and Nvidia hardware is a share s of it, a dollar of equity supports s ÷ (1 − a) dollars of Nvidia orders. a is the rise in CoreWeave’s debt in the first nine months of 2026 ($25.5B) over its capex on its own measure ($30.65B): 83%. It is a rough figure, since some debt refinances earlier debt and that capex includes assets not yet paid for, and the other 17% was not all equity. The range uses the lowest and highest advance rates on its facilities. CoreWeave said on November 10 that such loans finance anywhere from 70% to 100% of the capex they are tied to; at 100% no equity is needed and the formula has no limit, so the high case stops at 90%. s is the 50% (45% to 60%) used above. Nebius is assumed to finance itself the same way. For labs no advance rate is filed, so the chart divides Nvidia’s own count of their demand next year by the nearly $50B it has invested; the pink part is the 20% from the cash estimate.</p>
<p><strong>The loop.</strong> If a share r of each revenue dollar returns as revenue, a dollar from outside becomes 1 ÷ (1 − r) dollars of revenue: 1.03, 1.05 and 1.09 in the three cases of the cash estimate. The figure with the cloud stakes’ borrowing adds the central case for labs (3.4 cents) to the clouds’ orders with borrowing (6.7 cents).</p>
<p><strong>Balance-sheet support.</strong> Lab demand is 25% of the revenue analysts expected for the year to January 2028 when the <a href="https://06.ai/packs/nvidia-after-jensen">Nvidia-after-Huang</a> pack recorded it on November 6. The $189.5B is that pack’s total of what Nvidia has signed for customers: guarantees behind OpenAI’s Ohio campus of up to $105B, entered in August, and other guarantees of $3.5B; AI-cloud agreements of $36B; leases signed for partners of $20B; and equity commitments of $25B.</p>
<p><strong>Not counted.</strong> What suppliers, staff, governments and shareholders do with their cents, some of which also ends up funding AI companies. Stakes bought since July 26.</p>
<hr>
<p>Every chart above links to its interactive version on 06.ai, with the data behind it. The daily trackers are at <a href="https://06.ai/trackers">06.ai/trackers</a>.</p>
<p><em>Not investment advice. Estimates are labeled as estimates.</em></p>]]></content:encoded>
</item>
<item>
  <title>Which neoclouds depend most on Nvidia?</title>
  <link>https://inthesingularity.ai/packs/neocloud-nvidia-exposure</link>
  <guid isPermaLink="true">https://inthesingularity.ai/packs/neocloud-nvidia-exposure</guid>
  <pubDate>Fri, 13 Nov 2026 12:28:00 GMT</pubDate>
  <description>Nvidia’s latest filing holds $59.5B of promises made for AI clouds: $36B of capacity it must rent if they cannot sell it, $20B of leases signed for them and $3.5B of rent guarantees. Set against each company’s own contracts, the smallest clouds lean hardest. Nvidia’s floor is 56% of Sharon AI’s contract value; its order is 5% of CoreWeave’s backlog and 13% of its debt; Nebius has no Nvidia credit at all. With Jensen Huang gone, the exposure is in what was never finished: a guarantee of more than $5B in talks for Nscale, a lease Nvidia holds under Lambda’s $34.9B Anthropic agreement, and new equity for Firmus.</description>
  <category>AI infrastructure</category>
  <enclosure url="https://inthesingularity.ai/newsletter/neocloud-nvidia-exposure/cover.png" length="55394" type="image/png" />
  <media:content url="https://inthesingularity.ai/newsletter/neocloud-nvidia-exposure/cover.png" medium="image" type="image/png" />
  <content:encoded><![CDATA[<figure><a href="https://06.ai/packs/neocloud-nvidia-exposure"><img src="https://06.ai/newsletter/neocloud-nvidia-exposure/thumb.png" alt="Which neoclouds depend most on Nvidia?" width="421" height="392"></a><figcaption><a href="https://06.ai/packs/neocloud-nvidia-exposure">Interactive version and data on 06.ai →</a></figcaption></figure>
<ul><li><strong>$59.5B</strong> — Signed for AI clouds, in Nvidia’s July filing. $36B of capacity floors, $20B of leases to hand on, <strong>$3.5B</strong> of rent guarantees</li><li><strong>56%</strong> — Nvidia’s floor, share of Sharon AI’s contracts. $4.9B of <strong>$8.8B</strong>; the highest of the six clouds we could measure</li><li><strong>13%</strong> — CoreWeave’s Nvidia order, against its debt. $6.3B against <strong>$46.9B</strong>; 5% of its backlog</li><li><strong>&gt;$5B</strong> — In talks, not signed. A guarantee for Nscale’s Anthropic building. Lambda’s <strong>$34.9B</strong> agreement also waits on a lease Nvidia holds</li></ul>
<h2>What this adds</h2>
<p><strong>The </strong><a href="https://06.ai/packs/nvidia-after-jensen"><strong>Nvidia-after-Huang</strong></a><strong> pack</strong><strong> found that the first thing to fade after Jensen Huang is the bank of Nvidia.</strong> Nvidia announced Huang’s resignation on November 12 and Colette Kress, the finance chief, is interim chief executive (<a href="https://06.ai/packs/nvidia-after-jensen">the </a><a href="https://06.ai/packs/nvidia-after-jensen">Nvidia-after-Huang</a><a href="https://06.ai/packs/nvidia-after-jensen"> pack</a>). This pack takes that one job and goes company by company. How much of each neocloud’s debt, leases, contracts and equity rests on Nvidia, and which part of it is signed?</p>
<p><strong>Nvidia ties itself to a cloud in four ways, and they bind differently.</strong> It buys shares. It becomes a customer, either renting capacity for its own research or promising to rent whatever the cloud cannot sell. It puts its credit under a building, by guaranteeing the rent or by signing the lease itself to pass on. And it decides who gets chips first. The first three sit in filings and contracts, and a contract binds the company whoever runs it. Allocation, and every deal still being negotiated, were Huang’s to give.</p>
<h2>What Nvidia has signed for AI clouds</h2>
<figure><a href="https://06.ai/packs/neocloud-nvidia-exposure#lines"><img src="https://06.ai/newsletter/neocloud-nvidia-exposure/fig-01.png" alt="What Nvidia has signed for AI clouds" width="694" height="192"></a><figcaption>$ billions, from Nvidia’s filing for the quarter to July 26. The filing names no partner. Strong grey is the part outside reporting attaches to a company: SemiAnalysis’s count of the three capacity deals announced so far, Hut 8’s two leases at Beacon Point in Texas, whose tenant the Financial Times reported to be Nvidia, and the rent guarantee in Nscale’s IPO filing. Grey is the rest. The $105B guarantee behind OpenAI’s Ohio campus is separate and not shown.</figcaption></figure>
<h2>A floor, a lease and a guarantee</h2>
<p><strong>The largest tie is a floor under a cloud’s revenue: $36B.</strong> The filing calls these AI cloud agreements. A cloud buys Nvidia’s systems and Nvidia commits to rent the capacity, typically for six years. The cloud can stop supplying Nvidia whenever it finds a customer who pays more, and the commitment falls as it does. Where the cloud earns more than the floor Nvidia takes a share, half of revenue above a threshold according to the Wall Street Journal. Three deals have been announced. SemiAnalysis puts Nvidia’s obligation at $21.1B for Firmus, $4.2B for Sharon AI and about $2.2B for GMI Cloud, 76% of the line between them.</p>
<p><strong>The floor is about half the going rate.</strong> Sharon AI describes its $4.9B as minimum revenue over six years on up to 40,000 GPUs, which works out at $2.33 per GPU-hour. Firmus’s $21.1B on up to 170,000 is $2.36. SemiAnalysis’s average for the programme is $2.35, against $4.50 to $4.60 for five-year contracts. Our own index of rental prices has the B300, the same generation of chip sold in servers, at $5.89 on spot and $9.67 on demand. A floor at 52% of the contract price is not the cloud’s profit. Firmus’s is about 1.4 times the estimated $15B cost of its chips, earned over six years, before power, buildings and interest. What it buys is a lender’s confidence that the loan will be serviced.</p>
<p><strong>Then there is Nvidia’s name on the lease.</strong> Nvidia has signed about $20B of leases, of about 15 years, that it expects to hand to others, starting in its fiscal 2028 and 2029. One campus fits that description. Hut 8’s Beacon Point in Nueces County, Texas, is let on two 15-year leases of 352 MW each, worth $9.8B apiece, to a tenant Hut 8 does not name and the Financial Times identified in July as Nvidia. Nvidia has not confirmed it. Separately Nvidia guarantees up to $3.5B of rent on other clouds’ leases for five to seven years, with $712M of the tenants’ money in escrow. One of those is public: up to $860.3M on Nscale’s site in Ward County, Texas, the first five years of rent, given in return for warrants valued at $60M and with $470M of Nscale’s cash held in escrow against it.</p>
<h2>The floor against the market</h2>
<figure><a href="https://06.ai/packs/neocloud-nvidia-exposure#floor"><img src="https://06.ai/newsletter/neocloud-nvidia-exposure/fig-02.png" alt="The floor against the market" width="694" height="306"></a><figcaption>Dollars per GPU-hour. The two floors are our arithmetic: each company’s minimum spread over every GPU for the full six years. A slower ramp would make the hourly figure higher. The five-year contract range is the one SemiAnalysis sets against its GB300 figure. The index rows are the B300 in servers, from 06.ai’s <a href="https://06.ai/trackers/gpu-compute-price-index">GPU compute price index</a>, seven-day averages at the November 12 fixing.</figcaption></figure>
<h2>Company by company</h2>
<p><strong>Debt.</strong> We found no neocloud debt that Nvidia guarantees directly. What it backs is the revenue underneath. For the three clouds in the capacity programme that is nearly everything. GMI’s Taiwan loan is capped at 55% of the value of its Nvidia contracts and Nvidia’s share of revenue is paid ahead of the banks, IFR reported; Sharon AI raised $1.6B in June to fund its side. For CoreWeave it is little. The $6.3B order for unsold capacity is 13% of $46.9B of debt, and its secured loans are sized on its customers’ contracts (<a href="https://06.ai/wire/w046">W.046</a>). Nebius and IREN have no Nvidia promise under their borrowing that we could find.</p>
<p><strong>Leases.</strong> Two companies, Nscale and Lambda. Nvidia’s guarantee covers 44% of the contracted rent on Nscale’s Texas lease, $860.3M of about $1.95B, and that one guarantee is 24% of everything Nvidia has guaranteed for clouds. Lambda’s largest contract depends on a lease that is reported to be Nvidia’s. Anthropic’s August agreement, $34.9B, is for about 350 MW at Beacon Point. Anthropic’s filing of November 10 says Lambda did not secure the facility lease by the agreed date, so either side may now end the agreement (<a href="https://06.ai/wire/w084">W.084</a>). It is 70% of the $50B backlog Lambda showed investors in September. Anthropic says it does not intend to terminate.</p>
<p><strong>Contracts.</strong> Where Nvidia is itself the customer, its share of contracted revenue runs from 56% at Sharon AI to 17% at IREN, 5% at CoreWeave, 3% at Lambda and 1% at Nscale, where it rents $1.2B of capacity in Europe under a 2025 agreement. Nebius has no order from Nvidia that we could find. Firmus and GMI publish no backlog to measure against.</p>
<p><strong>Equity.</strong> Nscale leans hardest. Of the $3.36B it signed up in September, $2.0B is Nvidia’s, 60%: $1.0B of the convertible notes, and a further $1.0B due to close around November 16. In March Nvidia had put $777M into a round of about $2.1B. It supplied $2.0B of the $16.8B Nebius raised in the nine months to September, 12%, and $2.0B of the roughly $31.7B CoreWeave raised this year to mid-October, 6%. It took part in Firmus’s last three rounds; the Australian Financial Review reported in July that it would lead the one that closed in August with about A$720M. Its share of rounds at Lambda, Crusoe and GMI was not disclosed. IREN gave it a right to buy 30M shares at $70; the shares are near $19. Nvidia’s January purchase of CoreWeave at $87.20 a share stood 30% under water on November 11, after the shares had risen 25% that day.</p>
<h2>Nvidia’s commitment to each company</h2>
<figure><a href="https://06.ai/packs/neocloud-nvidia-exposure#commit"><img src="https://06.ai/newsletter/neocloud-nvidia-exposure/fig-03.png" alt="Nvidia’s commitment to each company" width="694" height="369"></a><figcaption>$ billions at the value announced or filed, not what remains: orders, floors, rent guarantees and leases, not shares. Dashed bars wait on Nvidia: Nscale’s guarantee is in talks and is “more than” $5B as reported; Lambda’s is one of the two Beacon Point leases, which Lambda had not secured by the date in its Anthropic agreement. Firmus’s and GMI’s figures are SemiAnalysis’s; Sharon AI’s is its own. Firmus’s unsigned item is equity and is not drawn.</figcaption></figure>
<ul><li><strong>Firmus</strong> — Shares: In each of its last three rounds; about A$720M of August’s, as reported · As a customer: Floor on up to 170,000 GPUs at Batam for six years: about $21.1B (SemiAnalysis) · Under the lease or the loan: None found · Not signed: New equity: a $2–3B private round, with Nvidia in talks for hundreds of millions</li><li><strong>Sharon AI</strong> — Shares: None found · As a customer: Floor on up to 40,000 GPUs for six years: $4.9B, 56% of its contract value · Under the lease or the loan: None found · Not signed: Nothing reported</li><li><strong>GMI Cloud</strong> — Shares: In its $223M round · As a customer: Rents unused capacity for up to six years: about $2.2B (SemiAnalysis). Takes half of what other customers pay, as reported · Under the lease or the loan: Its Taiwan loan is capped at 55% of the value of Nvidia’s contracts · Not signed: Nothing reported</li><li><strong>Nscale</strong> — Shares: More than 5%. $2.0B of the $3.36B signed up in September, 60%; $1.0B of it is due around November 16 · As a customer: Rents $1.2B of capacity in Europe under a July 2025 agreement, 1% of contract value · Under the lease or the loan: Guarantees the first five years of rent in Ward County, Texas: up to $860.3M, 44% of the lease · Not signed: A guarantee of more than $5B for a building at Monarch</li><li><strong>Lambda</strong> — Shares: An investor since 2025; amounts not disclosed · As a customer: Rents back about $1.5B of capacity, 3% of backlog · Under the lease or the loan: Reported to hold both leases at Beacon Point, $9.8B each, where Lambda’s Anthropic capacity is to go · Not signed: The lease behind $34.9B, 70% of backlog</li><li><strong>CoreWeave</strong> — Shares: Roughly 10–13%, as reported; $2.0B in January, 6% of what it raised this year · As a customer: Order for unsold capacity to April 2032: $6.3B, 13% of debt · Under the lease or the loan: None found · Not signed: Nothing reported</li><li><strong>Nebius</strong> — Shares: $2.0B of pre-funded warrants in March, 12% of nine months’ fundraising · As a customer: None found · Under the lease or the loan: None found · Not signed: Nothing reported</li><li><strong>IREN</strong> — Shares: A right to buy 30M shares at $70; the shares are near $19 · As a customer: Five-year cloud contract: $3.4B, 17% of backlog · Under the lease or the loan: None found · Not signed: The rest of a partnership of up to 5 GW</li><li><strong>Crusoe</strong> — Shares: One of more than two dozen investors in its $3.9B round · As a customer: None found · Under the lease or the loan: None found · Not signed: Nothing reported</li></ul>
<h2>How much of each company is Nvidia</h2>
<figure><a href="https://06.ai/packs/neocloud-nvidia-exposure#share"><img src="https://06.ai/newsletter/neocloud-nvidia-exposure/fig-04.png" alt="How much of each company is Nvidia" width="694" height="401"></a><figcaption>Top: Nvidia’s own order or floor as a share of the revenue each company has under contract. The dates differ: Sharon AI at August 6, CoreWeave and Nebius at September 30, IREN at November 5, Nscale at November 12, Lambda in September as reported. Bottom: Nvidia’s money, paid or signed for, as a share of what each raised from investors and lenders over the period shown. Firmus and GMI publish neither figure.</figcaption></figure>
<h2>What Huang’s exit changes</h2>
<p><strong>What is signed stays signed.</strong> Nvidia’s board said on Thursday that it expects customer commitments to continue without interruption. Goldman Sachs’s James Schneider expects announced commitments to be met and thinks multi-party deals on the old scale will be much harder to arrange without Huang. A floor, an order or a guarantee is an obligation of a company that produced $69.9B of free cash flow in its first half.</p>
<p><strong>What is not signed needs a yes, and the person who gives it has said no before.</strong> The Information reports that Kress pushed back on the backstop Huang first proposed for OpenAI’s Ohio campus, about $250B. The guarantee signed in August is capped at $105B. Of the financing as a whole she told investors in August that “we’re not making loans”, and that outside capital underwrites each deal. She also co-wrote the post that launched the capacity programme in July, so the programme is hers as much as his. The reported difference between them is size.</p>
<p><strong>The pullback began before he left.</strong> On August 27 the Journal reported that Nvidia had paused some capacity deals. Employees had warned that the programme could draw antitrust scrutiny, and some partners had been irritated by limits on which customers they could serve. Nvidia said the model was still in place. Origin Partners’ Adil Malik told CNBC this week that for about three months Nvidia has wanted a binding letter of intent or a signed customer contract before it will backstop a client.</p>
<p><strong>The vehicle meant to carry the load has not closed a deal.</strong> In August six financiers set out to raise more than $500B for Nvidia-based data centers, with Nvidia able to support up to 25% of a deal. Bloomberg reported on Thursday that none has closed, that partners want Nvidia to cover more, and that a first deal may not come before January. Rivals offer more. Broadcom backstopped more than 80% of a $35B structure for Anthropic, Reuters reported, and stands behind up to 72.5% of what is owed on a $63B one (<a href="https://06.ai/wire/w046">W.046</a>).</p>
<p><strong>Washington is the other change.</strong> The Commerce Department is preparing a rule that would require licences to bring advanced chips into seven countries and make buyers disclose who uses them, The Information reported on November 11. Indonesia is one of the seven. Firmus’s Batam campus, the largest floor Nvidia has announced, is in Indonesia, is being built with Singapore’s DayOne and is due to start in the first quarter of 2027. By one account of the report, data centers run by US companies are outside the rule; Firmus is Australian. About 30% of GMI’s revenue comes from Chinese customers, The Information reported. Bankers told IFR that loans of this kind in Asia are exposed to US export controls, which could lock or downgrade the chips that secure them.</p>
<h2>Who is most exposed</h2>
<p>Ordered by how much of a company’s next financing step needs something from Nvidia that is not signed. The ordering is ours. It is not a forecast of who fails, and a company low on the list can still struggle to borrow for other reasons.</p>
<ul><li><strong>1</strong> — Company: Nscale · What waits on Nvidia: A guarantee of more than $5B on a 460 MW building for Anthropic, in talks since October and unsigned at the end of it · Why it matters now: Its prospectus, amended on November 12, says none of the financing its Anthropic agreements need is committed; the first building costs $10–10.5B before chips. By the FT’s reading Anthropic may leave if financing misses a deadline. About 6% of its $105.8B of contracts is active, Microsoft passed on its Texas option and the IPO has slipped.</li><li><strong>2</strong> — Company: Lambda · What waits on Nvidia: The lease on about 350 MW at Beacon Point, which Nvidia is reported to hold · Why it matters now: Anthropic’s $34.9B agreement, 70% of Lambda’s backlog, has been open to termination by either side since Lambda missed the date to secure the lease. Anthropic says it does not intend to use the right. A $4B round is not reported closed.</li><li><strong>3</strong> — Company: Firmus · What waits on Nvidia: New equity. It pulled a $5B IPO on October 9 and is seeking $2–3B privately; Nvidia’s part is undecided · Why it matters now: The floor at Batam is signed. But analysts working on the IPO put its debt at about $30B, and Batam is in Indonesia, one of seven countries in a reported US licensing rule.</li><li><strong>4</strong> — Company: CoreWeave · What waits on Nvidia: Nothing it has disclosed · Why it matters now: The market trades it as Nvidia’s proxy. Its test is a loan of at least $4B that could launch as early as next week at 7–7.5%, against about 15% on its notes in October.</li><li><strong>5</strong> — Company: Nebius · What waits on Nvidia: Nothing it has disclosed · Why it matters now: No Nvidia credit to lose. It plans $40–45B of capex in 2027 with 75–85% of it funded, which leaves $6–11B, and has a $3.5B 364-day loan to replace.</li><li><strong>6</strong> — Company: Sharon AI, GMI Cloud · What waits on Nvidia: Later phases of a programme the Journal says was partly paused in August · Why it matters now: The most tied by share, and signed. GMI gets about 30% of its revenue from Chinese customers, which it expects to fall to about 10% next year.</li><li><strong>7</strong> — Company: IREN · What waits on Nvidia: The rest of a partnership of up to 5 GW · Why it matters now: Nvidia is a customer under a signed contract. IREN still has $6B of this fiscal year’s capex to fund.</li><li><strong>8</strong> — Company: Crusoe · What waits on Nvidia: Nothing · Why it matters now: Nvidia is one of more than two dozen investors in its last round. Its sites are being built for Oracle and Microsoft.</li></ul>
<h2>What a missing backstop costs</h2>
<p><strong>Lenders already price the difference.</strong> GPU-backed loans this year cost about 6% when an investment-grade customer stood behind the chips and 8–10% when none did (<a href="https://06.ai/wire/w046">W.046</a>). That is $20M to $40M a year on each $1B borrowed. CoreWeave expects its new loan, secured on a Meta contract, at 7–7.5%; its unsecured notes yielded about 15% in October.</p>
<p><strong>Others can lend their credit, and each ties the cloud to a tenant.</strong> Google backstops leases for Fluidstack and for Anthropic (<a href="https://06.ai/wire/w013">W.013</a>). Broadcom finances its lab customers through its own vehicles. Customers prepay: CoreWeave expects about $9B in three months and Nebius $17–20B in 2027. What Nvidia’s floor offered was different. A loan banker in Taiwan told IFR that it lets a cloud rent to any AI company, where earlier financings needed Google, Microsoft or OpenAI as the tenant.</p>
<p><strong>So the split is by business model.</strong> Clouds that rent to hyperscalers and large labs on long contracts borrow against their customers: CoreWeave, Nebius, IREN and Crusoe. Clouds that rent to many small customers borrow against Nvidia: Firmus at Batam, Sharon AI and GMI. Clouds building for one lab without an investment-grade rating need a third party’s credit, and had been getting Nvidia’s: Nscale and Lambda, both for Anthropic, which Blackstone has warned may be rated below investment grade after its IPO. For the first group Nvidia is a shareholder and a supplier, and neither changes on paper. The second keeps what it has signed and may not get more. The third is waiting for an answer.</p>
<h2>What to watch</h2>
<ul><li><strong>About Nov 16.</strong> Nvidia’s further $1.0B of Nscale notes or shares is due to close. It was signed on September 15 and is the nearest dated payment in this pack.</li><li><strong>Nov 17.</strong> Nvidia’s results and quarterly filing. Three lines say whether the bank is still open: $36B of capacity floors, $20B of leases to hand on and $3.5B of rent guarantees at the end of July. SemiAnalysis expected the first to reach about $51B by January. A lower figure can be good news, because the commitment falls when a cloud sells the capacity to someone else.</li><li><strong>Next week.</strong> CoreWeave’s loan of at least $4B, secured on chips and a Meta contract, could come to market. Management expects 7–7.5%. It would be the first large neocloud financing since Huang left.</li><li><strong>About Nov 19.</strong> Anthropic’s shares are due to start trading. Anthropic is the tenant behind the contracts that Nscale and Lambda most need to finance, and Nvidia was down to buy up to $10B alongside the IPO.</li><li><strong>This month.</strong> The Commerce Department’s rule on advanced chips in seven countries, Indonesia among them, if it is published as reported.</li><li><strong>Before it lists.</strong> Whether Nscale’s next amendment shows financing for the Anthropic building committed, and whether an Nvidia guarantee is in it.</li><li><strong>No date.</strong> Whether Lambda secures the Beacon Point lease and closes its $4B round, and whether Firmus’s private round includes Nvidia.</li></ul>
<h2>What this is anchored on</h2>
<ul><li><strong>Nvidia</strong> — AI cloud agreements $36B, typically six years, falling as capacity is used by others; a revenue share if criteria are met; data center leases not commenced for third parties $20B, about fifteen years, starting fiscal 2028–29; land, power and shell guarantees $3.5B, reduced as partners pay over five to seven years, with $712M in escrow; SB Energy guarantees capped at $105B; Data Center revenue $48.7B hyperscale and $40.3B AI clouds, industrial and enterprise · <a href="https://www.sec.gov/Archives/edgar/data/0001045810/000104581026000075/nvda-20260726.htm">Nvidia, Form 10-Q for the quarter to July 26</a></li><li><strong>Nvidia</strong> — Huang’s resignation announced on November 12; Kress interim chief executive; the board expects customer commitments to continue without interruption; results on November 17 · <a href="https://www.sec.gov/Archives/edgar/data/0001045810/000104581026000082/nvidiaannouncesceotransiti.htm">Nvidia, Form 8-K, Nov 12</a></li><li><strong>Nvidia</strong> — What Huang did for customers; the commitments in total, $29B of cloud service agreements among them; up to $10B beside Anthropic’s IPO · <a href="https://06.ai/packs/nvidia-after-jensen">06.ai, </a><a href="https://06.ai/packs/nvidia-after-jensen">the </a><a href="https://06.ai/packs/nvidia-after-jensen">Nvidia-after-Huang</a><a href="https://06.ai/packs/nvidia-after-jensen"> pack</a></li><li><strong>Capacity programme</strong> — Six-year floors at about $2.35 per GPU-hour for the GB300 against $4.50–4.60 for five-year contracts; obligations of $21.1B for Firmus, $4.2B for Sharon AI and an estimated $2.2B for GMI; the line expected near $51B by January · <a href="https://newsletter.semianalysis.com/p/nvidias-backstop-universe-heads-i">SemiAnalysis, Sep 11</a></li><li><strong>Capacity programme</strong> — Launched in July with Sharon AI and Firmus; Nvidia rents back unused GPUs at a fixed rate and shares in cloud revenue · <a href="https://www.datacenterdynamics.com/en/news/nvidia-acts-as-backstop-for-customer-gpus-in-return-for-cut-of-cloud-revenue/">DCD, Jul 2026</a></li><li><strong>Capacity programme</strong> — Some deals paused; employees’ warnings on antitrust; clouds could rent only to approved customers; half of revenue above a threshold; Nvidia says the model is still in place · <a href="https://www.thestar.com.my/tech/tech-news/2026/08/28/nvidia-pauses-revenue-sharing-deals-with-ai-cloud-companies-wsj-reports">Reuters on the WSJ, via The Star, Aug 28</a></li><li><strong>Capacity programme</strong> — Firmus expected $25–30B of committed offtake in six years; neither named deal is known to be among those paused · <a href="https://thenextweb.com/news/nvidia-pauses-ai-cloud-revenue-sharing-deals">TNW, Aug 28</a></li><li><strong>Rental prices</strong> — B300 at $5.89 spot and $9.67 on demand, seven-day averages at the November 12 fixing · <a href="https://06.ai/trackers/gpu-compute-price-index">06.ai, GPU compute price index</a></li><li><strong>Firmus</strong> — A 360 MW campus at Batam, Indonesia, with DayOne; up to 170,000 GPUs; partnership through 2034 · <a href="https://firmus.co/newsroom/firmus-to-build-170-000-gpu-ai-factor-y-campus-with-nvidia-for-global-ai-natives">Firmus, Jun 2026</a></li><li><strong>Firmus</strong> — The Batam campus is due to go live in the first quarter of 2027 · <a href="https://thenextweb.com/news/firmus-nvidia-indonesia-batam-data-center-30-billion">TNW, Jun 28</a></li><li><strong>Firmus</strong> — Nvidia in its last three rounds; more than $13B raised, about $10B of it debt; a backstop on chips estimated at $15B; talks to invest hundreds of millions in a private round of at least $2B · <a href="https://dealroom.co/news/info-17vezxs-nvidia-moves-to-back-gpu-cloud-provider-firmus-after-investors-balk-at-i">The Information, via Dealroom, Oct 20</a></li><li><strong>Firmus</strong> — Nvidia to lead the round that closed in August, with about A$720M · <a href="https://dealroom.co/news/137837-nvidia-leads-firmus-a-2b-raise-at-15-5b-valuation-ahead-of-asx-ipo/">AFR, via Dealroom, Jul 9</a></li><li><strong>Firmus</strong> — IPO withdrawn on October 9; it sought about $5B at a $30.6B valuation; $2B raised in August at more than $10.5B · <a href="https://www.cnbc.com/2026/10/09/nvidia-ai-firmus-ipo-.html">CNBC, Oct 9</a></li><li><strong>Firmus</strong> — Analysts working on the IPO put its debt at about $30B · <a href="https://finimize.com/content/nvidia-backed-firmus-pulled-its-5-billion-australia-ipo">Reuters, via Finimize, Oct 8</a></li><li><strong>Sharon AI</strong> — Total contract value about $8.8B at August 6; a six-year collaboration with Nvidia worth $4.9B for up to 40,000 GB300 GPUs, listed beside three other contracts worth $2.6B · <a href="https://www.sec.gov/Archives/edgar/data/0002068385/000149315226036254/ex99-1.htm">Sharon AI, Form 8-K, Aug 6</a></li><li><strong>Sharon AI</strong> — The $4.9B described as minimum revenue, an average of $817M a year, on an initial 72 MW; a floor, with revenue above it shared with Nvidia · <a href="https://allmind.ai/earnings/event/2951556">Sharon AI’s August 6 call, transcript via AllMind</a></li><li><strong>Sharon AI</strong> — $1.6B raised in June to fund the collaboration: about $900M of shares and warrants and $700M of 4.75% convertible notes · <a href="https://www.businesswire.com/news/home/20260617057283/en/Sharon-AI-Announces-Oversubscribed-US%241.6-Billion-Strategic-Financing-to-Accelerate-Expansion-of-AI-Factories-Across-Australia-and-the-Asia-Pacific">Sharon AI, Jun 17</a></li><li><strong>GMI Cloud</strong> — A Taiwan loan capped at 55% of the value of Nvidia’s contracts, drawn to 80% of server cost; Nvidia’s revenue share paid ahead of the banks; bankers on export controls and on who can be the tenant · <a href="https://www.ifre.com/loans/2457977/nvidia-helps-apacs-first-gpu-syndicated-loan">IFR, Jul 2026</a></li><li><strong>GMI Cloud</strong> — $223M of equity, Nvidia taking part, and a $445M credit facility; more than $600M of contracted annual revenue; Chinese customers about 30% of revenue, about 10% next year · <a href="https://dealroom.co/news/info-1pcrwid-exclusive-gpu-cloud-provider-gmi-raises-668-million-from-nvidia-and-othe/">The Information, via Dealroom, Sep 30</a></li><li><strong>GMI Cloud</strong> — $947M of loan commitments; Nvidia leases unused capacity for up to six years and takes half the revenue from other customers · <a href="https://dealroom.co/news/148352-gmi-cloud-lands-947m-in-loans-backed-by-nvidia-gpus/">Tech in Asia, via Dealroom, Sep 2</a></li><li><strong>Nscale</strong> — Active and contracted value $105.8B, of which $6.4B active; Anthropic up to about $44.6B; no binding commitments for the financing those agreements need; the first Monarch building at $10–10.5B before chips, about $1B of it spent by September · <a href="https://www.sec.gov/Archives/edgar/data/0002110365/000119312526458722/ck0002110365-20261112.htm">Nscale, Form S-1/A, Nov 12</a></li><li><strong>Nscale</strong> — Nvidia rents $1.2B of capacity under a July 2025 agreement and holds more than 5%; it guarantees up to $860.3M of rent for the first five years of the Ward County lease, for warrants valued at $60M, with $470M in escrow · <a href="https://www.sec.gov/Archives/edgar/data/0002110365/000119312526458722/ck0002110365-20261112.htm">Nscale, Form S-1/A, Nov 12</a></li><li><strong>Nscale</strong> — A $3.36B subscription on September 15: $2.36B of convertible notes, $1.0B of them Nvidia’s, and a further $1.0B from Nvidia to close on or around November 16; Nvidia put $400M into the Series B and $777M into a Series C of about $2.1B · <a href="https://www.sec.gov/Archives/edgar/data/0002110365/000119312526458722/ck0002110365-20261112.htm">Nscale, Form S-1/A, Nov 12</a></li><li><strong>Nscale</strong> — The Ward County lease: 126 months, 234 MW, about $1.95B of rent · <a href="https://theenergymag.com/news/market-news/nscale-3-billion-financing-texas-north-carolina-ai">The Energy Mag, on Ionic Digital’s filing</a></li><li><strong>Nscale</strong> — Microsoft declined its option in Texas · <a href="https://06.ai/wire/w019">06.ai, W.019</a></li><li><strong>Nscale</strong> — Talks on a guarantee of billions for a 460 MW building Anthropic is to use · <a href="https://techcrunch.com/2026/10/30/nvidia-weighs-playing-the-cosigner-for-nscales-ai-data-center-in-west-virginia/">TechCrunch, Oct 30</a></li><li><strong>Nscale</strong> — By the FT’s reading Anthropic may leave if financing misses a deadline; listing now more likely in December · <a href="https://thenextweb.com/news/nvidia-nscale-anthropic-data-centre-debt-guarantee">TNW, Oct 31</a></li><li><strong>Nscale</strong> — The guarantee, of more than $5B, was unsigned at the end of October; Firmus’s talks undecided · <a href="https://news.futunn.com/en/post/1004470221/nvidia-s-post-jensen-huang-era-has-begun-the-cfo">Futu News, on The Information, Nov 12</a></li><li><strong>Lambda</strong> — Anthropic’s three agreements, about $41.3B; the August one, $34.9B, open to termination since Lambda did not secure the facility lease by the agreed date; Anthropic does not intend to end it · <a href="https://06.ai/wire/w084">06.ai, W.084, from Anthropic’s S-1</a></li><li><strong>Lambda</strong> — Backlog $15B in June and $50B in September; a round of up to $4B at $14.5B before the money · <a href="https://techcrunch.com/2026/10/06/ai-computing-startup-lambda-to-raise-4b-ahead-of-planned-ipo/">TechCrunch, on the WSJ, Oct 6</a></li><li><strong>Lambda</strong> — Nvidia rents back $1.3B of capacity over four years and $200M more · <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/nvidia-signs-usd1-5-billion-deal-with-cloud-startup-lambda-to-rent-back-its-own-ai-chips-18-000-gpus-will-be-leased-over-4-years-as-lambda-gears-up-for-its-ipo">Tom’s Hardware, on The Information, Sep 2025</a></li><li><strong>Beacon Point</strong> — Two 15-year leases of 352 MW, $9.8B each, $19.6B in all, to one unnamed high-investment-grade tenant; second-phase delivery from the second quarter of 2028 · <a href="https://www.benzinga.com/node/60546113">Hut 8, Jul 20, via Benzinga</a></li><li><strong>Beacon Point</strong> — The tenant is Nvidia, according to five people · <a href="https://thenextweb.com/news/nvidia-50bn-texas-data-centre">TNW, on the FT, Jul 28</a></li><li><strong>Beacon Point</strong> — Lambda’s Anthropic capacity, about 350 MW, is at Hut 8’s site in Nueces County; the Journal reported that Nvidia holds the lease · <a href="https://www.thestandard.com.hk/innovation/article/341478/Anthropic-signs-US35-billion-cloud-deal-with-Nvidia-backed-Lambda-source-says">Reuters, via The Standard, Sep 1</a></li><li><strong>CoreWeave</strong> — Debt $46.9B, backlog $132.5B, interest 2.2 times adjusted operating income, about $9B of prepayments expected · <a href="https://06.ai/wire/w076">06.ai, W.076</a></li><li><strong>CoreWeave</strong> — Default-swap record of 1,267.9 basis points on November 5 · <a href="https://06.ai/wire/w029">06.ai, W.029</a></li><li><strong>CoreWeave</strong> — About $31.7B raised in 2026 to October 16, $3B of it equity; its 2031 notes at a 15.1% yield that day · <a href="https://06.ai/packs/neocloud-capex-funding">06.ai, </a><a href="https://06.ai/packs/neocloud-capex-funding">the </a><a href="https://06.ai/packs/neocloud-capex-funding">neocloud</a><a href="https://06.ai/packs/neocloud-capex-funding"> pack</a></li><li><strong>CoreWeave</strong> — A loan of at least $4B backed by chips and a Meta contract, which could be marketed from the week of November 16; management expects 7–7.5%; no unsecured debt at current yields · <a href="https://www.thestreet.com/investing/coreweave-debt-chip-backed-loan">TheStreet, Nov 13</a></li><li><strong>CoreWeave</strong> — Nvidia bought $2B of shares at $87.20 in January; its stake put at roughly 10% to 12.8% · <a href="https://www.tipranks.com/news/coreweave-stock-crwv-plunges-15-on-nvidia-contagion-after-ceo-jensen-huang-resigns">TipRanks, Nov 12</a></li><li><strong>CoreWeave</strong> — Shares at $60.77 on November 11, up 25% · <a href="https://stocktwits.com/news-articles/markets/equity/crwv-stock-best-week-nvidia-stake-reveal-debt-scare-hiding-bargain-morningstar/cNGzFYdR99u">Stocktwits, Nov 12</a></li><li><strong>CoreWeave</strong> — Nvidia’s order requires it to buy unsold capacity through April 13, 2032 · <a href="https://www.thestar.com.my/tech/tech-news/2025/09/15/coreweave-nvidia-sign-new-63-billion-deal-for-cloud-computing-capacity">Reuters, via The Star, Sep 2025</a></li><li><strong>Nebius</strong> — $16.8B raised in nine months, $2.0B of it pre-funded warrants; debt $14.2B; remaining obligations $38.7B; 2027 capex, funding and prepayments; a $3.5B 364-day facility · <a href="https://06.ai/wire/w096">06.ai, W.096</a></li><li><strong>Nebius</strong> — Nvidia’s $2B came as pre-funded warrants over 21.1M shares · <a href="https://thenextweb.com/news/nvidia-invests-2-billion-in-nebius">TNW, Mar 12</a></li><li><strong>IREN</strong> — Revenue backlog $20.3B; debt $10.2B; $6B of capex still to fund · <a href="https://06.ai/wire/w021">06.ai, W.021</a></li><li><strong>IREN</strong> — A five-year, $3.4B contract and a five-year right to buy 30M shares at $70; a partnership of up to 5 GW · <a href="https://www.lse.co.uk/news/nvidia-inks-deal-to-invest-up-to-21bn-in-iren-j5ohjvwxg2b5iix.html">Sharecast, via LSE, May 8</a></li><li><strong>IREN</strong> — Shares at $19.37 in early trading on November 12 · <a href="https://www.tradingview.com/news/moneycontrol:a779432bc094b:0-us-neocloud-stocks-crash-up-to-16-after-jensen-huang-quits-as-nvidia-ceo-nebius-leads-losses/">Moneycontrol, via TradingView, Nov 12</a></li><li><strong>Crusoe</strong> — $3.9B raised at a $30.9B valuation, Nvidia among more than two dozen investors; sites for Oracle and Microsoft · <a href="https://siliconangle.com/?p=847369">SiliconANGLE, Sep 17</a></li><li><strong>After Huang</strong> — Goldman’s James Schneider expects announced commitments to be met and the next multi-party deals to be much harder · <a href="https://www.cnbc.com/2026/11/13/nvidia-jensen-huang-resignation-ai-financing.html">CNBC, Nov 13</a></li><li><strong>After Huang</strong> — Kress pushed back on a proposed $250B backstop for OpenAI · <a href="https://fortune.com/2026/11/12/ai-industry-without-jensen-huang-investment-capital-regulation-neocloud-policy/">Fortune, on The Information, Nov 12</a></li><li><strong>After Huang</strong> — Until about three months ago a non-binding letter of intent was enough for a backstop; Nvidia now wants a binding one or a signed contract; the Ohio guarantee was discussed at $250B · <a href="https://financefeeds.com/nvidia-backstop-286-billion-colette-kress-openai-neoclouds/">FinanceFeeds, on CNBC, Nov 12</a></li><li><strong>After Huang</strong> — Kress in August: “we’re not making loans”; outside capital underwrites each deal · <a href="https://thenextweb.com/news/nvidia-q2-fy27-results-year-ahead-forecast">TNW, Aug 27</a></li><li><strong>After Huang</strong> — Kress co-wrote the post that launched the capacity programme · <a href="https://www.tomshardware.com/tech-industry/nvidia-to-take-a-cut-of-ai-cloud-revenue-on-top-of-hardware-sales">Tom’s Hardware, Jul 2</a></li><li><strong>Financing platforms</strong> — Six firms aiming at more than $500B; Nvidia may backstop up to 25% · <a href="https://www.khaleejtimes.com/business/nvidia-partners-with-wall-street-giants-to-raise-500-billion-for-ai-buildout">Reuters, via Khaleej Times, Aug 12</a></li><li><strong>Financing platforms</strong> — No deal in months; customers and partners asking Nvidia for a larger backstop; a first deal no earlier than January · <a href="https://www.bloomberg.com/news/articles/2026-11-12/nvidia-faced-pressure-to-boost-its-ai-backstop-before-huang-exit">Bloomberg, Nov 12</a></li><li><strong>Financing platforms</strong> — Lenders want more than the 25% Nvidia outlined; Broadcom backstopped more than 80% of a $35B structure for Anthropic · <a href="https://www.calcalistech.com/ctechnews/article/05nw8zhu9">Reuters, via Calcalist, Oct 3</a></li><li><strong>Credit</strong> — GPU loans at about 6% with an investment-grade customer and 8–10% without; Broadcom’s backstop of up to 72.5% of what is owed on a $63B lease financing · <a href="https://06.ai/wire/w046">06.ai, W.046</a></li><li><strong>Credit</strong> — Google’s backstops for Fluidstack’s and Anthropic’s leases · <a href="https://06.ai/wire/w013">06.ai, W.013</a></li><li><strong>Credit</strong> — Blackstone’s warning that Anthropic may be rated below investment grade · <a href="https://techcrunch.com/2026/10/30/nvidia-weighs-playing-the-cosigner-for-nscales-ai-data-center-in-west-virginia/">TechCrunch, Oct 30</a></li><li><strong>Anthropic</strong> — Trading expected to start around November 19 · <a href="https://06.ai/wire/w073">06.ai, W.073</a></li><li><strong>Export rule</strong> — A Commerce rule in preparation covering Singapore, Malaysia, Thailand, Indonesia, Brazil, Norway and South Africa: licences for advanced chips and disclosure of end users; data centers run by US companies exempt on conditions; it could be announced in November · <a href="https://www.forbes.com/sites/jonmarkman/2026/11/12/chinese-ai-firms-rush-to-rent-nvidia-chips-abroad-before-a-new-us-rule/">Forbes, on The Information, Nov 12</a></li></ul>
<h2>Method</h2>
<p><strong>What we read.</strong> Nvidia’s quarterly filing itself, for the three lines and the guarantees; the <a href="https://06.ai/packs/nvidia-after-jensen">Nvidia-after-Huang</a> pack had used others’ transcriptions of it. Company filings and releases for Sharon AI, CoreWeave, Nebius, IREN and Hut 8. Nscale’s prospectus as amended on November 12, read directly; its figures replace those W.019 took from the first filing. Anthropic’s as read for W.084. Firmus, GMI, Lambda and Crusoe are private, and what is here about them is press reporting, named in the table above.</p>
<p><strong>Estimates.</strong> $21.1B, $4.2B and about $2.2B are SemiAnalysis’s counts of Nvidia’s obligations; Nvidia’s filing gives only the $36B total, so $8.5B has no name on it. Sharon AI’s own figure for its agreement is $4.9B, which we use for its shares and its bar. Floors per GPU-hour: $4.9B ÷ (40,000 × 6 × 8,760) = $2.33; $21.1B ÷ (170,000 × 6 × 8,760) = $2.36. Both assume every GPU for the whole term. The filing names no counterparty in any line. Beside the three lines it lists $29B of cloud service agreements, compute Nvidia rents for itself (the <a href="https://06.ai/packs/nvidia-after-jensen">Nvidia-after-Huang</a> pack). We do not know which line holds CoreWeave’s order or what Nvidia rents from IREN, Lambda and Nscale, so those come from the companies’ own disclosures and press reports and are not added to the $59.5B.</p>
<p><strong>Shares.</strong> $4.9B ÷ $8.8B = 55.7%; Sharon AI’s release lists the agreement beside three other contracts worth $2.6B and does not break the total down, so we take the $4.9B to be inside it. $3.4B ÷ $20.3B = 16.7%. $6.3B ÷ $132.5B = 4.8%, and ÷ $46.9B = 13.4%; we do not know whether Nvidia’s order is inside CoreWeave’s reported backlog. $1.5B ÷ $50B = 3%; $34.9B ÷ $50B = 69.8%, which sets a figure from Anthropic’s filing against a backlog reported from a letter to investors. $1.2B ÷ $105.8B = 1.1%, taking Nvidia’s agreement to be inside Nscale’s total. $860.3M ÷ about $1.95B = 44.1%. $2.0B ÷ $3.36B = 59.5%; $777M ÷ about $2.1B = 37%. $2.0B ÷ $16.8B = 11.9%. $2.0B ÷ $31.7B = 6.3%, on the <a href="https://06.ai/packs/neocloud-capex-funding">neocloud</a> pack’s tally of 2026 financings to October 16.</p>
<p><strong>Beacon Point.</strong> That Nvidia’s $20B of leases is Hut 8’s campus is our reading. The amounts ($19.6B), the term (15 years) and the start dates match, and the FT and the Journal report Nvidia as the tenant. Neither company has said so, and Anthropic’s filing does not name the site whose lease Lambda missed; press reports place it at Beacon Point. Why Lambda missed the date is not public.</p>
<p><strong>Not covered.</strong> Who gets Vera Rubin first, which may matter more to these companies than any of the above (<a href="https://06.ai/packs/blackwell-per-dollar">the </a><a href="https://06.ai/packs/blackwell-per-dollar">Blackwell</a><a href="https://06.ai/packs/blackwell-per-dollar"> pack</a>). The $105B guarantee for OpenAI’s Ohio campus and Nvidia’s stakes in the labs (the <a href="https://06.ai/packs/nvidia-after-jensen">Nvidia-after-Huang</a> pack). Terms of private agreements we could not see. Whether any of the capacity deals paused in August has resumed.</p>
<hr>
<p>Every chart above links to its interactive version on 06.ai, with the data behind it. The daily trackers are at <a href="https://06.ai/trackers">06.ai/trackers</a>.</p>
<p><em>Not investment advice. Estimates are labeled as estimates.</em></p>]]></content:encoded>
</item>
<item>
  <title>How big is AI revenue when each dollar is counted once?</title>
  <link>https://inthesingularity.ai/packs/ai-revenue</link>
  <guid isPermaLink="true">https://inthesingularity.ai/packs/ai-revenue</guid>
  <pubDate>Thu, 12 Nov 2026 12:27:00 GMT</pubDate>
  <description>Aswath Damodaran puts all revenue from AI products and services at about $250B a year, and Sam Altman is reported to have said AI demand has been double-counted. Added up the way companies report it, AI revenue ran at $456B a year in the third quarter; $161B of that was one AI company paying another. Counted once it was $295B, and we estimate about $369B by the end of 2026 and $716B by the end of 2027, in a range of $507B to $929B. More than half of the 2027 figure is OpenAI and Anthropic.</description>
  <category>AI revenue</category>
  <enclosure url="https://inthesingularity.ai/newsletter/ai-revenue/cover.png" length="61635" type="image/png" />
  <media:content url="https://inthesingularity.ai/newsletter/ai-revenue/cover.png" medium="image" type="image/png" />
  <content:encoded><![CDATA[<figure><a href="https://06.ai/packs/ai-revenue"><img src="https://06.ai/newsletter/ai-revenue/thumb.png" alt="How big is AI revenue when each dollar is counted once?" width="421" height="391"></a><figcaption><a href="https://06.ai/packs/ai-revenue">Interactive version and data on 06.ai →</a></figcaption></figure>
<p>Aswath Damodaran puts all revenue from AI products and services at about $250B a year. Sam Altman is reported to have said AI demand has been double-counted. Added up the way companies report it, AI revenue ran at <strong>$456B a year in the third quarter, and $161B of that was one AI company paying another</strong>. Counted once it was $295B. We estimate about <strong>$369B by the end of 2026 and $716B by the end of 2027</strong>, in a range of $507B to $929B. More than half of the 2027 figure is OpenAI and Anthropic.</p>
<ul><li><strong>$369B</strong> — Counted once, end of 2026. <strong>$287B</strong> sold as AI, plus $82B of AI in Meta’s and Google’s ads. Third quarter: $295B</li><li><strong>$716B</strong> — Counted once, end of 2027. Range $507B–929B. OpenAI and Anthropic are <strong>54%</strong> of it</li><li><strong>$456B</strong> — Added up without netting, Q3. <strong>$161B</strong>, 35%, is one AI company paying another</li><li><strong>33%</strong> — Cloud AI and rented compute that survives. <strong>$70B</strong> of $212B a year. The rest is paid by the labs or sold on</li></ul>
<h2>Two numbers that don’t match the filings</h2>
<p>On November 9 Aswath Damodaran, who teaches valuation at NYU, wrote that revenue from AI products and services is “about $250 billion” a year, even on the run-rates the labs leak. He used the same figure in August, when it included more than $65B for Anthropic and about $40B for OpenAI. Two days after his post, Korea’s Financial News reported that Sam Altman had told Samsung executives “there has been double-counting of demand” (<a href="https://06.ai/wire/w091">W.091</a>).</p>
<p>Company filings point the other way. Microsoft, Amazon and Alphabet reported cloud growth of 39% to 103%, most of it AI by the measure below. Oracle, CoreWeave and SpaceX rent out compute by the gigawatt. Meta says AI is already adding to its ad revenue. Add every line and the total is far above $250B. The lines overlap, though. A customer pays Anthropic; Anthropic pays Amazon and SpaceX for compute; all three book revenue.</p>
<p>This pack counts each dollar once, where it comes in from a customer that is not itself selling AI. That leaves four layers: what customers pay the labs, what they pay the big clouds for AI directly, what they pay other software companies for AI products, and the extra ad revenue Meta and Google earn because AI ranks their ads. Chips, servers, power and buildings are left out altogether. They are the capex, which this revenue has to pay for.</p>
<h2>Added up, $456B a year. Counted once, $295B</h2>
<figure><a href="https://06.ai/packs/ai-revenue"><img src="https://06.ai/newsletter/ai-revenue/fig-01.png" alt="Added up, $456B a year. Counted once, $295B" width="694" height="285"></a><figcaption>Third quarter of 2026 at a yearly rate, $B. Solid: revenue from customers outside the AI supply chain, which we keep. Dashed: revenue paid by another company in the tally, which we take out. Labs: OpenAI’s and Anthropic’s third-quarter revenue times four. Clouds: AI revenue at Azure, AWS and Google Cloud, each against its no-AI trend. Other compute: Oracle’s cloud infrastructure, CoreWeave and what SpaceX sells Anthropic, latest quarter times four. Software: Menlo Ventures’ September estimate. Ads: calendar 2026.</figcaption></figure>
<h2>The labs are counted in full</h2>
<p>Anthropic’s amended S-1 puts third-quarter revenue at $16.96B. OpenAI told investors its own was about $10B. Together that is $108B a year. Both have grown since. Anthropic’s run-rate was about $81B on October 31, up $10B in a month, and our <a href="https://06.ai/trackers/openai-anthropic-revenue">revenue tracker</a> had the two at $143B on November 11. It puts them at <strong>$167B at the end of December</strong> (Anthropic $94B, OpenAI $73B) and <strong>$385B at the end of 2027</strong> (Anthropic $185B, OpenAI $199B).</p>
<p>The two are not on the same basis. Anthropic books the full price of Claude sold through Amazon’s, Google’s and Microsoft’s marketplaces, which carried 44% of its revenue in the first nine months, and expenses the clouds’ cut. OpenAI books only its share of some sales through Microsoft. We leave both as reported. Anthropic’s figure is what customers paid, and the clouds’ cut comes out of the cloud layer below.</p>
<h2>Two-thirds of the compute revenue comes from inside the industry</h2>
<p>The <a href="https://06.ai/packs/ai-capex-payback">ad-payback</a> and <a href="https://06.ai/packs/cloud-capex-payback">cloud-payback</a> packs measured AI revenue at Google Cloud, Azure and AWS by setting each against an estimate of what it would have grown without AI (22%, 17% and 14.5% a year) and crediting 90% of the gap to AI. With the third quarter in, that is $13.9B at Google Cloud, $12.8B at Azure and $10.2B at AWS for the quarter, <strong>$148B a year</strong>.</p>
<p><strong>Microsoft’s 10-Q says how much of Azure’s is one customer.</strong> It recorded $10.0B of “revenue from commercial arrangements with OpenAI, inclusive of revenue-sharing payments” in the quarter, about 11% of its revenue. That is 78% of Azure’s AI revenue as we measure it, if all of it sits in Azure, and as much as OpenAI’s own revenue for the quarter.</p>
<p><strong>Anthropic’s filing gives a ceiling for the other two.</strong> It spent $9.86B on compute in the quarter and $2.36B on sales, marketing and partnerships, a line that in the first half was mostly fees to channel partners, chiefly the cloud marketplaces. The two come to 72% of its revenue. SpaceX’s 10-Q shows one customer of its AI segment at 34.5% of revenue, about $5.03B, and Bloomberg has reported that Anthropic is SpaceX’s largest customer. That is 67% of the segment’s $7.56B, and it leaves $4.83B of Anthropic’s compute bill for every other supplier. If all of that and all of the sales and marketing line went to Amazon and Google, $7.20B together, it would be 30% of their $24.1B of AI revenue. OpenAI’s $38B contract with AWS, spread evenly over its seven years, adds $1.4B a quarter and takes that to 35%. We take a third. Across the three clouds that makes the two labs about half of AI revenue (49%).</p>
<p><strong>Google Cloud needs one more adjustment.</strong> Alphabet began booking sales of TPU systems to other data center operators in the second quarter and called them a meaningful contributor in the third, without giving a figure. That is hardware, the same kind of revenue as Nvidia’s. We take out a further 10% of Google Cloud’s AI revenue for it, which may be low: Alphabet’s inventory rose from $2.4B to $10.0B in the first half of the year.</p>
<p><strong>Oracle, CoreWeave and SpaceX’s compute sales are left out entirely.</strong> Oracle’s cloud infrastructure brought in $7.39B in its August quarter, and its backlog includes a contract with OpenAI reported at about $300B. Three customers were 72% of CoreWeave’s second-quarter revenue. Whatever these companies sell to enterprises and governments directly is revenue we do not count, so our total is a little low. Of the $212B a year of cloud AI and rented compute in the tally, <strong>$70B survives</strong>.</p>
<h2>Who paid for the compute in the third quarter</h2>
<figure><a href="https://06.ai/packs/ai-revenue"><img src="https://06.ai/newsletter/ai-revenue/fig-02.png" alt="Who paid for the compute in the third quarter" width="694" height="252"></a><figcaption>$B in the quarter. Pink: paid by OpenAI or Anthropic. Azure’s is Microsoft’s disclosed revenue from OpenAI; SpaceX’s is the one customer in its 10-Q; the third at AWS and Google Cloud is our estimate from Anthropic’s filing. Dashed: Google’s TPU systems, which are hardware, our estimate. Strong grey: everyone else, the part we keep. Faint grey: not split by customer and not kept.</figcaption></figure>
<h2>Software sold outside the labs</h2>
<p>Menlo Ventures estimates that enterprises were spending $176.8B a year on using AI in September, 5.4 times last year’s $33B. Of that, $122.7B went to Anthropic, OpenAI and Google, which sit in the first two layers here. The other <strong>$54.1B</strong> went to everyone else: $37.3B for applications (about $10B for coding tools other than Claude Code and Codex, $9.5B for assistants such as Microsoft’s Copilot, $8.8B for other horizontal software and $10.1B for vertical software) and $16.8B for other models, inference platforms and agent infrastructure.</p>
<p>These vendors buy models from the labs and compute from the clouds, so part of their revenue is already counted above. We keep 65%, or $35B. That share is a judgment. A coding tool that resells tokens keeps less of what it charges; a vertical application with its own workflow keeps more.</p>
<h2>The ads</h2>
<p>The <a href="https://06.ai/packs/ai-capex-payback">ad-payback</a> pack put AI’s part of Meta’s ad revenue at $55B in 2026, 22% of it, and of Google’s at $29.5B, using the same kind of trend. Meta’s figure already had its third quarter in it. Google’s did not. Search grew 12%, where we had pencilled in 15%, and added $79M on the second quarter even as Alphabet reported its fastest query growth in five years. With that quarter in, Google’s figure is $27B, 8% of its ad revenue.</p>
<p>Nobody buys this revenue as AI. Advertisers pay for clicks, and the estimate turns on what Meta and Google would have grown without the new ranking models. We include it because it is the return on the largest AI capex outside the clouds, and we show the total without it throughout: $213B in the third quarter and $287B at the end of 2026.</p>
<h2>About $369B by the end of 2026 and $716B by the end of 2027</h2>
<figure><a href="https://06.ai/packs/ai-revenue"><img src="https://06.ai/newsletter/ai-revenue/fig-03.png" alt="About $369B by the end of 2026 and $716B by the end of 2027" width="694" height="383"></a><figcaption>AI revenue counted once, $B a year. The third quarter is reported figures times four; the end of 2026 is our estimate seven weeks out; the end of 2027 is shown for three cases. The ads layer is a calendar-year figure (2026 for the first two bars, 2027 for the rest), because ad revenue is seasonal.</figcaption></figure>
<figure><a href="https://06.ai/packs/ai-revenue"><img src="https://06.ai/newsletter/ai-revenue/fig-04.png" alt="OpenAI and Anthropic — Q3 2026: $108B · End of 2026: $167B · End of 2027, low: $283B · Base: $385B · High: $489B; AI at the three clouds, outside the labs — Q3 2026: $70B · End of 2026: $81B · End of 2027, low: $89B · Base: $153B · High: $209B; AI software outside the labs — Q3 2026: $35B · End of 2026: $39B · End of 2027, low: $42B · Base: $66B · High: $98B; Sold as AI — Q3 2026: $213B · End of 2026: $287B · End of 2027, low: $414B · Base: $603B · High: $796B; AI in Meta’s and Google’s ads — Q3 2026: $82B · End of 2026: $82B · End of 2027, low: $94B · Base: $113B · High: $133B; Counted once — Q3 2026: $295B · End of 2026: $369B · End of 2027, low: $507B · Base: $716B · High: $929B" width="694" height="365"></a></figure>
<h2>What 2027 rests on</h2>
<p>The end-2027 figure is 1.9 times the end-2026 one. Of the $347B added, 63% comes from the two labs.</p>
<p><strong>Labs.</strong> The tracker carries each lab’s recent monthly growth forward and halves it every 18 months. On that path Anthropic books about $138B of revenue in 2027. Goldman Sachs, the lead bank on its IPO, is telling investors $136B, the FT reported. OpenAI’s path, about $130B, is less well supported. Its own projection in February was $62B, a level its December target for run-rate would already pass. FutureSearch, an outside forecaster, put its run-rate at $92B in June 2027 when it last updated on October 16, with a 10th-to-90th-percentile range of $63.5B to $132.5B. The tracker has $127B by then, near the top of that range. The low case halves both labs’ growth every six months, which leaves them at $283B together at the end of 2027. The high case does not fade it at all, $489B. The tracker’s own ranges are wider, $262B to $668B if both labs land at the same end.</p>
<p>For Anthropic, compute does not stop the base case. On the <a href="https://06.ai/packs/anthropic-s1">Anthropic S-1</a> pack’s lowest supply path, 8 GW by the end of 2027 with research keeping its share of the compute, it can serve about $185B a year, the tracker’s central figure. On that pack’s base path, 10 GW, it can serve $231B. We have no equivalent check for OpenAI.</p>
<p><strong>Clouds.</strong> Amazon projects AWS growth of over 40% in both 2027 and 2028. Microsoft guided Azure to 49–50% for the December quarter. Alphabet expects just over 55% of Google Cloud’s $561B backlog to become revenue within 24 months. The base case has the three growing 40%, 40% and 50% in the fourth quarter of 2027, with the labs’ shares unchanged. The low case has them growing 32%, 30% and 35%, with more of it paid by the labs.</p>
<p><strong>Software and ads.</strong> Software outside the labs grows 1.7 times in the base case; Menlo has its segments growing 2.3 to 4.3 times in the year to September. Meta’s and Google’s ad revenue grow 18% and 11%, the <a href="https://06.ai/packs/ai-capex-payback">ad-payback</a> pack’s assumptions.</p>
<h2>Against Damodaran’s $250B</h2>
<p>On November 11 the things sold as AI, counted once, were running at about <strong>$256B a year</strong>: $143B at the labs, with the cloud and software layers halfway between their third-quarter and year-end levels. Damodaran’s figure is about right for today.</p>
<p>It is a rate, and it is moving. The same $250B in August included about $105B for the two labs. They are at $143B now, and Anthropic alone added $10B in October. On our estimate the comparable figure is $287B at the end of December and $603B a year later.</p>
<p>His test for the industry is about $5T of revenue within a decade, 20 times $250B. From $603B it is 8 times.</p>
<h2>Against the capex</h2>
<p>Amazon, Alphabet, Meta and Microsoft have guided to about $730B of capex between them in 2026, and the Street case in the <a href="https://06.ai/packs/blue-wave-fcf">blue-wave cash flow</a> pack has $1.01T in 2027. Revenue counted once is <strong>0.51× the first at the end of 2026 and 0.71× the second at the end of 2027</strong> (0.50× to 0.92× across our cases).</p>
<p>That compares a yearly rate at one date with a year of spending, and a dollar of capex earns for several years, so a ratio under one is not a shortfall by itself. The <a href="https://06.ai/packs/ai-capex-payback">ad-payback</a> and <a href="https://06.ai/packs/cloud-capex-payback">cloud-payback</a> packs work through what the ad and cloud revenue returns on the capex behind it. The ratio does say which way things are going: the revenue is growing faster than the spending, from well under it.</p>
<h2>What would move it</h2>
<ul><li><strong>Compute that arrives late.</strong> Microsoft lowered its calendar-2026 capex to about $170B because third-party capacity was delayed, and the <a href="https://06.ai/packs/texas-data-center-risk">Texas data center</a> pack has Texas switching on 2.38 GW less by the end of 2027 under its incoming governor than on the path the outgoing one had set. Late capacity lowers the cloud layer first and the labs after it.</li><li><strong>Prices.</strong> Anthropic said price reductions and discounts took about $600M off its third-quarter revenue (<a href="https://06.ai/wire/w001">W.001</a>), and it listed Fable 5.2 at 20% below Fable 5.1 (the <a href="https://06.ai/packs/fable-5-2">Fable 5.2</a> pack).</li><li><strong>Budgets.</strong> In Menlo’s survey the average enterprise spent 4.7 times last year’s AI budget against a planned 1.9. A year spent catching budgets up would slow every layer.</li><li><strong>The accounting basis.</strong> If Anthropic booked marketplace sales net, as OpenAI does, the labs layer would fall by the clouds’ cut and the cloud layer would rise by the same amount. The total would not change.</li><li><strong>The trends.</strong> The cloud and ad layers depend on what each business would have grown without AI. At the low and high ends of the ranges the <a href="https://06.ai/packs/ai-capex-payback">ad-payback</a> and <a href="https://06.ai/packs/cloud-capex-payback">cloud-payback</a> packs used, the end-2026 total is $396B or $341B.</li></ul>
<h2>What is not counted</h2>
<p>China’s clouds and labs: Alibaba alone said its AI-related products were selling at an annualized RMB35.8B in the March quarter. Consumer AI outside the two labs, such as Gemini subscriptions and Grok. AI in other ad businesses: AppLovin’s revenue grew 55% in the third quarter, to $2.19B. Compute that Oracle and the neoclouds rent directly to enterprises and governments. Consulting, and robotaxis. Each would raise the total.</p>
<h2>What to watch</h2>
<p><strong>Anthropic’s first report as a public company.</strong> It expects to list next week. Whether it keeps publishing run-rate, and what share of revenue goes through the clouds’ marketplaces.</p>
<p><strong>OpenAI’s December run-rate</strong> against its target of at least $70B, which the tracker’s year-end figure follows.</p>
<p><strong>Microsoft’s next 10-Q,</strong> in late January: the line for revenue from OpenAI, $10.0B this quarter.</p>
<p><strong>Whether Alphabet puts a number on TPU system sales,</strong> which would settle how much of Google Cloud’s growth is hardware.</p>
<p><strong>AWS’s fourth quarter</strong> against Amazon’s projection of more than 40% growth in 2027.</p>
<h2>What this is anchored on</h2>
<p>Company figures are as reported or as the filings’ readers reported them. The no-AI trends are those of the <a href="https://06.ai/packs/ai-capex-payback">ad-payback</a> and <a href="https://06.ai/packs/cloud-capex-payback">cloud-payback</a> packs. The labs’ shares of cloud revenue, the share software vendors keep and every 2027 growth rate are ours.</p>
<ul><li><strong>Damodaran</strong> — Revenue from AI products and services “about $250 billion”, on the run-rates leaked by the leading companies · <a href="https://aswathdamodaran.substack.com/p/deconstructing-the-ai-trade-ai-builders">Musings on Markets, Nov 9</a></li><li><strong>Damodaran</strong> — August: no more than $250B, including more than $65B at Anthropic and about $40B at OpenAI; about $5T of revenue needed within a decade, about 20 times that · <a href="https://mitsloanindia.com/article/damodaran-puts-ai-valuations-through-a-5-trillion-revenue-test/">MIT Sloan Management Review India</a></li><li><strong>Altman</strong> — Remarks to Samsung executives on Oct 31, as Financial News reported them on Nov 11 · <a href="https://06.ai/wire/w091">06.ai W.091</a></li><li><strong>Labs</strong> — Run-rates on Nov 11: Anthropic $83.5B, OpenAI $59.6B; December 2026: $94.1B and $72.5B; December 2027: $185.2B ($123B–369B) and $199.3B ($140B–299B); the paths with growth halving every 6 months or not at all · <a href="https://06.ai/trackers/openai-anthropic-revenue">06.ai revenue tracker, run of Nov 11</a></li><li><strong>Anthropic</strong> — Third quarter: revenue $16,956M; compute and infrastructure $9,863M; sales, marketing and partnerships $2,361M; price reductions and discounts took about $600M off revenue · <a href="https://06.ai/wire/w001">06.ai W.001, from the S-1/A of Nov 2</a></li><li><strong>Anthropic</strong> — Cloud marketplaces carried 48% of revenue in the first half and 44% in the first nine months · <a href="https://filinginsights.substack.com/p/anthropics-amended-s-1-adds-a-third">Filing Insights, on the S-1/A</a></li><li><strong>Anthropic</strong> — Marketplace sales are booked gross, as principal; the clouds remit net of their fees · <a href="https://www.sec.gov/Archives/edgar/data/0002130357/000162828026067168/anthropicpbc-sx1.htm">Anthropic S-1, Oct 26</a></li><li><strong>Anthropic</strong> — First half: sales, marketing and partnerships $3.168B, including a $2.7B increase in fees paid to channel partners · <a href="https://ppc.land/amazon-google-and-microsoft-hold-269bn-of-anthropic-compute-contracts/">PPC Land, on the S-1</a></li><li><strong>Anthropic</strong> — Run-rate about $81B on Oct 31, up $10B from Sep 30 · <a href="https://www.marketscreener.com/news/anthropic-eyes-1-4-trillion-ipo-valuation-in-test-of-wall-street-s-ai-fervor-ce785edadd8bf32d">Reuters, via MarketScreener, on the S-1/A of Nov 10</a></li><li><strong>Anthropic</strong> — Goldman Sachs, lead bank on the IPO: revenue of about $52B in 2026 and $136B in 2027 (FT) · <a href="https://investinglive.com/stocks/anthropic-looks-to-be-pricing-its-ipo-for-success/">investingLive, Nov 11</a></li><li><strong>Anthropic</strong> — Commitments by supplier; the Google, Amazon and Microsoft minimums · <a href="https://06.ai/wire/w091">06.ai W.084 and W.091</a></li><li><strong>OpenAI</strong> — Third-quarter revenue about $10B, up about 50% on the second; run-rate about $50B at the end of September. Anthropic expects to list next week · <a href="https://06.ai/wire/w079">06.ai W.079, from the WSJ and Bloomberg</a></li><li><strong>OpenAI</strong> — Books only its share of some sales through Microsoft; Anthropic books marketplace sales in full · <a href="https://www.thesaascfo.com/openai-vs-anthropic-arr/">The SaaS CFO</a></li><li><strong>OpenAI</strong> — February projections: revenue of $30B in 2026 and $62B in 2027 · <a href="https://mlq.ai/news/openai-revises-projections-upward-with-112-billion-extra-cash-burn-by-2030/">MLQ, on The Information</a></li><li><strong>OpenAI</strong> — Forecast of its run-rate in June 2027: median $92B, 10th percentile $63.5B, 90th $132.5B, from about $50B at the end of September; last updated Oct 16 · <a href="https://futuresearch.ai/openai-financial-forecast/">FutureSearch</a></li><li><strong>Microsoft</strong> — “revenue from commercial arrangements with OpenAI, inclusive of revenue-sharing payments, of $10.0 billion” in the quarter to Sep 30; receivables from OpenAI $7.2B · <a href="https://www.sec.gov/Archives/edgar/data/0000789019/000119312526436133/msft-20260930.htm">Microsoft 10-Q, Oct 28</a></li><li><strong>Microsoft</strong> — That is about 11% of revenue; $24.1B on the same measure in the twelve months to June · <a href="https://thenextweb.com/news/microsoft-openai-revenue-10bn-quarter">The Next Web, on the 10-Q</a></li><li><strong>Microsoft</strong> — Azure revenue $33.17B, up 48%, the first figure in dollars; December quarter guided to 49–50% in constant currency · <a href="https://siliconangle.com/?p=852356">SiliconANGLE, Oct 28</a></li><li><strong>Microsoft</strong> — Calendar-2026 capex about $170B, lower because third-party capacity was delayed · <a href="https://walnutinvest.com/stats/ai-capex-statistics">Walnut, on the Oct 28 call</a></li><li><strong>Amazon</strong> — AWS sales $46.0B, up 39.3%; a $184B run-rate; growth projected at over 40% in both 2027 and 2028 · <a href="https://www.telecomtv.com/content/digital-platforms-services/ai-drives-cloud-services-sector-growth-to-55-in-q3-56566/amp/">TelecomTV, on Amazon’s release</a></li><li><strong>Alphabet</strong> — Google Cloud $30.71B from $15.16B; Search &amp; other $63.35B; YouTube ads $11.21B; Network $7.46B; Google advertising $82.01B from $74.18B · <a href="https://www.verdict.co.uk/?p=404597">Verdict, on Alphabet’s release</a></li><li><strong>Alphabet</strong> — Cloud backlog $561B, just over 55% of it expected as revenue within 24 months; TPU system sales a meaningful contributor · <a href="https://marketbeat.com/instant-alerts/transcript-alphabet-q3-earnings-call-highlights-2026-10-29/">MarketBeat, Q3 call highlights</a></li><li><strong>Alphabet</strong> — TPU system sales first booked in the second quarter; no figure given; inventory $10.0B at Jun 30, from $2.4B at the end of 2025 · <a href="https://www.tikr.com/blog/alphabets-tpus-now-rank-second-only-to-nvidia-in-compute-built-google-cloud-shows-it">TIKR, on the 10-Q</a></li><li><strong>Alphabet</strong> — Search revenue rose $79M on the second quarter while query growth was the fastest in five years · <a href="https://ppc.land/google-search-sales-gain-just-79m-on-q2-as-query-growth-hits-5-year-high/">PPC Land, Oct 28</a></li><li><strong>SpaceX</strong> — Third-quarter revenue $14.58B, AI segment $7.56B; one customer of that segment was 34.5% of revenue; Bloomberg reported that Anthropic is SpaceX’s largest customer · <a href="https://fourweekmba.com/ai-one-ai-customer-made-up-34-5-of-spacexs-q3-revenue/">FourWeekMBA, on the 10-Q; 06.ai W.043</a></li><li><strong>CoreWeave</strong> — Third-quarter revenue $3,661M; backlog about $132.5B · <a href="https://06.ai/wire/w076">06.ai W.076</a></li><li><strong>CoreWeave</strong> — Second quarter: three customers were 36%, 26% and 10% of revenue · <a href="https://sec-api.io/insights/coreweave-crwv-2026-revenue-analysis-top-customer-concentration-exposure">sec-api.io, on the 10-Q</a></li><li><strong>Oracle</strong> — Cloud infrastructure revenue $7.39B in the quarter to Aug 31, up 121%; remaining performance obligations $664B · <a href="https://investor.oracle.com/investor-news/news-details/2026/Oracle-Announces-Q1-Results-Driven-by-Triple-Digit-Growth-in-Cloud-Infrastructure-Revenues/default.aspx">Oracle, first-quarter fiscal 2027 results</a></li><li><strong>Oracle</strong> — OpenAI’s contract reported at about $300B, not confirmed by Oracle · <a href="https://06.ai/wire/w061">06.ai W.061</a></li><li><strong>Amazon</strong> — OpenAI’s AWS agreements: $38B over seven years (November 2025) and $100B over eight (February 2026) · <a href="https://06.ai/wire/w061">06.ai W.061</a></li><li><strong>Software</strong> — Enterprise spending on using AI $176.8B (September 2026, annualized; $33B in 2025): applications $77.3B, infrastructure $99.5B; $122.7B at Anthropic, OpenAI and Google, $40.0B of it their own applications; coding $34.0B, of which Claude Code and Codex $24.0B; assistants outside the labs $9.5B; other horizontal $8.8B; vertical $10.1B; the average enterprise spent 4.7 times last year’s AI budget against a planned 1.9 · <a href="https://menlovc.com/perspective/2026-the-state-of-enterprise-ai/">Menlo Ventures, 2026: The State of Enterprise AI</a></li><li><strong>Meta</strong> — AI investments “already delivering substantial incremental advertising revenue”; fourth-quarter revenue guided to $73–76B · <a href="https://investor.atmeta.com/investor-news/press-release-details/2026/Meta-Reports-Third-Quarter-2026-Results/default.aspx">Meta, third-quarter 2026 results</a></li><li><strong>Ads</strong> — Method, trends and Meta’s 2026 figure (nine months reported, the fourth quarter at the midpoint of $73–76B guidance) · <a href="https://06.ai/packs/ai-capex-payback">06.ai </a><a href="https://06.ai/packs/ai-capex-payback">ad-payback</a><a href="https://06.ai/packs/ai-capex-payback"> pack</a></li><li><strong>Clouds</strong> — Method, trends and the quarterly Azure and AWS series · <a href="https://06.ai/packs/cloud-capex-payback">06.ai </a><a href="https://06.ai/packs/cloud-capex-payback">cloud-payback</a><a href="https://06.ai/packs/cloud-capex-payback"> pack</a></li><li><strong>Capex</strong> — 2026 guidance: Amazon about $220B, Alphabet $195–205B, Meta $135–145B, Microsoft about $170B. 2027, Street case: Alphabet $300B, Amazon $290B, Microsoft $215B, Meta $205B · <a href="https://06.ai/packs/blue-wave-fcf">06.ai </a><a href="https://06.ai/packs/blue-wave-fcf">blue-wave cash flow</a><a href="https://06.ai/packs/blue-wave-fcf"> pack</a></li><li><strong>Compute</strong> — Anthropic’s supply paths to 2029 and the revenue each supports; Texas gigawatts switched on later · <a href="https://06.ai/packs/anthropic-s1">06.ai </a><a href="https://06.ai/packs/anthropic-s1">Anthropic S-1</a><a href="https://06.ai/packs/anthropic-s1"> and </a><a href="https://06.ai/packs/anthropic-s1">Texas data center</a><a href="https://06.ai/packs/anthropic-s1"> packs</a></li><li><strong>Not counted</strong> — Alibaba: AI-related product revenue at an annualized RMB35.8B in the March quarter, the eleventh of triple-digit growth · <a href="https://nationalcioreview.com/articles-insights/cio-field-notes/alibaba-reports-ai-driven-cloud-growth-as-agent-workloads-rise/">National CIO Review</a></li><li><strong>Not counted</strong> — AppLovin: third-quarter revenue $2,185M, up 55% · <a href="https://www.sec.gov/Archives/edgar/data/0001751008/000175100826000068/exhibit991-3q26earningspre.htm">AppLovin 8-K, Nov 4</a></li></ul>
<h2>Method</h2>
<p><strong>What counts.</strong> Revenue from a customer that is not itself in the tally. A run-rate is a quarter’s revenue times four, except for the labs at the two year-ends, where it is the tracker’s estimate (Anthropic’s own measure is the last 28 days of consumption times 13 plus monthly subscriptions times 12), and for ads, which are calendar years. Suppliers’ revenue (chips, servers, power, construction) is not AI revenue here.</p>
<p><strong>Labs.</strong> Third quarter: Anthropic $16.96B (S-1/A) plus OpenAI about $10B (investor documents, as reported), times four. Year-ends: the revenue tracker’s run of November 11, frozen, so this page does not move when the tracker does. Its December 2026 figure for OpenAI follows OpenAI’s own target. The low and high cases for 2027 are the tracker’s paths with monthly growth halving every 6 months or not fading.</p>
<p><strong>Clouds.</strong> AI revenue in a quarter is 90% of revenue above a no-AI path: for AWS and Google Cloud the same quarter of 2023 grown at 14.5% and 22% a year, for Azure the non-AI revenue in Microsoft’s own split for 2024 grown at 17%. These are the methods and rates of the <a href="https://06.ai/packs/cloud-capex-payback">cloud-payback</a> and <a href="https://06.ai/packs/ai-capex-payback">ad-payback</a> packs, applied by quarter. What we keep: at Azure in the third quarter, AI revenue less Microsoft’s $10.0B from OpenAI, and at the year-ends 20% of it (10% and 30% in the low and high cases); at AWS and Google Cloud together, two-thirds (55% and 75%), less 10% of Google Cloud’s for TPU systems sold as hardware (20% and 5%). OpenAI’s $100B AWS agreement of February 2026 has no schedule and is left out of the arithmetic for the labs’ share. Microsoft’s figure includes revenue-sharing payments and the filing does not say which line it sits in. OpenAI’s share of Microsoft’s sales of its models is in both OpenAI’s revenue and Azure’s; we do not adjust for it. Oracle’s figure is all of its cloud infrastructure revenue, not only AI, and appears only in the sum.</p>
<p><strong>Software.</strong> Menlo’s $77.3B of applications less the $40.0B that is the three labs’ own, plus its $99.5B of infrastructure less the $82.7B of the three labs’ other revenue ($122.7B less $40.0B): $54.1B. Menlo’s categories do not sum exactly to its totals; we use the totals. We keep 65% (50% and 75% in the low and high cases), raise it 10% from September to December, and grow it 1.7 times in 2027 (1.4 and 2.2).</p>
<p><strong>Ads.</strong> AI’s part is 80% of Meta’s ad revenue above 2023’s grown at 11% a year, and 75% of Google’s above 2023’s grown at 7.5%, as in the <a href="https://06.ai/packs/ai-capex-payback">ad-payback</a> pack. Meta’s 2026 is nine months reported plus the fourth quarter at the middle of its guidance. Google’s is three quarters reported plus a fourth with Search up 11%, YouTube 9% and Network 1%, our estimate. In 2027 Meta’s ad revenue grows 18% (12% and 24%) and Google’s 11% (8% and 14%).</p>
<p><strong>Cases.</strong> The low and high cases move every assumption the same way at once, so they are wider than a range in which the layers vary independently. The no-AI trends are held at their central values in all three; their effect is given under “What would move it”.</p>
<p><strong>Against earlier packs.</strong> The <a href="https://06.ai/packs/ai-capex-payback">ad-payback</a> and <a href="https://06.ai/packs/cloud-capex-payback">cloud-payback</a> packs were written before Alphabet, Microsoft and Amazon reported the third quarter. With those quarters in, the same method gives calendar-2026 AI revenue of $41.5B at Azure (it was $39.4B), $35.5B at AWS ($31.0B), $45.4B at Google Cloud ($38.0B) and $26.8B in Google’s ads ($29.5B). The <a href="https://06.ai/packs/cloud-capex-payback">cloud-payback</a> pack took the labs’ share of AWS’s AI revenue as 50% for its stress cases; the arithmetic above suggests about a third for AWS and Google Cloud together. Neither pack’s conclusions are re-run here.</p>
<hr>
<p>Every chart above links to its interactive version on 06.ai, with the data behind it. The daily trackers are at <a href="https://06.ai/trackers">06.ai/trackers</a>.</p>
<p><em>Not investment advice. Estimates are labeled as estimates.</em></p>]]></content:encoded>
</item>
<item>
  <title>How long can Nvidia run without Jensen Huang?</title>
  <link>https://inthesingularity.ai/packs/nvidia-after-jensen</link>
  <guid isPermaLink="true">https://inthesingularity.ai/packs/nvidia-after-jensen</guid>
  <pubDate>Thu, 12 Nov 2026 12:26:00 GMT</pubDate>
  <description>Jensen Huang resigned on November 12 and Nvidia’s finance chief is running the company. About two years of it are already decided: revenue for the year to January 2028 is capped by supply at about 70% growth, three chip generations are defined and $279B of supply is on order. What decays first is what Huang did for everyone else: rationing chips, lending Nvidia’s balance sheet and dealing with governments. At about $165 the shares cost about 10 times next fiscal year’s expected earnings, against 15 at October’s record.</description>
  <category>Nvidia</category>
  <enclosure url="https://inthesingularity.ai/newsletter/nvidia-after-jensen/cover.png" length="62050" type="image/png" />
  <media:content url="https://inthesingularity.ai/newsletter/nvidia-after-jensen/cover.png" medium="image" type="image/png" />
  <content:encoded><![CDATA[<figure><a href="https://06.ai/packs/nvidia-after-jensen"><img src="https://06.ai/newsletter/nvidia-after-jensen/thumb.png" alt="How long can Nvidia run without Jensen Huang?" width="421" height="391"></a><figcaption><a href="https://06.ai/packs/nvidia-after-jensen">Interactive version and data on 06.ai →</a></figcaption></figure>
<p>Jensen Huang resigned on November 12 and Nvidia’s finance chief is running the company. <strong>About two years</strong> of it are already decided: revenue for the year to January 2028 is capped by supply at about <strong>70%</strong> growth, three chip generations are defined and <strong>$279B</strong> of supply is on order. What decays first is what Huang did for everyone else: rationing chips, lending Nvidia’s balance sheet (<strong>$189.5B</strong> committed for customers, with about $240B more under discussion) and dealing with governments. At about <strong>$165</strong> the shares cost <strong>10.4 times</strong> next fiscal year’s expected earnings, against 15.0 at October’s record. That price fits a plateau in AI spending or a botched succession, but not both.</p>
<ul><li><strong>−31%</strong> — Since the October 6 record. $239.24 to about <strong>$165</strong>; about $1.79T of market value</li><li><strong>10.4×</strong> — Price to next year’s expected earnings. <strong>15.0×</strong> at the record; analysts’ new targets imply 12.6–14.1×</li><li><strong>≈70%</strong> — Next year’s revenue growth, capped by supply. Nvidia’s outlook for the year to January 2028; it says demand is nearer <strong>double</strong></li><li><strong>$530.5B</strong> — Commitments and guarantees. <strong>$189.5B</strong> of it for customers; about $240B more was not yet signed</li></ul>
<h2>What happened</h2>
<p><strong>Nvidia said on Thursday morning that Jensen Huang had resigned as president, chief executive and director, effective immediately.</strong> Colette Kress, finance chief since 2013, becomes interim chief executive and keeps the finance job. Harvey Jones, a director since 1993, becomes the company’s first board chair. The board will consider internal and external candidates and gave no timetable. It gave no reason either. The resignation followed two weeks of reporting on Huang’s personal conduct and a board inquiry, opened on November 8, into reports that he had been the target of a foreign intelligence effort; <a href="https://06.ai/wire/w054">W.054</a> and <a href="https://06.ai/wire/w087">W.087</a> set out what was reported. This pack adds nothing to those reports. It is about what Nvidia, and the companies built around it, lose.</p>
<p><strong>The shares opened 13% lower and won back more than half of it.</strong> They opened at $150.99, touched $148.80 and were near $165 by late morning, 5% below Wednesday’s $173.73 close and 31% below the record close of October 6. On about 24.1B shares that is about $1.79T of market value in five weeks.</p>
<p><strong>More than half of that is Nvidia’s own.</strong> Chip stocks fell together after the midterm elections (<a href="https://06.ai/wire/w026">W.026</a>) and then recovered: from November 6, the last close before the board announced its inquiry, to Wednesday the PHLX chip index rose about 12% while Nvidia fell 7%. Had Nvidia kept pace it would have closed Wednesday near $210. At $165 it is 21% below that, about $45 a share or $1.08T. An analyst once put the cost of losing Steve Jobs at a quarter of Apple’s value. The market has charged Nvidia about that.</p>
<h2>Nvidia’s share price since June</h2>
<figure><a href="https://06.ai/packs/nvidia-after-jensen#price"><img src="https://06.ai/newsletter/nvidia-after-jensen/fig-01.png" alt="Nvidia’s share price since June" width="694" height="322"></a><figcaption>Daily closes, with about $165 on the morning of November 12 (open circle). The dashed grey line is where Nvidia’s November 6 close would have gone had it moved with the PHLX chip index to November 11. The New York Times first reported on October 29 that US officials believed Huang had been targeted.</figcaption></figure>
<h2>Five jobs, decaying at different speeds</h2>
<p><strong>Nvidia has no second-in-command.</strong> An internal chart seen by Business Insider in October 2025 showed 36 people reporting to Huang, down from 55, and accounts this week put the number near 60; he has said he avoids one-to-one meetings so that everyone hears the same thing. That design puts every cross-company decision in one head. Apple in 2011 passed to an operating chief who had already run the company through three medical leaves. Nvidia passes, for now, to its finance chief. Its annual report names no key person. It says the company is highly dependent on its longstanding executive team and lists a failure of succession planning among its risks.</p>
<p><strong>He did at least five jobs, and they do not wear off at the same rate.</strong> The two that set Nvidia’s own results for the next two years, the roadmap and the supply chain, are the slowest to fade because they are written into tape-outs and purchase orders. The three that mattered most to everyone else, deciding who gets chips, who gets Nvidia’s credit and what governments will allow, were personal and stop now. Allocation was personal enough that Larry Ellison has described asking Huang for chips over dinner.</p>
<h2>What Huang did, and how long each part lasts without him</h2>
<ul><li><strong>Architect</strong> — What it was: A new architecture every year; 36 people reporting to him and no deputy · What is left: Three generations already defined, and the engineers who built them · Wears off in: Two to three years</li><li><strong>Supply-chain principal</strong> — What it was: TSMC’s largest customer, 19% of its 2025 revenue; $279B of supply on order · What is left: The contracts and the scale · Wears off in: Two years, then at the margin</li><li><strong>Allocator</strong> — What it was: Customers want about twice this year’s volume and supply allows 70% more, so each shipment is a choice of customer · What is left: Purchase orders for the next few quarters · Wears off in: Months</li><li><strong>Banker</strong> — What it was: $189.5B of guarantees, buy-backs, leases and equity committed for customers; a $99B portfolio of 242 stakes · What is left: Everything already signed · Wears off in: At once, for anything not signed</li><li><strong>Diplomat</strong> — What it was: Export licences, sovereign buyers and the White House, in person; the outlook already assumes no data center chips sold to China · What is left: The licences in hand · Wears off in: At once</li></ul>
<h2>How long the rails run</h2>
<p><strong>The year to January 2028 is unusually well protected.</strong> In August Kress told investors to expect revenue growth of about 70% and said supply, not demand, sets that figure at least through that year: customers’ forecasts pointed to demand roughly doubling. Analysts have $692B for the year, on estimates from $507B to $779B. Doubling this year’s $412B would be about $823B, so about 16% of next year’s demand, $131B, could disappear before a chip went unsold. Kress also said the demand from AI labs that Nvidia supports with its own balance sheet is about a quarter of next year’s business, about $173B. The cushion covers three-quarters of it.</p>
<p><strong>The supply is bought.</strong> Nvidia had $279B of supply and capacity on order at the end of July, up from $119B three months earlier and $16B at the start of 2025: mostly memory, wafers and packaging, with $92B falling due by January and $87B in the year after. TSMC’s annual report shows its largest customer, read as Nvidia, at 19% of 2025 revenue, up from 12%, and Morgan Stanley has Nvidia taking about 45% of 2027’s advanced packaging. The relationship is nearly three decades old and was personal: it began with a letter Huang wrote to Morris Chang in 1997, and Chang’s memoir says he offered Huang his own job in 2013. The orders are not, and at this size they outlast him. They also cut the other way. If demand did fall, Nvidia would still owe for the parts.</p>
<p><strong>Three generations are defined.</strong> Vera Rubin’s six chips taped out by August 2025 and went into production a year later; Kress expected it to be about a fifth of data center revenue in the quarter just ended. Rubin Ultra is due in 2027 and Feynman in 2028, on the roadmap Huang showed in March. On Rubin’s timing a chip’s design is frozen about a year before it ships and its architecture well before that, so Feynman is Huang’s design, finished by someone else. The first generation defined without him would ship in 2029 or 2030.</p>
<p><strong>So is the capital return.</strong> The board added $150B to the buyback on September 28, leaving $235B to be used by January 2028. At $165 that is 5.9% of the company.</p>
<p><strong>What is not on rails is people, and the decisions still in flight.</strong> Nvidia reported staff turnover of 2.7% in fiscal 2024. That loyalty was built with stock now 31% below its high and by a founder who has left; Oppenheimer’s downgrade cites the risk of departures. California does not enforce non-compete agreements, and Elon Musk has said he would be honoured to hire Huang. Nor is the half of the data center business that is not the hyperscalers on rails: $40B last quarter against their $49B, growing twice as fast, and sold by Huang deal by deal to AI clouds, governments and labs. Open decisions include Rubin Ultra’s memory configuration, which SemiAnalysis reports is still moving (<a href="https://06.ai/packs/ai-memory-balance">the </a><a href="https://06.ai/packs/ai-memory-balance">memory</a><a href="https://06.ai/packs/ai-memory-balance"> pack</a>), next year’s prices, and which customers get backed.</p>
<h2>What is already decided</h2>
<figure><a href="https://06.ai/packs/nvidia-after-jensen#rails"><img src="https://06.ai/newsletter/nvidia-after-jensen/fig-02.png" alt="What is already decided" width="694" height="386"></a><figcaption>Each chip sits in its launch year on Nvidia’s roadmap; the shaded years are those whose products would be defined after Huang. The strong grey bars are commitments with dates on them. The dashed line is November 12, 2026.</figcaption></figure>
<h2>The bank of Nvidia</h2>
<p><strong>The commitments and guarantees in Nvidia’s latest filing add up to $530.5B, and $189.5B of that was made for customers.</strong> The rest, $341B, is for its own supply and operations. The largest item for a customer is a guarantee of up to $105B, entered in August, behind the Ohio campus that SB Energy is building for OpenAI to rent. Another $36B obliges Nvidia to buy capacity from AI clouds if they cannot sell it; CoreWeave’s $6.3B order, which runs to April 2032, is one such. Nvidia also holds about $99B of stakes in 229 private and 13 public companies by Hunterbrook’s count, has put nearly $50B into frontier labs, and borrowed $25B in June. On the charge that this is circular financing, Huang said in August “we see it differently”: the compute can be redeployed, so the risk is limited.</p>
<p><strong>About $240B more was under discussion and unsigned.</strong> A preliminary agreement to put $100B of equity into OpenAI in ten installments tied to chip orders. A backstop of up to a quarter of the $500B that six financiers (Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR) aim to raise for Nvidia-based data centers, a figure Huang gave in a post and no filing contains. Up to $10B beside Anthropic’s IPO. Talks on more than $5B of guarantees for Nscale. None of it binds the company. Each item needs someone to say yes, and that person is now an interim chief with an inquiry open and every reason to be careful. UBS estimates a pause could take about $50B out of the ecosystem for a time.</p>
<p><strong>The comparison people reach for is J.P. Morgan,</strong> who decided in 1907 which banks would be saved. He died in March 1913 and the Federal Reserve Act was signed that December: the function outlived the man by becoming an institution. That is already under way here. Broadcom finances its two lab customers through a vehicle with Apollo and Blackstone, whose first tranche was $35B. Alphabet’s backstops for partners’ leases reached $48B in September (<a href="https://06.ai/wire/w090">W.090</a>). The six-firm platform was to be Nvidia’s version, and whether it is signed with Nvidia’s backstop is the first thing to watch.</p>
<p><strong>Against revenue the support is large; against cash flow it is not.</strong> Lucent’s financing commitments to customers peaked at $8.1B in 2000, about 24% of a year’s revenue. Nvidia’s $189.5B is 46% of this fiscal year’s expected revenue. But analysts expect $190B of free cash flow this year and $327B next, which Lucent never had. The exposure that matters is its concentration in a few labs without investment-grade credit, OpenAI above all.</p>
<h2>Nvidia’s commitments and guarantees</h2>
<figure><a href="https://06.ai/packs/nvidia-after-jensen#bank"><img src="https://06.ai/newsletter/nvidia-after-jensen/fig-03.png" alt="Nvidia’s commitments and guarantees" width="694" height="582"></a><figcaption>$ billions. Signed items are from the filing for the quarter to July 26, as tabulated by FourWeekMBA and Hunterbrook; the Ohio guarantee was entered in August and is a maximum. The dashed bars were reported or preliminary and appear in no filing; “up to” and “more than” are as reported.</figcaption></figure>
<h2>What happened to the others</h2>
<p><strong>No earlier case matches on all four counts: a founder, the company’s chief innovator, at the peak, overnight.</strong> Jobs in 2011 was the first three, but he had been ill for years and handed over to a deputy. Gates in 2000 was a founder at the peak who chose his moment and stayed as chairman. Krzanich at Intel and Hurd at HP left overnight over conduct and were replaced for a time by finance chiefs, but neither had founded the company. Jobs in 1985 was forced out abruptly, from a company that was not at a peak.</p>
<p><strong>The businesses kept growing.</strong> Apple’s revenue went from $108.2B in the year Jobs resigned to $233.7B four years later. Microsoft’s more than tripled under Ballmer and its profit more than doubled. Even after 1985 Apple grew from $1.9B to $11.1B in ten years on the Macintosh before it nearly failed.</p>
<p><strong>What investors would pay for them did not.</strong> From the day Jobs resigned to April 18, 2013, Apple’s trailing earnings per share rose 66% and its share price 4%: the multiple fell from 14.9 to 9.4. Microsoft’s market value fell from $604B to $269B over Ballmer’s 13 years while its profit rose 133%.</p>
<p><strong>Product and margin slowed inside two years.</strong> Apple’s sales growth went from 45% in fiscal 2012 to 9% and 7% in the next two years, and its gross margin from 43.9% to 37.6% in one. Its first new category, the Watch, arrived 44 months after Jobs left. What Cook added was financial. From fiscal 2012 to 2023 Apple’s net income rose 2.3 times and its earnings per share 3.9 times, because the share count fell 40%.</p>
<p><strong>Where it went wrong, the successor was the problem.</strong> Intel made its interim finance chief permanent after seven months, though he had said he was not a candidate. In July 2020 it disclosed that its next manufacturing process was six months late; the shares fell 16% in a day and AMD’s rose 17%. HP hired an outsider who lasted 11 months. TSMC, which Morris Chang handed to two engineers after years of preparation, has grown its revenue about 3.7 times since. Yahoo Finance counts six chief executives removed over conduct: the median first-day move was −3%, and four of the six lagged the S&amp;P 500 over the next year, by a median of about 12 points. Nvidia’s 13% opening fall was outside the range of all six.</p>
<h2>Earnings kept rising. The value did not</h2>
<figure><a href="https://06.ai/packs/nvidia-after-jensen#precedents"><img src="https://06.ai/newsletter/nvidia-after-jensen/fig-04.png" alt="Earnings kept rising. The value did not" width="694" height="244"></a><figcaption>Change in earnings (strong grey) and in what the market paid (pink) after two founders stepped back at a peak. Apple: trailing four-quarter earnings per share and share price, Aug 24, 2011 to Apr 18, 2013. Microsoft: annual net income and market value, January 2000 to August 2013.</figcaption></figure>
<ul><li><strong>Apple, 1985</strong> — Who left, and how: Steve Jobs, co-founder, forced out by the board · Who followed: John Sculley, already chief executive · What followed: Revenue rose from $1.9B to $11.1B in ten years on the Macintosh. Then the company nearly failed and Jobs came back.</li><li><strong>Microsoft, 2000</strong> — Who left, and how: Bill Gates hands over at the peak and stays as chairman · Who followed: Steve Ballmer, his deputy · What followed: Revenue $23B to $77.9B and net income $9.4B to $21.9B by 2013. Market value $604B to $269B; $1,000 became $767 with dividends.</li><li><strong>HP, 2010</strong> — Who left, and how: Mark Hurd resigns after a conduct and expenses inquiry · Who followed: The finance chief as interim, then an outsider who lasted 11 months · What followed: The shares had their largest fall in six years and lagged the S&amp;P 500 by nearly 40 points over the next year.</li><li><strong>Apple, 2011</strong> — Who left, and how: Jobs resigns, ill, at the peak · Who followed: Tim Cook, the operating chief who had run the company during his leaves · What followed: Revenue $108.2B to $233.7B in four years; the first new category 44 months later. In the first 20 months earnings per share rose 66% and the share price 4%.</li><li><strong>Intel, 2018</strong> — Who left, and how: Brian Krzanich resigns over a past relationship with an employee · Who followed: The finance chief as interim, made permanent after seven months · What followed: Its next process ran six months late in 2020 and the shares fell 16% in a day. AMD’s share of server processors went from about 1% to 27%.</li><li><strong>TSMC, 2018</strong> — Who left, and how: Morris Chang retires at 86 after a planned handover · Who followed: Two engineers, as chairman and chief executive · What followed: Revenue up about 3.7 times by 2025, with Nvidia its largest customer.</li></ul>
<h2>How far behind the others are</h2>
<p><strong>On the last public benchmark, AMD’s best measured chip did per megawatt what Nvidia’s 2024 chip did.</strong> SemiAnalysis’s InferenceX has AMD’s MI355X at 2.01M tokens a second per megawatt on an agentic workload, against 2.26M for Nvidia’s H200, 28.5M for a GB300 rack and 59.4M for a Vera Rubin rack on pre-release software. The gap is the rack: 14 times, then 30. AMD had no rack-scale system until Helios, which first shipped in the quarter to September, more than a year after Nvidia’s first racks. Helios has no public result yet. AMD says its MI455X chip leads on compute, memory and bandwidth.</p>
<p><strong>In revenue AMD is about three years behind.</strong> Its data center segment, processors included, will be about $34B this year and at least $75B in 2027 on its own guidance. Nvidia sold $96B in its July quarter alone. AMD’s whole 2027, at least $94B, is about three-quarters of what Nvidia sold in the year to January 2025.</p>
<p><strong>The nearer rival is custom silicon.</strong> Broadcom expects AI revenue of $58B this fiscal year, about $115B next and $230B the year after, and says the supply for both is secured. On current outlooks that takes Broadcom from 14% of Nvidia’s revenue to 25% in two years. Its chief executive says the chip it builds for OpenAI is comparable to Vera Rubin on OpenAI’s own workloads and that custom chips cost less than half as much, claims with no public benchmark behind them. The buyers’ contracts say more. Of the $644B in Anthropic’s seven largest disclosed compute contracts, 61% is with Broadcom, Google and Amazon, largely on their own chips, and 39% is on Nvidia’s (<a href="https://06.ai/wire/w001">W.001</a>).</p>
<p><strong>So how many years to parity?</strong> On the chip, AMD says none. On the system, about one generation, 12 to 18 months, and with both companies on an annual cadence that gap holds unless Nvidia misses a beat. That is the part the succession decides. On software the moat is being drained from outside: Lisa Su said Anthropic used Claude to bring a model up on AMD racks in a weekend. On volume it is not close: AMD’s whole 2027 is one Nvidia quarter, and what it can ship is set by the same TSMC and memory lines Nvidia has booked. Intel shows how slowly a leader loses even after a slip: AMD’s share of server processors went from about 1% in 2017 to 27% of units and 41% of revenue by mid-2025, three to four points a year.</p>
<h2>Tokens per megawatt, by chip</h2>
<figure><a href="https://06.ai/packs/nvidia-after-jensen#bench"><img src="https://06.ai/newsletter/nvidia-after-jensen/fig-05.png" alt="Tokens per megawatt, by chip" width="694" height="275"></a><figcaption>Tokens a second per megawatt of utility power, serving DeepSeek V4 Pro at 100 tokens a second per user, in millions, from SemiAnalysis’s InferenceX in September; values are interpolated between measured points and the Vera Rubin figure used pre-release software. AMD’s Helios rack and Broadcom’s custom chips have no public result. The label under each name is the maker and the year it shipped.</figcaption></figure>
<h2>Revenue on each company’s own outlook</h2>
<figure><a href="https://06.ai/packs/nvidia-after-jensen#outlook"><img src="https://06.ai/newsletter/nvidia-after-jensen/fig-06.png" alt="Revenue on each company’s own outlook" width="694" height="358"></a><figcaption>$ billions, three years each, lighter to darker. Nvidia is all revenue on analysts’ consensus (it guides to about 70% growth in the middle year); Broadcom is AI chips only, on its own outlook; AMD is its data center segment, from three reported quarters, its guidance and its statement that 2027 revenue grows more than 80% with data center above 80% of it. Fiscal years differ: Nvidia’s end in January, Broadcom’s around November, AMD’s in December.</figcaption></figure>
<h2>What slowing down would cost</h2>
<p><strong>Two questions are easy to confuse.</strong> One is how much the industry buys. The other is what a change of leadership costs Nvidia within that. We run three paths for each. For demand: the analysts’ path ($692B then $903B, then 12% and 6% growth, which is <a href="https://06.ai/packs/ai-compute-balance">the </a><a href="https://06.ai/packs/ai-compute-balance">compute-balance</a><a href="https://06.ai/packs/ai-compute-balance"> pack</a>’s accelerator spending less three points of share a year); a plateau, where revenue stops growing after January 2028 as in <a href="https://06.ai/packs/memory-prices">the </a><a href="https://06.ai/packs/memory-prices">memory-price</a><a href="https://06.ai/packs/memory-prices"> pack</a>; and a digestion, where it falls 25%, then 10%, then recovers 10%. For succession: the handover holds, as at Apple in 2011 or TSMC in 2018; drift, where revenue runs 3%, 7% and 11% below the path in the three years to January 2031 and gross margin gives up as much as 2.5 points, about the pace at which Intel lost servers; and disorder, with departures and a late generation, where the shortfall reaches 25% and 5.5 points.</p>
<p><strong>A point of share is worth 22 cents.</strong> On the analysts’ path, 1% of revenue in the year to January 2029 is $9.0B and $0.22 of earnings per share. A point of gross margin is $0.31.</p>
<p><strong>Succession moves value by 13% to 30%. Demand moves it more.</strong> With the handover holding, a rough present value runs from $227 a share on the analysts’ path to $148 on the plateau and $107 in a digestion. On any of them drift takes off about 13% and disorder about 30%.</p>
<p><strong>At $165 the market is charging for one of the two problems, not both.</strong> The price sits a little above the plateau with a clean handover ($148) and the analysts’ path with disorder ($159). Both at once would be $103. Put our own weights on the nine cells, 35%, 45% and 20% on the demand paths and 45%, 40% and 15% on succession, and the answer is $151, which is where the shares opened. Every ten points moved from the plateau to the analysts’ path adds about $7.</p>
<p><strong>The discount for Huang is larger than the earnings damage we can model.</strong> Our weights take 10% off the value for succession. Certain drift takes off 13%. The shares stand 21% below where the chip index would have carried them, which is about the cost of an even bet between drift and disorder (22%). Either investors think the handover is very unlikely to hold, or part of the discount is a lower price for each dollar of earnings, which the Apple and Microsoft cases suggest.</p>
<h2>What each path is worth, against the share price</h2>
<figure><a href="https://06.ai/packs/nvidia-after-jensen#values"><img src="https://06.ai/newsletter/nvidia-after-jensen/fig-07.png" alt="What each path is worth, against the share price" width="694" height="391"></a><figcaption>A rough present value per share: earnings for the four years to January 2031 plus 10 times the last year’s, discounted at 10% (the ten-year Treasury yields about 5%). Earnings stand in for cash. The dashed line is about $165, the price on the morning of November 12. Analysts’ path: consensus to January 2029, then 12% and 6% growth. Plateau: revenue stops growing after the year to January 2028. Digestion: revenue falls 25%, then 10%, then recovers 10%.</figcaption></figure>
<h2>Earnings per share on each path</h2>
<p>Pick a demand path. The dashed grey line is analysts’ consensus to January 2029, extended two years. The slider changes the multiple put on the last year’s earnings, which is where a view on growth after 2031 goes: at a 10% discount rate a multiple of 10 assumes none.</p>
<p><em>Sliders at their defaults: Analysts’ path · Multiple on the last year 10×.</em></p>
<figure><a href="https://06.ai/packs/nvidia-after-jensen#explorer"><img src="https://06.ai/newsletter/nvidia-after-jensen/fig-08.png" alt="Earnings per share on each path" width="694" height="348"></a></figure>
<p>The handover holds · $227 · 14.2× next year’s expected earnings; <strong>+37%</strong> against $165</p>
<p>Drift · $196 · 12.3× next year’s expected earnings; <strong>+19%</strong> against $165</p>
<p>Disorder · $159 · 10.0× next year’s expected earnings; <strong>−4%</strong> against $165</p>
<h2>What multiple</h2>
<p><strong>On next year’s expected earnings of $15.91 a share, Nvidia traded at 15.0 times at the record, 10.9 on Wednesday, 9.5 at Thursday’s open and 10.4 at $165.</strong> LSEG’s 12-month measure, about 12, is the lowest since November 2012. Analysts’ targets averaged $327 a week ago, 20.6 times. The ones cut this week sit at $200 to $225, 12.6 to 14.1 times. Bernstein halved its multiple from about 25 to 12.7 without changing a forecast and called the shares “almost absurdly cheap” while advising against buying in size. The challenger gets three times the multiple: AMD closed Monday at 31 times its expected 2027 earnings.</p>
<p><strong>Each multiple is a statement about a path.</strong> At a 10% discount rate and no growth after 2031, 14.2× is the analysts’ path with a clean handover. 12.3× is the analysts’ path with drift, which is where the sell side has reset. About 10 is the analysts’ path with disorder, or a plateau with a clean handover, which is where the market is. 8.0× and below is a plateau with drift or worse, and under 7 needs a digestion. Twenty needs growth after 2031: a new market of the kind Huang found three times, in games, in data centers and in networking. That part of the multiple left with him, and no successor gets it back by being competent. On our weights the answer is 9.5×, against 10.4× in the market.</p>
<p><strong>The Tim Cook case is arithmetic, and it is available.</strong> Analysts expect $327B and $452B of free cash flow in the next two fiscal years, 8.2% and 11.4% of the market value at $165. Spending 60% of it on stock at an average of $200 would retire about 9.7% of the shares in two years, which adds as much to earnings per share as drift takes away by January 2030 (10%). A finance chief can run that, and this one already does. (President Trump said on Thursday that Cook himself should take the job, Yahoo Finance reported.) It works only if the earnings arrive, and it does nothing for the next architecture. The part of Cook’s record that does not transfer is the customer. Apple sold to consumers; about half of Nvidia’s data center revenue comes from a few hyperscalers that design chips of their own.</p>
<h2>Who else is exposed</h2>
<p><strong>The companies that fell furthest on Thursday were Nvidia’s dependants.</strong> In the first hour or so CoreWeave was down about 12%, IREN 10% and Nebius 16% or more, against Nvidia’s 5%. Broadcom rose about 4.5% and AMD 2.8%; William Blair upgraded AMD. TSMC fell 2%.</p>
<p><strong>The AI clouds exist because Nvidia chose to supply and back them.</strong> Four of the eight clouds it named first for Vera Rubin are neoclouds (<a href="https://06.ai/packs/blackwell-per-dollar">the </a><a href="https://06.ai/packs/blackwell-per-dollar">Blackwell</a><a href="https://06.ai/packs/blackwell-per-dollar"> pack</a>), and customers other than the hyperscalers are now nearly half of its data center revenue, $40B against $49B last quarter, and grew twice as fast in it. CoreWeave owes $46.9B, pays 2.2 times its adjusted operating income in interest and saw its bonds yield 15.08% in October (<a href="https://06.ai/wire/w076">W.076</a>). What is signed stays signed, including Nvidia’s promise to buy its unsold capacity. What changes is the next deal. A new backstop needs an approval nobody is placed to give, and Kress said in August that Nvidia means its chips to go to customers with investment-grade credit or with a backer that has it. The reporting behind the inquiry centres on a Southeast Asian customer (<a href="https://06.ai/wire/w054">W.054</a>), so checks on who buys will tighten first for buyers of that kind. One offset: if Nvidia’s cadence slips, the chips these companies own stay current for longer (<a href="https://06.ai/packs/gpu-decommission">the </a><a href="https://06.ai/packs/gpu-decommission">GPU-decommission</a><a href="https://06.ai/packs/gpu-decommission"> pack</a>).</p>
<p><strong>Among the labs OpenAI is the most exposed and Anthropic the least.</strong> OpenAI has the unsigned $100B, the $105B guarantee behind its Ohio campus and an operating loss larger than its revenue last quarter (<a href="https://06.ai/wire/w079">W.079</a>). Anthropic is raising its own money this month and has most of its compute on other chips.</p>
<p><strong>Broadcom and AMD gain a customer base with more reason to dual-source, but not more chips to sell.</strong> Both say demand exceeds what they can supply. Share moves only as fast as TSMC and the memory makers allot capacity, and Nvidia’s orders for 2027 are placed. TSMC and the memory makers sell the wafer either way.</p>
<p><strong>Does total investment fall?</strong> Not by much on the arithmetic. The five largest hyperscalers will spend about $800B this year and $1.3T next, by Nvidia’s own count, from their cash flow and the bond market (<a href="https://06.ai/packs/hyperscaler-capex-funding">the </a><a href="https://06.ai/packs/hyperscaler-capex-funding">hyperscaler</a><a href="https://06.ai/packs/hyperscaler-capex-funding"> pack</a>). The support at risk is the unsigned $240B, spread over several years; UBS’s $50B is 4% of next year’s hyperscaler spending. Output is capped by supply, so a chip a lab cannot finance goes to a hyperscaler that can. What changes is who gets the chips, and what lenders charge the ones who lose them. The Fed’s Bowman put AI-related debt above $800B this year (<a href="https://06.ai/wire/w090">W.090</a>). Some of the riskiest of it leans on Nvidia’s guarantees and promises to buy.</p>
<h2>The first hours of trading on November 12</h2>
<figure><a href="https://06.ai/packs/nvidia-after-jensen#moves"><img src="https://06.ai/newsletter/nvidia-after-jensen/fig-09.png" alt="The first hours of trading on November 12" width="694" height="276"></a><figcaption>Share-price change from Wednesday’s close, %, as reported by the outlets in the table below between the open and about 10:50 a.m. in New York. The readings were taken at different times and moved while this was written. Nvidia opened 13.1% lower and is shown at about $165. CoreWeave had risen 27% the day before on its results, and Nebius also said its 2026 revenue would land at the low end of its range.</figcaption></figure>
<h2>What to watch</h2>
<ul><li><strong>Nov 17.</strong> Nvidia’s results. Its guide was $108B, plus or minus 2%. What matters is whether the new management repeats the outlook for about 70% growth next year, and what happened to the $279B of supply on order.</li><li><strong>Nov 18.</strong> The closed briefing three House chairmen asked Huang for. Three China chip bills are in the Senate’s draft defense bill.</li><li><strong>This month.</strong> Anthropic’s IPO, where Nvidia was down for up to $10B; the preliminary $100B for OpenAI; and whether the financing platforms sign with Nvidia’s backstop.</li><li><strong>Next months.</strong> Who gets the job. Intel took seven months and kept its finance chief; an engineer from inside is the nearest thing to Apple’s or TSMC’s handover. Press accounts mention Kress herself and Ian Buck, who built Nvidia’s CUDA software, from inside, and chip executives at Apple, Arm and AMD from outside. Departures among the 36 people who reported to Huang would be the first sign of disorder.</li><li><strong>Any day.</strong> Huang’s shares. He held 870.6M in March, 3.58% of the company, about $144B at $165. He has not said what he will do next.</li><li><strong>March 2027.</strong> The first GTC without him, and whether Rubin Ultra is on schedule.</li></ul>
<h2>What this is anchored on</h2>
<ul><li><strong>The handover</strong> — Huang resigned as president, chief executive and director, effective immediately; Kress interim chief executive and still finance chief; Jones the first board chair; internal and external candidates; results still due November 17 · <a href="https://www.globenewswire.com/news-release/2026/11/12/3401635/0/en/nvidia-announces-ceo-transition.html">Nvidia, Nov 12</a></li><li><strong>The reporting before it</strong> — The board’s inquiry of November 8 and the reports of November 11 · <a href="https://06.ai/wire/w087">06.ai, W.054 and W.087</a></li><li><strong>Share prices</strong> — Daily closes June 1 to November 5 (Yahoo Finance, as collected for the <a href="https://06.ai/packs/unit-growth">unit-growth</a> pack); $186.64, $178.67, $168.89 and $173.73 on November 6 to 11 · <a href="https://06.ai/wire/w087">06.ai, W.087</a></li><li><strong>Share prices</strong> — November 12: open $150.99, low $148.80 · <a href="https://robinhood.com/us/en/stocks/NVDA/">Robinhood, Nov 12</a></li><li><strong>Share prices</strong> — Down as much as 14%, then about $161 at 10:11 a.m. ET; IREN −9.8% · <a href="https://www.coindesk.com/markets/2026/11/12/nvidia-shares-plunge-10-as-jensen-huang-resigns">CoinDesk, Nov 12</a></li><li><strong>Share prices</strong> — $165.05, down 5.00%, at 10:51 a.m. ET; fell more than 13% in early trading; the president’s remark about Tim Cook · <a href="https://finance.yahoo.com/technology/live/nvidia-live-updates-jensen-huang-resigns-cfo-colette-kress-to-serve-as-interim-ceo-140341430.html">Yahoo Finance live blog, Nov 12</a></li><li><strong>Chip index</strong> — PHLX Semiconductor Index 9,390.25 on November 6 (W.066); about 10,560 on November 11, 17% above its November 5 low; Nvidia lost about $190B from November 5 to 11 while TSMC, Broadcom and AMD each gained about $250B · <a href="https://finance.yahoo.com/markets/article/nvidias-boardroom-bombshell-hits-a-fragile-chip-rally-chart-of-the-day-110000355.html">Yahoo Finance, Nov 12</a></li><li><strong>Analysts’ estimates</strong> — Revenue $411.7B and $692.2B, earnings per share $9.31 and $15.91, for the years to January 2027 and 2028; 57 analysts, with next year’s revenue from $507B to $779B; mean price target $327.14 · <a href="https://finance.yahoo.com/quote/NVDA/analysis/">Yahoo Finance, fetched Nov 6</a></li><li><strong>Analysts’ estimates</strong> — Year to January 2029: revenue $903.2B, earnings per share $20.52; free cash flow $190.2B, $326.8B and $451.7B for the three years · <a href="https://www.marketscreener.com/quote/stock/NVIDIA-CORPORATION-57355629/finances/">MarketScreener, Oct 23</a></li><li><strong>Results</strong> — Quarter to July 26: revenue $96.2B, Data Center $89.0B, gross margin 75%; first-half free cash flow $69.9B, repurchases $39.0B; outlook $108B plus or minus 2%; Vera Rubin in full production · <a href="https://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-second-quarter-fiscal-2027">Nvidia, Aug 26</a></li><li><strong>Outlook</strong> — Revenue growth of about 70% in the year to January 2028, limited by supply at least through that year; customers’ forecasts point to demand roughly doubling; gross margin 71–72% in the January quarter and 72–73% next year; lab demand supported by Nvidia’s balance sheet about a quarter of next year’s business; nearly $50B invested in frontier labs; hyperscale revenue $49B and all other data center revenue $40B · <a href="https://www.webull.com/news/15497781742806016">Nvidia’s August 26 call, transcript via Webull</a></li><li><strong>Commitments</strong> — The filing’s three tables: $279B of supply and capacity among $366B of future commitments, $56B of additional commitments, $108.5B of guarantees, $530.5B in all; supply was $119B a quarter earlier · <a href="https://fourweekmba.com/ai-nvidia-10-q-lists-530b-in-commitments-and-guarantees/">FourWeekMBA, on the 10-Q</a></li><li><strong>Commitments</strong> — $99B of stakes in 229 private and 13 public companies; a preliminary agreement for $100B of OpenAI equity in ten installments; Huang’s statement that Nvidia could backstop up to a quarter of the $500B platforms; talks with Nscale; CoreWeave’s $6.3B order · <a href="https://newsletter.hntrbrk.com/p/the-comprehensive-guide-to-the-bank">Hunterbrook, Nov 6</a></li><li><strong>Commitments</strong> — The $105B guarantee, entered in August for the Ohio campus OpenAI will rent, ends if the tenant earns a satisfactory credit rating; $25B of notes sold in June; CoreWeave’s order runs to April 13, 2032 · <a href="https://www.capefearadvisors.com/strategy-lab/nvidia-the-scarce-thing">Cape Fear Advisors, Aug 27</a></li><li><strong>Commitments</strong> — Supply commitments $16.1B, then $95.2B, at the last two fiscal year-ends; $92B due in the rest of this fiscal year and $87B in the next · <a href="https://finance.yahoo.com/markets/stocks/articles/michael-burry-called-nvidia-95-180332650.html">Barchart, via Yahoo Finance, Sep 2</a></li><li><strong>Buyback</strong> — $150B added on September 28, $235B remaining, to be used through the year to January 2028 · <a href="https://nvidianews.nvidia.com/news/nvidia-announces-a-150-billion-share-repurchase-authorization-increase">Nvidia, Sep 28</a></li><li><strong>Thursday’s reporting</strong> — Nvidia set to invest up to $10B beside the IPO; about 60 direct reports; the tally of six conduct exits; Bernstein’s and Yorkville’s notes · <a href="https://finance.yahoo.com/technology/live/nvidia-live-updates-jensen-huang-resigns-cfo-colette-kress-to-serve-as-interim-ceo-140341430.html">Yahoo Finance live blog, Nov 12</a></li><li><strong>Analysts</strong> — Bernstein to market-perform, $400 to $200, its multiple from about 25.4 to 12.7 on unchanged forecasts; Baird to neutral, $500 to $200 · <a href="https://investinglive.com/stocks/bernstein-halves-nvidia-target-to-200-the-3rd-broker-in-2-days-to-step-back-over-huang/">investingLive, Nov 12</a></li><li><strong>Analysts</strong> — Oppenheimer’s downgrade and its warning on departures; Stifel $205, Citi $225; UBS on a pause in Nvidia’s funding taking about $50B out for a time · <a href="https://uk.investing.com/news/analyst-ratings/wall-street-reacts-to-nvidia-saga-bernstein-and-oppenheimer-downgrade-stock-4977374">Investing.com, Nov 12</a></li><li><strong>Thursday’s moves</strong> — CoreWeave −13.6%, Nebius −20%, AMD +2%; Nvidia at about 12 times 12-month forward earnings, the lowest since November 2012, per LSEG · <a href="https://www.marketscreener.com/news/global-ai-stocks-sink-as-nvidia-crisis-deepens-sector-selloff-ce785ed8dc80f526">Reuters, via MarketScreener, Nov 12</a></li><li><strong>Thursday’s moves</strong> — Broadcom +3.8% at 9:50 a.m. ET, AMD +2%, TSMC −2%, CoreWeave −12%, Nebius −16% to −19%, Micron and Intel about −3% · <a href="https://sg.finance.yahoo.com/news/amd-broadcom-stocks-jump-as-nvidia-ceo-jensen-huang-resigns-143725561.html">Yahoo Finance, Nov 12</a></li><li><strong>Thursday’s moves</strong> — AMD +2.8% and Broadcom +4.5% after the open; William Blair’s upgrade of AMD · <a href="https://uk.investing.com/news/stock-market-news/amd-broadcom-gain-after-reports-say-nvidia-board-preparing-to-oust-huang-4976645">Investing.com, Nov 12</a></li><li><strong>Organization</strong> — 36 direct reports in October 2025, down from 55, per an internal chart seen by Business Insider · <a href="https://www.tipranks.com/news/nvidia-nvda-reshapes-leadership-team-as-ai-expansion-speeds-up">TipRanks, Oct 16, 2025</a></li><li><strong>Organization</strong> — Names mentioned as possible successors, a roundup of press reports · <a href="https://en.bloomingbit.io/feed/news/124282">Korea Economic Daily, via Bloomingbit, Nov 12</a></li><li><strong>Organization</strong> — Elon Musk’s remark that it would be an honour to hire Huang · <a href="https://www.newsquawk.com/headlines/daily-us-equity-opening-news---12th-november-2026-nvda-board-weighs-ceo-ouster-on-fresh-affair-claims-crwv-reportedly-to-raise-4b-in-ig-loans">Newsquawk, Nov 12</a></li><li><strong>Organization</strong> — Staff turnover of 2.7% in fiscal 2024 · <a href="https://sherwood.news/business/earnings-nvidia-employees-millionare-rsu-talent-retention/">Sherwood, Aug 29, 2024</a></li><li><strong>Board</strong> — Directors, tenure and ages; Jones a director since 1993 · <a href="https://edgar-mirror.sec-api.io/1045810/000104581026000036/nvda-20260512.htm">Nvidia, 2026 proxy statement</a></li><li><strong>Board</strong> — The risk factors name no individual; they say Nvidia is highly dependent on its longstanding executive team and list a failure of succession planning as a risk · <a href="https://fortune.com/company-assets/1936/quartr/annual-report-10-k-4ccfc-2026-02-25-09-13-43.pdf">Nvidia, annual report for fiscal 2026, as hosted by Fortune</a></li><li><strong>TSMC</strong> — Its largest customer was 19% of 2025 revenue, NT$726.9B, from 12%; read as Nvidia · <a href="https://view.asiae.co.kr/en/article/2026030610061501773">Asia Economy, Mar 9, on TSMC’s annual report</a></li><li><strong>TSMC</strong> — Morris Chang’s memoir says he offered Huang the top job at TSMC in 2013 · <a href="https://thestar.com.my/tech/tech-news/2024/11/29/tsmc-founder-morris-chang-offered-top-job-to-jensen-huang-memoir-shows">Reuters, via The Star, Nov 29, 2024</a></li><li><strong>TSMC</strong> — Huang’s 1997 letter to Morris Chang began the partnership; Nvidia was among TSMC’s five largest customers within a few years · <a href="https://benzinga.com/z/43517209">Benzinga, Feb 2025</a></li><li><strong>TSMC</strong> — Advanced-packaging demand of about 2.7M wafers in 2027, Nvidia about 45% · <a href="https://www.kucoin.com/news/flash/morgan-stanley-report-cowos-demand-to-double-by-2027-amd-and-asics-drive-ai-chip-growth">Morgan Stanley, via KuCoin, Jul 8</a></li><li><strong>Allocation</strong> — Larry Ellison’s account of asking Huang for chips over dinner · <a href="https://fortune.com/2024/09/16/larry-ellison-elon-musk-begged-nvidias-jensen-huang-more-gpus-fancy-sushi-dinner/">Fortune, Sep 16, 2024</a></li><li><strong>Roadmap</strong> — Six Vera Rubin chips taped out at TSMC, as Huang said on a visit to Taiwan · <a href="https://www.notebookcheck.net/Nvidia-Rubin-architecture-taped-out-with-six-chips-at-TSMC-marks-major-platform-overhaul.1094897.0.html">Notebookcheck, Aug 24, 2025</a></li><li><strong>Roadmap</strong> — Rubin Ultra on the Kyber rack in 2027; Feynman with the Rosa CPU in 2028; a new architecture every year · <a href="https://www.tweaktown.com/news/110521/nvidia-updates-roadmap-with-new-details-on-its-next-gen-gpu-feynman-coming-in-2028/index.html">TweakTown, on GTC 2026</a></li><li><strong>Benchmark</strong> — Tokens a second per megawatt on DeepSeek V4 Pro at 100 tokens a second per user, interpolated between measured points; Vera Rubin on pre-release software · <a href="https://inferencex.semianalysis.com/blog/vera-rubin-nvl72-agentic-inference">SemiAnalysis InferenceX, Sep 14</a></li><li><strong>AMD</strong> — Quarterly revenue and Data Center revenue; fourth-quarter guidance of about $17B · <a href="https://06.ai/wire/w004">06.ai, W.004, from AMD’s releases</a></li><li><strong>AMD</strong> — Data Center up about 50% in the fourth quarter; revenue growth above 80% in 2027 with data center more than 80% of it; first Helios revenue in the third quarter; the weekend port · <a href="https://www.fool.com/earnings/call-transcripts/2026/11/04/amd-amd-q3-2026-earnings-call-transcript/">AMD’s November 3 call, transcript via Motley Fool</a></li><li><strong>AMD</strong> — Closed at $508.58 on November 9; analysts expect $16.30 a share in 2027 (Yahoo Finance, November 6) · <a href="https://06.ai/wire/w066">06.ai, W.066</a></li><li><strong>AMD</strong> — 27.3% of server processors by units and 41% by revenue in the second quarter of 2025, per Mercury Research · <a href="https://www.tomshardware.com/pc-components/cpus/amds-desktop-pc-market-share-hits-a-new-high-as-server-gains-slow-down-intel-now-only-outsells-amd-2-1-down-from-9-1-a-few-years-ago">Tom’s Hardware, Aug 14, 2025</a></li><li><strong>Broadcom</strong> — AI revenue of $58B in fiscal 2026, about $115B in 2027 and $230B in 2028 with supply secured; six custom-chip customers; the financing vehicle with Apollo and Blackstone; Tan’s comparison with Vera Rubin · <a href="https://www.fool.com/earnings/call-transcripts/2026/09/09/broadcom-avgo-q3-2026-earnings-call-transcript/">Broadcom’s September call, transcript via Motley Fool</a></li><li><strong>Anthropic</strong> — Its seven largest disclosed compute contracts · <a href="https://06.ai/wire/w001">06.ai, W.001, from its prospectus</a></li><li><strong>CoreWeave</strong> — Debt of $46.9B, interest 2.2 times adjusted operating income; bonds at 15.08% in October · <a href="https://06.ai/wire/w076">06.ai, W.076 and </a><a href="https://06.ai/wire/w076">the </a><a href="https://06.ai/wire/w076">neocloud</a><a href="https://06.ai/wire/w076"> pack</a></li><li><strong>Apple</strong> — After hours on August 24, 2011, the shares fell about 5% to $357, from a close near $376; market value nearly $350B · <a href="https://money.cnn.com/2011/08/24/technology/apple_after_hours_shares/index.htm">CNN Money, Aug 24, 2011</a></li><li><strong>Apple</strong> — Record close of $702.10 on September 19, 2012; $392.05 on April 18, 2013 · <a href="https://www.macworld.com/article/670415/whats-happening-to-apples-share-price.html">Macworld, Apr 19, 2013</a></li><li><strong>Apple</strong> — Revenue, gross margin, quarterly earnings per share, net income and share counts, fiscal 2010 to 2023 · <a href="https://investor.apple.com/sec-filings/default.aspx">Apple, annual and quarterly reports</a></li><li><strong>Apple</strong> — An analyst’s 2007 estimate that losing Jobs could take a quarter off Apple’s value · <a href="https://fortune.com/2011/08/25/wall-street-shrugs-off-steve-jobs-resignation">Fortune, Aug 25, 2011</a></li><li><strong>Microsoft</strong> — Market value $604B in January 2000 and $269B in August 2013; $1,000 became $767 with dividends (FactSet) · <a href="https://phys.org/news/2013-08-microsoft-stock-slumped-ballmer.pdf">AP, via Phys.org, Aug 2013</a></li><li><strong>Microsoft</strong> — Revenue about $23B to $77.9B and net income $9.4B to $21.9B, fiscal 2000 to 2013 · <a href="https://www.buzzfeednews.com/article/mattlynley/steve-ballmers-tenure-as-microsoft-ceo-in-numbers">BuzzFeed News, Aug 2013</a></li><li><strong>Intel</strong> — Swan made permanent after a seven-month search, having said he was not a candidate · <a href="https://fortune.com/2019/01/31/intel-ceo-search-bob-swan/">Fortune, Jan 31, 2019</a></li><li><strong>Intel</strong> — A six-month delay to its 7nm process; the shares closed 16% lower and AMD’s nearly 17% higher · <a href="https://amp.cnn.com/cnn/2020/07/24/investing/intel-earnings-stock-selloff">CNN, Jul 24, 2020</a></li><li><strong>HP</strong> — The largest fall in six years after Hurd resigned; the finance chief as interim · <a href="https://www.bloomberg.com/news/articles/2010-08-09/hp-s-lesjak-inherits-slowing-growth-job-search-distraction-as-hurd-leaves">Bloomberg, Aug 9, 2010</a></li><li><strong>Lucent</strong> — Customer financing commitments of $8.1B, about 24% of a year’s revenue · <a href="https://tomtunguz.com/nvidia_nortel_vendor_financing_comparison/">Tomasz Tunguz, Oct 3, 2025</a></li></ul>
<h2>Method</h2>
<p><strong>Prices.</strong> About $165 is the price at about 10:50 a.m. in New York on November 12; it had moved between $148.80 and about $165 since the open. Market value is price times about 24.1B shares, the count on the cover of the latest quarterly filing. Next year’s earnings are analysts’ consensus for the fiscal year to January 2028, $15.91 a share, from Yahoo Finance on November 6; the year to January 2029, $20.52, is from MarketScreener on October 23. Neither had been revised for Thursday’s news.</p>
<p><strong>The chip index.</strong> $186.64 × (10,560 ÷ 9,390.25) = $209.89. The November 11 index level is read from Yahoo Finance’s chart, about 17% above the November 5 low, and is approximate. Nvidia is itself a large part of the index, which makes the gap look smaller than it is; some of the index’s rise was rivals gaining at Nvidia’s expense, which makes it look larger. The index’s move on November 12 is not included.</p>
<p><strong>The cushion.</strong> Demand roughly doubling is taken as 2 × this year’s expected $411.7B, against $692.2B expected for next year: 1 − 692.21 ÷ 823.3 = 15.9%. A quarter of $692.2B is $173.1B.</p>
<p><strong>Commitments.</strong> We did not read the filing itself; the tables are as transcribed by FourWeekMBA, Hunterbrook and Cape Fear Advisors, which agree. “For customers” is the guarantees ($108.5B), the AI-cloud agreements ($36B), the leases Nvidia expects to reassign ($20B) and equity commitments ($25B): $189.5B. The unsigned items are as reported and are maxima or minima, so their $240B sum is an order of magnitude, not a forecast.</p>
<p><strong>Scenarios.</strong> Earnings per share on each path are analysts’ figures scaled by modelled operating income: revenue × gross margin − operating cost, with the analysts’ path at 72.5% and the costs in the table under the explorer. The analysts’ path with no succession effect reproduces their consensus to January 2029 and is extended at 12% and 6%. The share count is held flat, as it is in the consensus, so buybacks are not in the scenario figures. Value = Σ EPS ÷ 1.1<sup>t</sup> for the four years, with the first ending 1.2 years out, plus 10 × the last year’s EPS, discounted the same way. It leaves out the quarter now under way, net cash and the $99B of stakes, together about $8 a share. The paths and the weights are ours; they were set before the values were computed and have not been tuned to the share price.</p>
<p><strong>Precedents.</strong> Apple’s trailing earnings are the four quarters to June 2011 ($4.64, $6.43, $6.40, $7.79) and to March 2013 ($9.32, $8.67, $13.81, $10.09), on a close of about $376.18 on August 24, 2011 and $392.05 on April 18, 2013, before later share splits. Per-share figures for 2012 and 2023 are split-adjusted. AMD’s 2027 data center figure is 80% of 1.8 times its 2026 revenue of $52.2B, which is three reported quarters plus the fourth-quarter guide; both 80% figures are floors AMD gave. Anthropic’s split treats its Google and Amazon contracts as largely on those companies’ own chips, which is our reading, not a figure in its prospectus.</p>
<p><strong>Not covered.</strong> The inquiry’s findings, any legal or regulatory cost to Nvidia, and what Huang does next. Any of them could matter more than everything above.</p>
<hr>
<p>Every chart above links to its interactive version on 06.ai, with the data behind it. The daily trackers are at <a href="https://06.ai/trackers">06.ai/trackers</a>.</p>
<p><em>Not investment advice. Estimates are labeled as estimates.</em></p>]]></content:encoded>
</item>
<item>
  <title>What does Fable 5.2 do to the price of the frontier?</title>
  <link>https://inthesingularity.ai/packs/fable-5-2</link>
  <guid isPermaLink="true">https://inthesingularity.ai/packs/fable-5-2</guid>
  <pubDate>Wed, 11 Nov 2026 12:25:00 GMT</pubDate>
  <description>Claude Fable 5.2 took an estimated $1.37M on its first full day on OpenRouter, the biggest first day of the 20 launches we track, at a list price 20% below Fable 5.1’s. The list price matters less than it did. One model now sells at five effort levels whose cost per task spans as much as 21 times, and at high effort Fable 5.2 outscores Opus 5.5 at max for less money. Each of Anthropic’s last three flagships was matched at half its price within seven weeks and at a tenth within five months. Anthropic shipped nine new best models in twelve months, and the frontier moved 23 points on Epoch’s index against a trend of 17.</description>
  <category>AI labs</category>
  <enclosure url="https://inthesingularity.ai/newsletter/fable-5-2/cover.png" length="56006" type="image/png" />
  <media:content url="https://inthesingularity.ai/newsletter/fable-5-2/cover.png" medium="image" type="image/png" />
  <content:encoded><![CDATA[<figure><a href="https://06.ai/packs/fable-5-2"><img src="https://06.ai/newsletter/fable-5-2/thumb.png" alt="What does Fable 5.2 do to the price of the frontier?" width="421" height="391"></a><figcaption><a href="https://06.ai/packs/fable-5-2">Interactive version and data on 06.ai →</a></figcaption></figure>
<ul><li><strong>$1.37M</strong> — First full day on OpenRouter estimated spend on Nov 10: <strong>9.1%</strong> of all OpenRouter spend and 20.6% of Anthropic’s. Fable 5.1’s first day: $0.65M</li><li><strong>$8 / $40</strong> — List price per million tokens input / output: <strong>20% below</strong> Fable 5.1 and twice Opus 5.5. The top of Anthropic’s list is still 60% above a year ago</li><li><strong>62 v 59</strong> — Fable 5.2 at high, Opus 5.5 at max on Artificial Analysis’s index, for <strong>$9.08</strong> a task against $10.04. Low to max effort spans up to 21 times in cost</li><li><strong>174</strong> — Epoch Capabilities Index. <strong>+22</strong> on Opus 4.5, last year’s best. DeepSeek’s latest scores 159, level with Opus 4.8 from May</li></ul>
<h2>One full day in</h2>
<p><strong>Fable 5.2’s first full day on OpenRouter was the biggest of any launch we track in dollars, and ordinary as a share of Anthropic.</strong> It was listed on November 9 at about 19:00 UTC. On November 10, its first full day, it processed 179 billion tokens, which at list price is an estimated $1.37M: 9.1% of everything spent on OpenRouter that day, from 0.6% of the tokens. Only Opus 5.5 took more.</p>
<p>That is 2.1 times Fable 5.1’s first day in dollars and 2.6 times in tokens. But OpenRouter is 1.6 times the size it was at the start of September, and Fable 5.2 took 20.6% of Anthropic’s own spend there, a little less than Fable 5 (23.6%) and Fable 5.1 (22.4%) did. The launch is bigger because the market is, not because Fable took more of Anthropic.</p>
<h2>The first full day of every launch</h2>
<p>The ten largest of the 20 launches on 06.ai’s <a href="https://06.ai/trackers/frontier-launches">frontier launch tracker</a>, each on its first full UTC day on OpenRouter. Dollars are our estimate: OpenRouter’s token counts at list price.</p>
<p>Fable 5.2’s first full day brought in an estimated <strong>$1.37M</strong>, the most of the 20 launches tracked. Fable 5 took $1.19M in June and Fable 5.1 $0.65M in September.</p>
<figure><a href="https://06.ai/packs/fable-5-2#first-day"><img src="https://06.ai/newsletter/fable-5-2/fig-01.png" alt="The first full day of every launch" width="694" height="281"></a></figure>
<figure><a href="https://06.ai/packs/fable-5-2#first-day"><img src="https://06.ai/newsletter/fable-5-2/fig-02.png" alt="The first full day of every launch — Share of OpenRouter" width="694" height="312"></a><figcaption><strong>Share of OpenRouter</strong>: As a share of everything spent on OpenRouter that day, 9.1% ranks 6th of 20. Fable 5 took 16.1% of a market half the size.</figcaption></figure>
<figure><a href="https://06.ai/packs/fable-5-2#first-day"><img src="https://06.ai/newsletter/fable-5-2/fig-03.png" alt="The first full day of every launch — Share of Anthropic" width="694" height="336"></a><figcaption><strong>Share of Anthropic</strong>: Fable 5.2 took 20.6% of Anthropic’s spend on OpenRouter on day one, 7th of Anthropic’s 11 launches here and a little less than Fable 5 (23.6%) and Fable 5.1 (22.4%).</figcaption></figure>
<p>Nothing was taken from Opus 5.5 on day one: it processed 1,113 billion tokens, against 1,098 billion the Tuesday before. Fable 5.1 had already faded. Its last day in OpenRouter’s daily top 50 was October 28; the cut-off for that list has been 45 to 63 billion tokens a day since. And the traffic is mostly one product: OpenRouter’s page for the model attributes 111 billion tokens to Claude Code, 83% of what it lists by app (it gives no time window).</p>
<p>OpenRouter is a small window on Anthropic. Its spend there annualizes to about $2.4B, 3% of the roughly $81B run-rate in its IPO filing. What the window shows clearly is the split between tokens and dollars.</p>
<h2>Tokens against dollars on November 10</h2>
<figure><a href="https://06.ai/packs/fable-5-2#mix"><img src="https://06.ai/newsletter/fable-5-2/fig-04.png" alt="Tokens against dollars on November 10" width="694" height="320"></a><figcaption>The nine models with the most estimated spend on OpenRouter that day. Anthropic took <strong>44%</strong> of the spend from 7% of the tokens. DeepSeek’s V4.1 Flash 1017 carried 27% of the tokens; we price it at the cheapest host OpenRouter lists, which puts it at 3% of spend. At DeepSeek’s own prices it would be 7–13%.</figcaption></figure>
<p>One day settles nothing. Every premium launch on the tracker has run the same course: a few weeks as the most capable model on sale, then a cheaper model that is nearly as good, usually from the same lab, takes the volume. Fable 5’s best week was 8.0% of OpenRouter spend and Fable 5.1’s 8.9%; Opus 5.5, at 40% of the price, reached 33.4%. The tracker’s bar for a game-changer is 15% in the best of the first four weeks with 60% of that kept in week six, and its verdict comes in December.</p>
<p>What is different this time is the distance to the tier below. Fable 5 launched 4.3 points ahead of the best Opus on Epoch’s index and Fable 5.1 2.6 ahead; Opus 5.5 passed it three weeks later. Fable 5.2 is 5.6 points ahead of Opus 5.5, more than any single Opus update has added (the most is 4.6).</p>
<h2>A cut at the top of the list</h2>
<p><strong>Fable 5.2 is the first cut to a Fable list price, and it completes a repricing of every Claude tier in under five months.</strong> Haiku fell 90%, Sonnet 33%, Opus 20% and now Fable 20%. The cuts are deepest at the bottom, where open-weight models compete, and shallowest at the top, where nothing does.</p>
<p>The list is now a ladder that doubles at each rung: $2 / $10 for Sonnet, $4 / $20 for Opus, $8 / $40 for Fable, and fast mode doubles Fable again to $16 / $80. Cached input is the exception, at $0.20 per million tokens on both Opus 5.5 and Fable 5.2.</p>
<p>The top rung is not getting cheaper year on year. Anthropic’s most expensive model cost $75 per million output tokens in May 2025, $25 after Opus 4.5 that November, $50 when Fable 5 added a tier above Opus in June, and $40 now: 60% more than a year ago. The sticker at the frontier fell with a price cut and rose with a new tier, and the new tier was the bigger move.</p>
<h2>Anthropic’s list price by tier</h2>
<figure><a href="https://06.ai/packs/fable-5-2#ladder"><img src="https://06.ai/newsletter/fable-5-2/fig-05.png" alt="Anthropic’s list price by tier" width="694" height="374"></a><figcaption>List price per million output tokens of the newest model in each tier, log scale. Each dot is the launch that changed the tier’s price. Input prices moved in step.</figcaption></figure>
<p>The list today</p>
<figure><a href="https://06.ai/packs/fable-5-2#ladder"><img src="https://06.ai/newsletter/fable-5-2/fig-06.png" alt="Fable 5.2 — Input: $8 · Output: $40 · Cached input: $0.20 · Fast mode, in / out: $16 / $80 · Default effort: high · −20% with Fable 5.2, Nov 2026; Opus 5.5 — Input: $4 · Output: $20 · Cached input: $0.20 · Fast mode, in / out: $8 / $40 · Default effort: medium · −20% with Opus 5.5, Sep 2026; Sonnet 5.5 — Input: $2 · Output: $10 · Cached input: $0.10 · Default effort: high · −33% with Sonnet 5, Jun 2026; Haiku 5.5 — Input: $0.10 · Output: $0.50 · Cached input: $0.01 · Default effort: medium · −90% with Haiku 5.5, Oct 2026" width="694" height="333"></a></figure>
<p>Per million tokens. Haiku 5.5’s price is for prompts up to 100,000 tokens; longer ones cost $0.50 / $2.50. Batch requests are half price. Default effort is the API’s; in Claude’s apps and Claude Code Fable 5.2 defaults to medium.</p>
<p>Against rivals, $8 / $40 sits 20% under GPT-6 Astra’s $10 / $50. Google answered within a day: Gemini 4 Carbon became generally available on November 10 at $2 / $10 until November 30 and $4 / $20 after, which is Opus 5.5’s price.</p>
<h2>Five strengths at one price</h2>
<p><strong>The token price is no longer the price. Fable 5.2 is sold at five effort levels, and what a task costs depends more on the level than on the list.</strong> On Artificial Analysis’s index the same model costs $3.10 a task at low effort and $26.54 at max, 8.6 times as much, for 54 points and 65. On Cognition’s FrontierCode 2.0, as published in Anthropic’s system card, the range is $2.10 to $44.70: 21 times.</p>
<p><strong>Above about $3 a task, Fable 5.2 is nearly always the better buy than Opus 5.5 on that index, at twice the price per token.</strong> Fable at low effort scores 54 for $3.10; Opus at high scores 52 for $3.46. Fable at high scores 62 for $9.08; Opus at max scores 59 for $10.04. Only Opus at xhigh (57 for $6.05) keeps a place, between Fable’s medium (56 for $3.80) and its high.</p>
<h2>Score against cost per task</h2>
<p>Each dot is one model at one effort level. The first two views are the Artificial Analysis Intelligence Index v5.0.1 and its cost per index task, read November 11; the third is Cognition’s FrontierCode 2.0 as Anthropic’s system card reports it.</p>
<p>Fable at low effort (<strong>54 for $3.10</strong>) beats Opus at high (52 for $3.46) on score and cost; Fable at high (<strong>62 for $9.08</strong>) beats Opus at max (59 for $10.04). Hollow dots are settings another one beats on both.</p>
<figure><a href="https://06.ai/packs/fable-5-2#effort"><img src="https://06.ai/newsletter/fable-5-2/fig-07.png" alt="Score against cost per task" width="694" height="417"></a></figure>
<p>The two ladders</p>
<figure><a href="https://06.ai/packs/fable-5-2#effort"><img src="https://06.ai/newsletter/fable-5-2/fig-08.png" alt="low — 54 · $3.10 · 6 · 35 · $0.98 · 3 · 34.2% · $2.10; medium — 56 · $3.80 · 8 · 48 · $2.45 · 7 · 37.6% · $2.48; high — 62 · $9.08 · 18 · 52 · $3.46 · 10 · 45.3% · $6.72; xhigh — 62 · $14.83 · 26 · 57 · $6.05 · 16 · 49.1% · $14.19; max — 65 · $26.54 · 44 · 59 · $10.04 · 25 · 45.7% · $44.70" width="694" height="340"></a></figure>
<p>Greyed scores are settings another one beats on both score and cost. On FrontierCode 2.0, Opus 5.5’s best run (max) is 45.5% for $12.82, GPT-6 Astra’s 42.8% for $6.68 and Fable 5.1’s (xhigh) 32.4% for $8.68.</p>
<figure><a href="https://06.ai/packs/fable-5-2#effort"><img src="https://06.ai/newsletter/fable-5-2/fig-09.png" alt="Score against cost per task — Every model" width="694" height="448"></a><figcaption><strong>Every model</strong>: From 29 points to 55, OpenAI’s GPT-6.1 Terra and Sol are the cheapest way to every score. From 56 up, six settings are unbeaten, and Fable 5.2 has three of them. Open-weight models are in blue.</figcaption></figure>
<figure><a href="https://06.ai/packs/fable-5-2#effort"><img src="https://06.ai/newsletter/fable-5-2/fig-10.png" alt="Score against cost per task — FrontierCode 2.0" width="694" height="448"></a><figcaption><strong>FrontierCode 2.0</strong>: Anthropic’s own figures: at high effort Fable 5.2 matches Opus 5.5’s best (45.3% against 45.5%) for 48% less a task. At max it costs $44.70 a task and scores below xhigh.</figcaption></figure>
<p>Two other measures agree. On FrontierCode, Fable at high matches Opus at max (45.3% against 45.5%) for 48% less a task. On Vals AI’s index, which runs every model through one harness, Fable 5.2 scores higher than Opus 5.5 (71.3% against 70.1%) and costs 28% less a test ($17.94 against $24.77).</p>
<p>The top of the ladder is poor value. From high to max, Fable gains 3 index points for 2.9 times the money. Xhigh adds nothing over high on Artificial Analysis’s index, and on FrontierCode max scores below xhigh (45.7% against 49.1%) at 3.2 times the cost. Anthropic’s documentation says as much: start at high, move to xhigh only where a gain has been measured, and keep max for finding the model’s ceiling.</p>
<p>Anthropic’s advice is still to start most work on Opus 5.5 and move up to Fable for harder jobs. Below about $3 a task the numbers agree: Opus at medium, its default, scores 48 for $2.45 in about 7 minutes, and nothing Fable offers is that cheap. Above $3 the advice costs money. Fable is also rationed. On Pro and standard Team plans it is included for launch week only, for up to half of weekly limits. Our reading, which Anthropic hasn’t stated, is that capacity is part of why it steers volume to Opus; the <a href="https://06.ai/packs/anthropic-s1">Anthropic S-1</a> pack asks whether its compute can keep up with its demand.</p>
<p>Across labs the same thing holds: token prices rank models badly. GPT-6.1 Sol at max scores 55 for $1.27 a task, a point under Fable at medium for a third of the cost. Gemini 4 Carbon at high scores 60 for $7.76 at its standard price, half Fable’s per token: 85% of what Fable at high costs a task, for two points less. And the cheapest tokens are not the cheapest tasks. DeepSeek’s V4.1 Flash 1017 at max scores 33 for $0.60 a task; GPT-6.1 Terra at medium scores 38 for $0.17.</p>
<h2>What deflates and what doesn’t</h2>
<p><strong>Frontier cost deflation is three different prices, and they are moving in different directions.</strong></p>
<ul><li><strong>The sticker at the top</strong> moves in steps and is up on the year: $25 per million output tokens then, $40 now. Within a tier it has been falling 20% every five to ten months.</li><li><strong>A fixed result</strong> gets cheaper fast. A FrontierCode score of 45% cost $12.82 a task on September 22 (Opus 5.5 at max) and $6.72 on November 9 (Fable 5.2 at high): 48% less in 48 days. Fable 5.1’s best run, 32.4% for $8.68, is beaten by Fable 5.2 at low for $2.10: 76% less in 69 days. On Artificial Analysis’s index, the 50 points that cost $10.85 a task with Fable 5.1 at xhigh, its best setting, cost $0.59 with GPT-6.1 Sol at high: an 18th of the cost, from a model that arrived 28 days after Fable 5.1.</li><li><strong>The best available result</strong> has just got much dearer. The top score on the index cost $10.04 a task when Opus 5.5 took it in September and $7.76 when Gemini 4 Carbon took it in October. With Fable 5.2 at max it costs $26.54, 2.6 times what Opus 5.5’s cost, and the last 3 points account for $17.46 of it.</li></ul>
<p>Put together, a lab keeps its revenue per task at the frontier by selling more tokens per task at a slowly falling price, while everything behind the frontier falls toward what it costs to serve. OpenRouter’s own figures show the second half: the blended price paid per million tokens for closed models fell from $1.21 in January to $0.52 in October, and for open-weight ones from $0.28 to $0.084. Anthropic describes its own pricing in cost per task: its S-1 says Opus 5.5 cut customers’ cost per task by about 40% against Opus 5. The list price fell 20%.</p>
<p>For deflation from here, that means three things. The top list price keeps falling in small steps. New rungs keep appearing above it; fast mode already doubles it. And most of the fall in the cost of a given result comes from the tier below catching up, not from the price list.</p>
<h2>From frontier to affordable</h2>
<p><strong>Each of Anthropic’s last three flagships was matched at half its price within seven weeks, and at a tenth within five months.</strong> We take each model that launched at the top of Anthropic’s price list, the Opus models up to 4.8 and then the Fables, and find the first later model, from any lab, to match it on Epoch’s index at a fraction of its list price.</p>
<p>Half price has taken 21 to 145 days; for the last three, 42, 45 and 21. A tenth has taken 55 to 289 days, and less with almost every generation: 289 days for Opus 4.5, 216 for Opus 4.6, 174 for Opus 4.7, 132 for Opus 4.8, 139 for Fable 5 and 55 for Fable 5.1.</p>
<h2>How long each flagship took to be matched for less</h2>
<figure><a href="https://06.ai/packs/fable-5-2#stair"><img src="https://06.ai/newsletter/fable-5-2/fig-11.png" alt="How long each flagship took to be matched for less" width="694" height="362"></a><figcaption>Days from launch until a later model matched it on Epoch’s Capabilities Index at half, a fifth, a tenth and a fiftieth of its list price. Under each name is its launch price per million output tokens. Where one model cleared several thresholds on the same day, the deepest is shown.</figcaption></figure>
<p>Who matched it</p>
<figure><a href="https://06.ai/packs/fable-5-2#stair"><img src="https://06.ai/newsletter/fable-5-2/fig-12.png" alt="Opus 4.5 Nov 24, 2025 · $5 / $25 — Half the price: 23 days Gemini 3 Flash · A fifth: 23 days Gemini 3 Flash · A tenth: 289 days DeepSeek V4.1 Flash · A fiftieth: 302 days GPT-6 Luna; Opus 4.6 Feb 5, 2026 · $5 / $25 — Half the price: 145 days Sonnet 5 · A fifth: 189 days Gemini 3.7 Flash · A tenth: 216 days DeepSeek V4.1 Flash · A fiftieth: 229 days GPT-6 Luna; Opus 4.7 Apr 16, 2026 · $5 / $25 — Half the price: 84 days GPT-5.6 Terra · A fifth: 119 days Gemini 3.7 Flash · A tenth: 174 days Haiku 5.5 · A fiftieth: 174 days Haiku 5.5; Opus 4.8 May 28, 2026 · $5 / $25 — Half the price: 42 days GPT-5.6 Terra · A fifth: 132 days Haiku 5.5 · A tenth: 132 days Haiku 5.5 · A fiftieth: 132 days Haiku 5.5; Fable 5 Jun 9, 2026 · $10 / $50 — Half the price: 45 days Opus 5 · A fifth: 111 days Sonnet 5.5 · A tenth: 139 days GPT-6.1 Terra; Fable 5.1 Sep 1, 2026 · $10 / $50 — Half the price: 21 days Opus 5.5 · A fifth: 28 days GPT-6.1 Sol · A tenth: 55 days GPT-6.1 Terra; Fable 5.2 Nov 9, 2026 · $8 / $40" width="694" height="492"></a></figure>
<p>A match is an index score at or above the flagship’s, at a list price at or below the fraction on both input and output. Candidates are the 32 models whose list prices we could verify; a cheaper open-weight model outside that list may have got there sooner. Opus 5 and Opus 5.5 launched below Fable on the list, so they appear here only as the models that did the matching.</p>
<p>Opus 4.8 cost $5 / $25 on May 28. On October 7, 132 days later, Haiku 5.5 matched it on the index at $0.10 / $0.50, a fiftieth of the price. Fable 5.1 cost $10 / $50 on September 1; Opus 5.5 passed it at $4 / $20 three weeks later, GPT-6.1 Sol at $2 / $10 a week after that, and GPT-6.1 Terra at $1 / $5 on October 26.</p>
<p>The higher the sticker, the faster this goes: a tenth of Fable’s $50 is the price of a mid-tier model, and a tenth of Opus’s $25 was not. A match on the index is also not a match on everything, which the last sections come back to.</p>
<p>For Fable 5.2 the clock started on November 9. If it runs like the last three, a model at $4 / $20 or less matches it between the end of November and Christmas, and one at $0.80 / $4 by April. We’d bet on slower. Fable 5.2 opened 5.1 points clear of the next model on the index, the widest lead a new frontier model has had since o1 in December 2024.</p>
<h2>A new best model every six or seven weeks</h2>
<p><strong>Anthropic released nine models that set a new high for it on Epoch’s index in the twelve months to November 9, against four in the twelve months before.</strong> The median gap between them fell from 81 days to 45. One of the nine, Opus 5, set its high by 0.2 of a point, which is inside the index’s noise. Counting every tier, Anthropic launched 13 models against 6.</p>
<h2>Every Claude release since February 2025</h2>
<figure><a href="https://06.ai/packs/fable-5-2#cadence"><img src="https://06.ai/newsletter/fable-5-2/fig-13.png" alt="Every Claude release since February 2025" width="694" height="258"></a><figcaption>By tier. A filled dot set a new high for Anthropic on Epoch’s Capabilities Index when it launched; a hollow one did not. Mythos models, which are not on general sale, are left out.</figcaption></figure>
<p>The Fable line has moved twice, after 84 and 69 days; the Opus line moves every 42 to 73. Carried forward, that gives the table below: nine to twelve new best models in the next twelve months if the pace of the past year or half-year holds, four or five of them Fables, and about twice that in two years. These are extrapolations, not forecasts.</p>
<figure><a href="https://06.ai/packs/fable-5-2"><img src="https://06.ai/newsletter/fable-5-2/fig-14.png" alt="Back to 2025’s rhythm — New best models a year: 4.5 · Next 12 months: 5 · Next 24 months: 9; The past twelve months — New best models a year: 9.0 · Next 12 months: 9 · Next 24 months: 18; The past six months — New best models a year: 12.0 · Next 12 months: 12 · Next 24 months: 24" width="712" height="189"></a></figure>
<p><strong>Is it accelerating? On the evidence so far, yes.</strong></p>
<ul><li>The frontier gained 23.1 points on Epoch’s index in the twelve months to Fable 5.2, against 15.8 in the twelve months before and a trend of 16.6 since reasoning models arrived. Most of the difference is Fable 5.2 itself: the day before it shipped, the twelve-month gain was 18.0. Epoch says the model sits above the trend that earlier reasoning models followed. Its score is two days old, with an interval of 170 to 179.</li><li>METR, after an eleven-week assessment, estimates Anthropic is improving capabilities on well-specified tasks at about twice its rate of March 2024 to October 2025 (1.9 times on the benchmark trend, 95% interval 1.5–2.8). Its rough guess is that about 80% of that is due to AI, and it found no convincing bottleneck for the next twelve months. Anthropic redacted the evidence METR found most informative, and METR says it would not be surprised by 1.3 times or 3.</li><li>Anthropic’s amended S-1 says its internal estimates indicate it “may imminently meet” its own threshold for dramatically accelerated progress: roughly a doubling of the pace, driven by AI automating research. It also says no current model has crossed it. More than 90% of the changes merged into its core codebase are now written by Claude.</li></ul>
<p>Three things pull the other way. Dario Amodei’s September essay called for the industry to slow the pace at which it improves models; “We called for pacing, not pausing,” Dianne Penn, Anthropic’s head of product management for research, told CNBC about this launch. Fable 5.2 was held in training to about Opus 5.5’s level on two skills that most directly speed up AI development, accelerator kernel optimization and LLM architecture improvement. And the S-1 discloses that Anthropic delayed pre-training a new frontier base model by more than a month to harden security. Point releases will keep coming; the next generation is later than it would have been.</p>
<h2>What a year bought</h2>
<p><strong>Fable 5.2 scores 174.2 on Epoch’s index, 22.4 points above Opus 4.5, the best Anthropic model of a year ago. The 22 points before that took 17 months.</strong> Opus 4.5 arrived on November 24, 2025; on the exact anniversary Anthropic’s best was Sonnet 4.5, 26.6 points back.</p>
<p>The index is abstract, but Epoch’s model turns it into scores. On a benchmark where Opus 4.5 scored 20%, a model 22 points higher is expected to score between 60% and 97%, depending on how sharply that benchmark separates models. In odds that is 6 to 110 times.</p>
<h2>Epoch’s Capabilities Index since 2025</h2>
<figure><a href="https://06.ai/packs/fable-5-2#eci"><img src="https://06.ai/newsletter/fable-5-2/fig-15.png" alt="Epoch’s Capabilities Index since 2025" width="694" height="409"></a><figcaption>The best score to date on Epoch AI’s Capabilities Index, which combines 59 benchmarks on one scale (Claude 3.5 Sonnet = 130, GPT-5 = 150). The frontier is the best closed model from any lab. Scores carry 90% intervals: about two points either way for most models here, and more than four for Fable 5.2, which is new.</figcaption></figure>
<p>On Epoch’s own tests the change is plainer. Research-level maths (FrontierMath, tiers 1 to 3) went from 34% to 95%. Its hardest tier went from 5% to 100%. Facts recalled without a search (SimpleQA Verified) went from 46% to 94%. The older, easier tests barely register it: GPQA Diamond went from 86% to 94%.</p>
<h2>The same tests, a year apart</h2>
<figure><a href="https://06.ai/packs/fable-5-2#tests"><img src="https://06.ai/newsletter/fable-5-2/fig-16.png" alt="The same tests, a year apart" width="694" height="322"></a><figcaption>Epoch AI’s own runs, from its model pages. Opus 4.5 was Anthropic’s best a year ago. DeepSeek V4.1 Flash is the September version, the newest DeepSeek model Epoch lists test by test; the October refresh rates 2.9 points higher on the index.</figcaption></figure>
<p>The measure closest to work is the Remote Labor Index: 240 real freelance projects worth $143,991, averaging 29 hours of human time, each scored on whether a reasonable client would accept the result. Opus 4.5 delivered 3.75% of the 230 that are scored. Fable 5.2 delivers 41.3%, 11 times as many, and nearly twice the next model, Opus 5.5 at 21.25%.</p>
<h2>Freelance projects delivered to a client’s standard</h2>
<figure><a href="https://06.ai/packs/fable-5-2#rli"><img src="https://06.ai/newsletter/fable-5-2/fig-17.png" alt="Freelance projects delivered to a client’s standard" width="694" height="233"></a><figcaption>Remote Labor Index automation rate, from Scale AI’s leaderboard: Anthropic’s models in order of release. The best model from another lab is GPT-6 Astra, at 20.83%. Entries run in different agent harnesses.</figcaption></figure>
<p>The best-known yardstick can no longer be read. METR put Opus 4.5’s time horizon, the length of task it completes half the time, at 4 hours 49 minutes. In May it put an early Claude Mythos Preview at 16 hours or more, with a warning that anything above 16 hours is unreliable on its task suite. It stopped updating the page in September.</p>
<h2>Last year’s frontier against DeepSeek’s latest</h2>
<p><strong>DeepSeek’s latest model is clearly better than Anthropic’s best of a year ago, and about level with its best of last May.</strong> V4.1 Flash 1017, released on October 17 with open weights, scores 159.3 on Epoch’s index. That is 7.5 points above Opus 4.5, with intervals that don’t overlap, and within a point of Opus 4.8 (159.9) and Haiku 5.5 (160.3).</p>
<p>It costs $0.30 / $1.20 per million tokens at DeepSeek’s peak rate and half that off-peak, against the $5 / $25 Opus 4.5 launched at: about a twentieth of the price on output at peak and a fortieth off-peak, eleven months later. On Epoch’s own tests the September version already beat Opus 4.5 on research maths (67% against 34% on FrontierMath tiers 1 to 3), chess puzzles and mystery games, and roughly matched it on recall.</p>
<p>Against today’s frontier it has lost ground. A year ago DeepSeek’s best model trailed Anthropic’s best by 3.2 points on the index; today it trails by 14.9. DeepSeek gained 14.9 points in the year, a little under the frontier’s long-run pace. Anthropic gained 26.6.</p>
<p>The best open-weight model, Moonshot’s Kimi K3.1 Pro at 162.0, stands where the closed frontier stood on April 23, six months earlier, and where Anthropic stood in June. Epoch’s own published figure is shorter: four months on average since January. In points the gap was 7.1 the day before Fable 5.2 and is 12.2 now. Epoch counts Kimi K3.1 Pro as open-weight; when Artificial Analysis tested it in late October its weights were not yet out.</p>
<h2>Where benchmark parity ends</h2>
<p><strong>Open-weight models match the frontier on tests that are public, short or saturated, and fall far behind on ones that are new, private or long.</strong></p>
<p>Vals AI runs every model through the same harness. On its legal index DeepSeek’s V4.1 Flash 1017 scores 71.9% and Fable 5.2 69.2%, at $0.59 a test against $8.88. On legal research and finance the two are within three points, and on tax DeepSeek leads by twelve. On Terminal-Bench 4.0, hard multi-step work in a terminal, DeepSeek scores 31.8% and Fable 65.2%.</p>
<h2>DeepSeek’s latest against Fable 5.2, test by test</h2>
<figure><a href="https://06.ai/packs/fable-5-2#parity"><img src="https://06.ai/newsletter/fable-5-2/fig-18.png" alt="DeepSeek’s latest against Fable 5.2, test by test" width="694" height="346"></a><figcaption>DeepSeek V4.1 Flash 1017 and Claude Fable 5.2 on the same tests, ordered from DeepSeek’s largest lead to Fable’s. Nine are Vals AI’s runs (percent); the tenth is the Artificial Analysis Intelligence Index (points, both at max effort).</figcaption></figure>
<p>Epoch’s tests show the same split. On GPQA Diamond, DeepSeek’s September V4.1 Flash is four points behind Fable 5.2 (90% against 94%). On FrontierMath’s hardest tier it scores 27% against 100%; on spotting assembly mistakes in photographs, 34% against 90%; on recalling facts, 43% against 94%.</p>
<p>Artificial Analysis ran the experiment in October. When it rebuilt its index so that about 60% rests on private or held-out tests, the best open-weight model’s gap to the leader widened from 3 points to 12, and to 26 for the best model whose weights could be downloaded. Today DeepSeek’s latest scores 33 on it, about half of Fable 5.2’s 65.</p>
<p>Even a matched score can be local. Kimi K3.1 Pro is level with the frontier on Terminal-Bench 4.0: 63.6% against 65.2% for Fable 5.2 on Vals AI’s runs. On Terminal-Bench-Science, a newer set of 70 research tasks, Artificial Analysis has it at 34% against 59% for Opus 5.5 and 63% for GPT-6 Astra. Epoch finds the same by domain: four months behind in general, six to seven on cyber, and about seven on a visual task in September, when most Chinese open-weight models still took no images (DeepSeek’s V4.1 Flash does).</p>
<p>Length is where small gaps become large ones. A job of a thousand steps finishes only if every step does: at 99.9% reliability a step it finishes 37% of the time, and at 99.99%, 90%. That is how a model can look level on short tests and be unusable on long ones. Anthropic’s example at launch: a thousand Fable 5.2 agents formalized Fermat’s Last Theorem in 30 hours, while Opus 5.5 in the same set-up had done 26% after 18 hours and barely moved after 30. Opus 5.5 is six points behind Fable 5.2 on Artificial Analysis’s index, and ahead of it on Vals AI’s legal index (72.4% against 69.2%).</p>
<p>Buyers already price this in. Open-weight models carry most of OpenRouter’s paid tokens and, by OpenRouter’s count, about a quarter of its paid spend, and the price gap between closed and open-weight tokens widened from 4.3 times to 6.2 between January and October.</p>
<p>One thing may lengthen the lag. Anthropic says the defences it added this autumn cut the reasoning data that distillers harvest from Claude by 90% to 99%, named Zhipu and Moonshot among those doing it, and says Fable 5.2 has its strongest protections yet. If copying is part of how open-weight models stay four to six months behind, the next few months will show it.</p>
<h2>What else is in this launch</h2>
<ul><li><strong>Google answered in a day.</strong> At its standard price Gemini 4 Carbon costs $7.76 a task for 60 points on Artificial Analysis’s index; Fable at high costs $9.08 for 62. Until November 30 Carbon’s tokens are half price.</li><li><strong>The unrestricted version is not on general sale.</strong> Mythos 5.2 is the same model without Fable’s cybersecurity and biology safeguards, open to vetted organizations and, through Claude Security, to enterprise customers. On Fable those safeguards block about 55% less than Fable 5.1’s did.</li><li><strong>A drag on business use is being removed.</strong> Fable comes with 30 days of data retention for safety monitoring. Enterprise Frontier Safeguards, in private beta with Bridgewater, Salesforce and Stripe, runs the monitoring on data kept in the customer’s own infrastructure, which makes zero retention possible.</li><li><strong>It shipped below its own ceiling.</strong> Anthropic says it held two AI-research skills at about Opus 5.5’s level during training, so on those its best model is deliberately no better than the tier below. Agentic coding and general machine-learning research were not limited.</li><li><strong>It strays less.</strong> In Anthropic’s sandbox tests, Mythos 5.2 used credentials or files from outside its task in 3.1% of more than 1,800 scenarios, against 21% for Mythos 5.1 and 39% for Mythos 5. That matters for anyone leaving an agent to run for hours. The figures are Anthropic’s and haven’t been checked by anyone else.</li><li><strong>Fable as an advisor.</strong> Haiku, Sonnet or Opus can consult Fable on the hard steps of a task, which buys Fable’s tokens only where they matter. It is the price ladder used inside one job.</li><li><strong>The IPO.</strong> Anthropic set a range of $70–78 a share the day after this launch, which values it at up to $1.4T, with run-rate revenue of about $81B at the end of October, up from about $71B a month earlier.</li></ul>
<h2>What to watch</h2>
<p><strong>Weeks four and six on the tracker.</strong> 15% of OpenRouter spend by December 7, then 60% of that kept by December 21. The numbers update twice a day on the <a href="https://06.ai/trackers/frontier-launches">frontier launch tracker</a>.</p>
<p><strong>Opus 5.5’s share.</strong> If Fable’s middle settings are the better buy, some Opus spend should move up. On day one none did.</p>
<p><strong>The first cheaper match.</strong> History says a model at half Fable 5.2’s price matches it by Christmas. A 5-point lead says later.</p>
<p><strong>December 1,</strong> when Gemini 4 Carbon doubles to $4 / $20.</p>
<p><strong>The next open-weight release,</strong> the first trained since Anthropic’s distillation defences took hold.</p>
<p><strong>Epoch’s and METR’s next reports on acceleration,</strong> and the changes Anthropic is said to be considering to its Responsible Scaling Policy.</p>
<h2>What this is anchored on</h2>
<p>Figures are as reported. The OpenRouter dollar estimates, the dominated settings, the days-to-match table, the release counts and the scenarios are ours, built from these.</p>
<ul><li><strong>Fable 5.2 launch</strong> — Released November 9, 2026 at $8 / $40 per million tokens; fast mode at $16 / $80; outputs up to 30% faster than Fable 5.1; about 55% fewer safeguard blocks; advances in accelerator kernel optimization and LLM architecture held to about Opus 5.5’s level in training; more than 90% of changes merged into Anthropic’s core codebase written by Claude; most work should start on Opus 5.5; a thousand agents formalized Fermat’s Last Theorem in 30 hours · <a href="https://www.anthropic.com/claude-fable-and-mythos-5-2">Anthropic</a></li><li><strong>Price list</strong> — Input, output and cache-read prices for every Claude model; fast mode prices; batch at half price · <a href="https://platform.claude.com/docs/en/about-claude/pricing">Claude Platform pricing</a></li><li><strong>Release dates</strong> — Every Claude launch and price change since Claude 3.7 Sonnet, with dates · <a href="https://platform.claude.com/docs/en/release-notes/overview">Claude Platform release notes</a></li><li><strong>Effort guidance</strong> — API default is high; xhigh for the hardest long-running work, where a gain has been measured; max for measuring the model’s ceiling · <a href="https://platform.claude.com/docs/en/models/fable-5-2/whats-new-fable-5-2">Claude Platform docs</a></li><li><strong>Subscriptions</strong> — Fable included for Pro and standard Team seats for one week, November 9–16, up to 50% of weekly limits · <a href="https://support.claude.com/en/articles/17506494-claude-fable-5-2-launch-offers">Claude Help Center</a></li><li><strong>Score and cost by effort</strong> — Intelligence Index score and cost per task at every effort level for Fable 5.2, Opus 5.5 and nine other models · <a href="https://artificialanalysis.ai/leaderboards/models">Artificial Analysis leaderboard</a></li><li><strong>Fable 5.2 vs Gemini 4 Carbon</strong> — Fable 5.2: 54, 56, 62, 62 and 65 at $3.10, $3.80, $9.08, $14.83 and $26.54 a task; Carbon at standard price: 54, 58 and 60 at $7.19, $6.59 and $7.76 · <a href="https://artificialanalysis.ai/models/releases/comparisons/claude-fable-5-2-vs-gemini-4-carbon">Artificial Analysis</a></li><li><strong>Index v5</strong> — About 60% of the index now rests on private or held-out tests; the best open-weight model’s gap to the leader widened from 3 points to 12, and to 26 for the best with released weights; Terminal-Bench-Science (70 tasks): GPT-6 Astra 63%, Opus 5.5 59%, Kimi K3.1 Pro 34% · <a href="https://artificialanalysis.ai/articles/artificial-analysis-intelligence-index-v5">Artificial Analysis, Oct 27</a></li><li><strong>Fable 5.1 on the index</strong> — 50 at xhigh, its best setting, for $10.85 a task on v5.0.1 · <a href="https://artificialanalysis.ai/models/claude-fable-5-1-xhigh">Artificial Analysis</a></li><li><strong>FrontierCode 2.0</strong> — Cognition’s benchmark, 132 tasks: Fable 5.2 scores 34.2%, 37.6%, 45.3%, 49.1% and 45.7% from low to max effort at $2.10, $2.48, $6.72, $14.19 and $44.70 a task; Opus 5.5 at max 45.5% for $12.82; Fable 5.1 at xhigh 32.4% for $8.68; no uncertainty intervals · <a href="https://venturebeat.com/technology/anthropic-releases-claude-fable-5-2-and-mythos-5-2-pairing-new-record-high-benchmarks-with-20-api-cost-reduction">Anthropic’s system card, via VentureBeat</a></li><li><strong>FrontierCode 2.0</strong> — The low-effort row; 660 attempts in all, 22 of which timed out at max effort (45.7% falls to 44.4% if they count as failures) · <a href="https://llm-stats.com/blog/research/claude-fable-5-2-launch">LLM Stats</a></li><li><strong>Capabilities Index</strong> — ECI scores, 90% intervals and release dates for 239 models, updated November 10, 2026 (CC BY 4.0) · <a href="https://epoch.ai/data/eci_scores.csv">Epoch AI</a></li><li><strong>Capabilities Index</strong> — Scale fixed at Claude 3.5 Sonnet = 130 and GPT-5 = 150; the frontier has advanced 16.6 points a year (14.6–18.7) since the first reasoning models, 8.0 before · <a href="https://epoch.ai/publications/eci-bayesian-methodology">Epoch AI, Nov 4</a></li><li><strong>Capabilities Index</strong> — How a gap in ECI maps to an expected benchmark score; slopes of 0.08 (WeirdML) and 0.21 (ARC-AGI) per point · <a href="https://epoch.ai/data/eci-documentation/methodology">Epoch AI methodology</a></li><li><strong>Fable 5.2 on the index</strong> — A record 174, five points clear of GPT-6 Astra and above the trend earlier reasoning models followed · <a href="https://x.com/EpochAIResearch/status/2119903415363113088">Epoch AI, Nov 9</a></li><li><strong>Epoch’s own tests</strong> — Scores on the nine benchmarks Epoch has run on Fable 5.2, and the same tests for Opus 4.5, Haiku 5.5 and DeepSeek V4.1 Flash · <a href="https://epoch.ai/models/claude-fable-5-2">Epoch AI model pages</a></li><li><strong>Open-weight lag</strong> — Open-weight models have trailed the closed frontier by four months, or 8 points, on average since January 2026 · <a href="https://epoch.ai/data-insights/open-closed-eci-gap">Epoch AI, May 29</a></li><li><strong>Open-weight lag, cyber</strong> — The open-weight frontier on cyber sits where the closed frontier stood six to seven months earlier · <a href="https://epoch.ai/data-insights/open-closed-cyber-eci-gap">Epoch AI, Oct 16</a></li><li><strong>Open-weight lag, vision</strong> — About seven months behind on a furniture-assembly test, against 4.4 on the general index; image input still rare in Chinese open-weight models as of September · <a href="https://epoch.ai/publications/furniture-assembly">Epoch AI, Sep 23</a></li><li><strong>Remote Labor Index</strong> — Fable 5.2 41.30%, Opus 5.5 21.25%, GPT-6 Astra 20.83%, Opus 4.5 3.75%; 240 freelance projects worth $143,991, averaging 28.9 hours of human work, 230 of them held privately and scored · <a href="https://labs.scale.com/leaderboard/rli">Scale AI</a></li><li><strong>Vals Index</strong> — Fable 5.2 71.31% at $17.94 a test; Opus 5.5 70.14% at $24.77; DeepSeek V4.1 Flash 1017 61.26% at $0.66 · <a href="https://www.vals.ai/benchmarks/vals_index">Vals AI, Nov 9</a></li><li><strong>Terminal-Bench 4.0</strong> — Fable 5.2 and Opus 5.5 65.15%; Kimi K3.1 Pro 63.64%; DeepSeek V4.1 Flash 1017 31.82%; Opus 4.8 23.23% · <a href="https://www.vals.ai/benchmarks/terminal-bench-4">Vals AI, Nov 9</a></li><li><strong>Vals Legal Index</strong> — DeepSeek V4.1 Flash 1017 71.86% at $0.59 a test; Fable 5.2 69.17% at $8.88 · <a href="https://www.vals.ai/benchmarks/legal_index">Vals AI, Nov 9</a></li><li><strong>Other Vals tests</strong> — Tax Agent, Legal Research, Finance Index, Finance Agent, Vibe Code and Code Migration scores for both models · <a href="https://openrouter.ai/deepseek/deepseek-v4.1-flash-1017">OpenRouter model pages</a></li><li><strong>AI R&amp;D acceleration</strong> — Anthropic is improving capabilities on well-specified tasks at about twice its March 2024–October 2025 rate: 1.9 times on the benchmark trend (95% interval 1.5–2.8), METR would not be surprised by 1.3 or 3 times; its rough guess is that about 80% is due to AI; no convincing bottleneck for the next 12 months; the most informative evidence redacted by Anthropic · <a href="https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment/">METR, Nov 5</a></li><li><strong>Time horizons</strong> — Measurements above 16 hours are unreliable with the current task suite; the page is no longer actively updated · <a href="https://metr.org/time-horizons/">METR</a></li><li><strong>Time horizons</strong> — An early Claude Mythos Preview measured at 16 hours or more (95% interval 8.5 to 55), the top of what METR’s tasks can measure · <a href="https://the-decoder.com/metr-says-it-can-barely-measure-claude-mythos-palo-alto-networks-warns-of-autonomous-ai-attackers/">The Decoder</a></li><li><strong>Amended S-1</strong> — Revenue of $33.2B in the nine months to September 30, against $2.5B a year earlier; $895.4B of purchase commitments; Opus 5.5 cut customers’ cost per task about 40% against Opus 5 · <a href="https://www.sec.gov/Archives/edgar/data/0002130357/000162828026068676/anthropicpbc-sx1ax1.htm">Anthropic S-1/A, Nov 2</a></li><li><strong>Amended S-1</strong> — Internal estimates indicate Anthropic “may imminently meet” its threshold for dramatically accelerated AI-driven progress, roughly a doubling in the pace of AI progress, and says no current model has crossed it; IPO range of $70–78, up to $1.4T; run-rate revenue of about $81B at the end of October, up from about $71B in late September · <a href="https://fortune.com/2026/11/10/anthropic-ipo-price-range-ai-stock-sell-off/">Fortune, Nov 10</a></li><li><strong>S-1</strong> — Anthropic “recently” delayed pre-training a new frontier base model by more than a month to harden security against theft of model weights · <a href="https://upstox.com/news/business-news/latest-updates/anthropic-paused-next-gen-ai-model-training-for-over-a-month-over-security-concerns/article-202453/">Upstox, Oct 27</a></li><li><strong>Pacing</strong> — Dario Amodei’s essay “We Must Pace the Frontier” (September 12): the industry should slow the pace at which it improves models, without halting training; outside evaluators embedded in the labs · <a href="https://x.com/DarioAmodei/status/2098773920774074715">Dario Amodei</a></li><li><strong>Pacing</strong> — “We called for pacing, not pausing,” Dianne Penn, Anthropic’s head of product management for research, to CNBC · <a href="https://www.implicator.ai/anthropic-fable-5-2-amodei-pacing-ipo/">Implicator</a></li><li><strong>Sandbox tests</strong> — Across more than 1,800 scenarios Mythos 5.2 used credentials or files from outside its task in 3.1%, against 21% for Mythos 5.1 and 39% for Mythos 5 (Anthropic’s figures, not independently verified) · <a href="https://www.implicator.ai/anthropic-fable-5-2-amodei-pacing-ipo/">Implicator</a></li><li><strong>Distillation</strong> — New defences cut the reasoning data distillers harvest from Claude by roughly 90–99%; Zhipu and Moonshot named · <a href="https://www.anthropic.com/news/distillation-defenses">Anthropic, Oct 12</a></li><li><strong>Gemini 4 Carbon</strong> — Generally available November 10 at $2 / $10 through November 30, then $4 / $20 · <a href="https://venturebeat.com/technology/google-gemini-4-carbon-is-now-publicly-available-and-it-undercuts-claude-fable-5-2-and-opus-5-5-on-api-price">VentureBeat</a></li><li><strong>OpenRouter, October</strong> — 790T tokens, up 37% on September; blended price per million tokens down from $1.21 to $0.52 for closed models and $0.28 to $0.084 for open-weight ones since January; cache hit rates above 90% for agents · <a href="https://openrouter.ai/data/october-2026">OpenRouter</a></li><li><strong>OpenRouter apps</strong> — Apps sending Fable 5.2 the most tokens: Claude Code 111B, Hermes Agent 9.22B · <a href="https://openrouter.ai/anthropic/claude-fable-5.2">OpenRouter</a></li><li><strong>DeepSeek prices</strong> — V4.1 Flash 1017 at $0.30 / $1.20 per million tokens on DeepSeek’s own endpoint, half that off-peak; accepts images · <a href="https://openrouter.ai/deepseek/deepseek-v4.1-flash-1017">OpenRouter</a></li></ul>
<h2>Method</h2>
<p><strong>OpenRouter.</strong> Tokens are OpenRouter’s daily top 50 models, through November 10. Spend is our estimate: each model’s tokens at its list price that day, assuming 10% of tokens are output and 50.7% of input is billed at the cached rate, the mix every 06.ai tracker uses. The listing day is a partial day and stays out. Open-weight models are priced at the catalog price OpenRouter shows, which is the cheapest host’s and understates what is paid for them; the range for DeepSeek uses its own off-peak and peak prices.</p>
<p><strong>Effort ladders.</strong> Scores and cost per task are the Artificial Analysis Intelligence Index v5.0.1 as its leaderboard showed them on November 11, with Gemini 4 Carbon at its standard price. Fable 5.1’s score and cost are from its Artificial Analysis page. FrontierCode 2.0 figures are from Anthropic’s system card as reported; Cognition ran the benchmark and gave no uncertainty intervals. At max effort 22 of 660 attempts timed out, and counting those as failures lowers that score from 45.7% to 44.4%. A setting is “beaten” when another scores at least as much for no more money. Effort settings are each vendor’s own and aren’t equivalent across labs.</p>
<p><strong>Epoch’s index.</strong> Scores are from Epoch AI’s file of November 10, under its Bayesian method of November 4. The frontier is the best closed or limited-access model to date. Fable 5.2’s score is two days old and rests on the nine benchmarks Epoch has run on it so far; scores move as more results come in, and its interval overlaps GPT-6 Astra’s. Our lags are strict: the date a closed model first scored at or above an open-weight model’s point estimate. That gives longer lags than the four-month average Epoch publishes. The conversion to scores uses the two benchmark slopes Epoch’s documentation gives, 0.08 and 0.21 per point.</p>
<p><strong>Days to match.</strong> For each model that launched at the top of Anthropic’s price list (Opus 4.5 to 4.8, then the Fables), the first later model with an index score at or above it and a list price at or below the fraction on both input and output. Prices are list prices at launch; later cuts aren’t counted. The comparison rests on point estimates whose intervals overlap, so a match means level within the noise.</p>
<p><strong>Cadence.</strong> A new best model is one that scored above every earlier Anthropic model on Epoch’s index when it launched; two launches on one day count once. Windows run from Nov 10, 2025 to Nov 9, 2026 and the same span a year earlier. The scenarios multiply a pace by one and two years and are not forecasts. Mythos models are excluded because they aren’t generally available.</p>
<p><strong>Tests.</strong> Epoch’s test-by-test scores are from its model pages on November 11. Vals AI’s are from its leaderboards of November 9 where noted and otherwise as shown on OpenRouter’s model pages. The Remote Labor Index is Scale AI’s leaderboard. Benchmarks Anthropic reported about its own model are described as Anthropic’s.</p>
<p><strong>Left out.</strong> Claude’s own apps and API, where most of Anthropic’s revenue is: OpenRouter is about 3% of it and leans toward developers. What Fable costs Anthropic to serve. Any test of whether the capacity reading is right.</p>
<p><strong>Disclosure.</strong> This analysis was produced with Claude Fable 5.2, the model it is about.</p>
<hr>
<p>Every chart above links to its interactive version on 06.ai, with the data behind it. The daily trackers are at <a href="https://06.ai/trackers">06.ai/trackers</a>.</p>
<p><em>Not investment advice. Estimates are labeled as estimates.</em></p>]]></content:encoded>
</item>
<item>
  <title>When does replacing old GPUs beat building new power?</title>
  <link>https://inthesingularity.ai/packs/gpu-decommission</link>
  <guid isPermaLink="true">https://inthesingularity.ai/packs/gpu-decommission</guid>
  <pubDate>Sun, 08 Nov 2026 12:24:00 GMT</pubDate>
  <description>A megawatt of H100s still earns 2.6–6.5 times what a new powered megawatt costs, so where power can be had, building it beats ripping Hopper out. Where it can’t arrive in time, the megawatt is worth what the newest chips would earn in it: on that test, swapping H100s for B300s has paid since mid-March at on-demand prices, though not at three-year prices. If compute stops being scarce, Hopper lasts until renters pay about 3.4 times as much per kilowatt for the newest chips; today they pay 1.7–2.1 times. Replacement becomes half of new GPU deployments once powered capacity grows less than about 15–26% a year.</description>
  <category>AI infrastructure</category>
  <enclosure url="https://inthesingularity.ai/newsletter/gpu-decommission/cover.png" length="64162" type="image/png" />
  <media:content url="https://inthesingularity.ai/newsletter/gpu-decommission/cover.png" medium="image" type="image/png" />
  <content:encoded><![CDATA[<figure><a href="https://06.ai/packs/gpu-decommission"><img src="https://06.ai/newsletter/gpu-decommission/thumb.png" alt="When does replacing old GPUs beat building new power?" width="421" height="391"></a><figcaption><a href="https://06.ai/packs/gpu-decommission">Interactive version and data on 06.ai →</a></figcaption></figure>
<ul><li><strong>6.5×</strong> — A megawatt of H100s earns the cost of a new powered megawatt, at on-demand prices. <strong>2.6×</strong> at implied three-year prices</li><li><strong>0.80×</strong> — H100 against the power-wall line. $3.90 an hour on demand against a <strong>$4.85</strong> line set by the B300. Below it every week since Mar 16</li><li><strong>3.4×</strong> — Where Hopper loses even with power to spare the B300’s rent per kW over the H100’s. Today <strong>1.7–2.1×</strong></li><li><strong>15–32%</strong> — Replacement share of new GPU megawatts at about 46% a year growth in US data-center power. <strong>Half</strong> once growth slows to 15–26%</li></ul>
<h2>The rule</h2>
<p><strong>An old fleet should keep its megawatt while its cash margin there beats what that megawatt is worth to the next use.</strong> If new power can arrive in time, that is the cost of a new powered megawatt. If it can’t, it is what the newest chips would clear in this one after paying for themselves. Size enters only through which of the two applies: a bigger build-out makes the next megawatt slower to get, which moves more sites from the first case to the second.</p>
<p>Divided by the kilowatt-hours a megawatt delivers in a year, each case becomes a price per kWh of IT capacity, and any GPU’s rent can be set against it. A GPU’s rent per kWh is its hourly rent over its power, counting its share of the server and network. There are three lines:</p>
<ul><li><strong>Power available</strong>, $0.56 per kWh: a new powered megawatt plus power and upkeep. Below it, a GPU isn’t worth its slot anywhere.</li><li><strong>Power wall</strong>: the newest GPU’s rent per kWh less its capital charge, $1.33 for a B300. Below it, a site that can’t get new power in time earns more by swapping in the newest chips.</li><li><strong>New chips idle</strong>: the newest GPU’s whole rent per kWh. Below it, frontier chips already bought and waiting for power should take the slot.</li></ul>
<h2>What a megawatt earns</h2>
<p><strong>A megawatt of H100s clears $7.9M–$19.5M a year after power and upkeep; a new powered megawatt costs about $3.0M a year.</strong> At every point on the price curve, Hopper out-earns its slot.</p>
<h2>What a megawatt of H100s earns a year</h2>
<p>H100 rents from 06.ai’s <a href="https://06.ai/trackers/gpu-compute-price-index">GPU Compute Price Index</a> at the Nov 5, 2026 fixing: the trimmed mean across providers on demand, and term prices implied from providers’ own term-to-on-demand ratios. A megawatt holds about 714 H100s at 1.4 kW each, server and network included, rented 85% of the hours.</p>
<figure><a href="https://06.ai/packs/gpu-decommission#per-mw"><img src="https://06.ai/newsletter/gpu-decommission/fig-01.png" alt="On-demand composite — $ per GPU-hour: $3.90 · Revenue a year: $20.7M · After power and upkeep: $19.5M · × a new powered megawatt: 6.5×; One-year, implied — $ per GPU-hour: $2.69 · Revenue a year: $14.3M · After power and upkeep: $13.1M · × a new powered megawatt: 4.4×; Three-year, implied — $ per GPU-hour: $1.70 · Revenue a year: $9.1M · After power and upkeep: $7.9M · × a new powered megawatt: 2.6×" width="694" height="251"></a></figure>
<p>The new megawatt is mostly building, not power. <a href="https://epoch.ai/data-insights/ai-datacenter-cost-breakdown">Epoch AI</a> puts a 1 GW AI site at $38B upfront and $8.5B a year all-in, with servers about $5B and operations $0.9B of that. The rest annualizes to about $2.6M per MW, between <a href="https://build.inc/insights/hyperscale-data-center-lease-terms-2026">hyperscale leases and wholesale colocation</a> at $100–250 per kW-month.</p>
<p>Generation is the small part. <a href="https://www.bloomberg.com/news/articles/2026-04-23/cost-to-build-natural-gas-fired-power-plant-surges-66-bnef-says">BNEF put new combined-cycle gas plants</a> at $2,157/kW in 2025, and <a href="https://247wallst.com/investing/2026/10/29/ai-data-centers-have-driven-gas-turbine-prices-up-50/">GE Vernova’s CEO expects</a> 2026 gas-turbine orders to price about 50% higher per kW than 2025’s. Even at $5,000/kW, more than twice the BNEF figure, on-site generation adds only about $0.6M per MW-year over 20 years.</p>
<p>Spread over the kilowatt-hours a megawatt delivers, the slot and upkeep ($3.0M + $1.2M) come to <strong>$0.56 per kWh of IT capacity</strong>. An H100 or H200 keeps its slot above $0.79 an hour, an A100 above $0.51, a B200 above $1.07. Every GPU in the index clears that by 1.8× or more, on every price basis.</p>
<h2>Rent per kWh against the three lines</h2>
<p>Each GPU’s rent per GPU-hour over its power in kW, server and network share included, at the Nov 5, 2026 fixing. The GB200 is left out: too few providers quote it. Term prices are implied from providers’ term-to-on-demand ratios; the B300’s rest on one or two providers.</p>
<p>On on-demand prices, <strong>4 of 7</strong> GPUs rent below the power-wall line the B300 sets, the H100 at <strong>0.80×</strong> of it. Every one clears the power-available line, the lowest (L40S) by 4.0×.</p>
<figure><a href="https://06.ai/packs/gpu-decommission#lines"><img src="https://06.ai/newsletter/gpu-decommission/fig-02.png" alt="Rent per kWh against the three lines" width="694" height="339"></a></figure>
<h2>A hard power wall</h2>
<p><strong>When new power arrives after the old chips’ earning life ends, a megawatt is worth what the newest chips would earn in it, net of their own capital cost.</strong> By that measure, at on-demand prices, swapping H100s for B300s has paid since mid-March 2026.</p>
<p>The test, per kWh: swap when the newest chip’s rent beats the incumbent’s by more than the newest chip’s capital charge. For a B300 that charge is <strong>$1.33 per kWh</strong>: about $63,000 per GPU all-in over 2.0 kW, recovered over 4 years at 10%. The server price is <a href="https://www.amcompute.com/gpu-prices">American Compute’s</a> $400,000–450,000 per new 8-GPU B300 machine at volume, plus about 20% for network, storage and installation.</p>
<p>On Nov 5 a B300 rented for $9.59 an hour on demand, or $4.80 per kWh. That puts the power-wall line at $3.46 per kWh: $4.85 an hour for an H100, which rented for $3.90, or <strong>0.80×</strong> of its line. A100s sat lower still at 0.71×; MI300Xs (1.06×), H200s (1.13×) and B200s (1.18×) cleared it.</p>
<h2>A year against the power-wall line</h2>
<p>GetDeploying’s weekly median on-demand rent per kWh, against the power-wall line the B300 set that week: its own rent per kWh less today’s $1.33 capital charge. Shaded: weeks the GPU rented below the line. Only one to five providers listed the B300 before April 2026, so read the early weeks loosely.</p>
<p>The H100 has rented below its power-wall line every week since <strong>Mar 16, 2026</strong>, at <strong>0.81×</strong> of it in the week of Nov 2.</p>
<figure><a href="https://06.ai/packs/gpu-decommission#history"><img src="https://06.ai/newsletter/gpu-decommission/fig-03.png" alt="A year against the power-wall line" width="694" height="322"></a></figure>
<p>The crossing came when B300 rents jumped in March. In GetDeploying’s weekly medians the B300 went from $5.74 to $7.13 an hour between Mar 2 and Mar 16, when only 4 providers listed it. H100s have rented below their line every week since.</p>
<p>The three-year curve disagrees. At implied three-year prices the B300’s premium over the H100 shrinks to $0.95 per kWh, below its capital charge, and the H100 clears its line by 1.46×. That curve rests on a single provider’s B300 term ratio, but it says the premium is priced to fade before a B300 pays for itself.</p>
<p>If the new chips are already bought and idle, there’s no capital charge left to clear. The line becomes the B300’s whole rent, $6.72 an hour for an H100. That was Microsoft’s position when <a href="https://www.techzine.eu/news/infrastructure/135966/satya-nadella-power-shortage-not-gpu-inventory-is-the-problem">Satya Nadella said on the BG2 podcast in October 2025</a> that it had chips in inventory and no powered shells to plug them into.</p>
<p>The wall is real for the next few years. <a href="https://www.sec.gov/Archives/edgar/data/0001996810/000199681026000162/gev3q2026form8-k10282026.pdf">GE Vernova’s gas-power backlog and slot reservations</a> reached 122 GW at the end of the third quarter of 2026, and management says it has sold slots for 2032 delivery. <a href="https://www.amcompute.com/appraisals/h100">American Compute’s conservative residual forecast</a> puts used H100s at 11–19% of list in 2027 and 5–7% in 2028, so most of Hopper’s earning life ends before those turbines arrive. Evicted chips still have buyers: a used 8-GPU H100 node was worth about $195,000 in August.</p>
<h2>If compute stops being scarce</h2>
<p><strong>Then Hopper survives as long as renters value its kilowatt at more than about 30% of the newest chip’s.</strong> Today they value it at about half.</p>
<p>In that world the newest chip rents at its full cost: its $1.33 capital charge plus the $0.56 slot and upkeep, or $1.90 per kWh. If renters keep paying the same ratio between generations, an old GPU rents at $1.90 divided by that ratio. It keeps its slot while that beats $0.56, so the break comes at a ratio of <strong>3.4×</strong>.</p>
<p>Today’s prices are far from that equilibrium. A B300 rents for $4.80 per kWh on demand, 2.5× its full cost; even its implied three-year price ($2.17) sits 14% above. Per kilowatt, though, a B300 rents for only 1.7–2.1× an H100, depending on the price basis. If scarcity ended at those ratios, an H100 would rent for about $1.25–1.54 an hour, still well above its $0.79 line.</p>
<p>Vendor claims point the other way. <a href="https://developer.nvidia.com/blog/nvidia-vera-rubin-and-blackwell-set-a-new-standard-for-agentic-ai-performance-per-watt/">In Nvidia’s own tests</a>, GB300 NVL72 delivers up to 15x the throughput per megawatt of an H200 NVL8 on DeepSeek V4 Pro, and <a href="https://blogs.nvidia.com/blog/vera-rubin/">Vera Rubin NVL72 up to 10x the tokens per megawatt</a> of GB200 NVL72. Those are chosen benchmarks at favorable operating points, but even half the claimed multiples clears 3.4×.</p>
<p>The gap is about what’s being sold. GPU-hours rent across a mix of training, fine-tuning and smaller models, where newer chips win by far less than on frontier mixture-of-experts inference. If you rent out GPUs by the hour, Hopper survives the end of scarcity. If you sell frontier-model tokens, it doesn’t.</p>
<p>The tell is the per-kilowatt rent ratio between the newest chip and Hopper. When it climbs toward 3.4×, the market has started paying for tokens rather than hours, and Hopper is done even with power to spare.</p>
<h2>The halls don’t fit</h2>
<p><strong>Rack power has more than tripled from Hopper to Blackwell and is headed for about 600 kW, so most Hopper halls can’t take the racks that would replace Hopper without new electrical gear and liquid cooling.</strong> The scarce asset is the grid connection, not the building.</p>
<p><a href="https://www.moduledge.com/blog/nvidia-vera-rubin">Rack power runs</a> from about 30–40 kW for Hopper to 120–140 kW for GB200 and GB300 NVL72 and an estimated 190–230 kW for Vera Rubin NVL72, with Rubin Ultra’s Kyber rack at about 600 kW in 2027. Halls built for earlier densities need their power and cooling rebuilt before they can host Rubin-class racks.</p>
<p>The retrofit is cheap next to what the slot earns. <a href="https://stlpartners.com/press/liquid-cooling-retrofits-can-cost-roughly-80-less/">STL Partners puts a liquid-cooling retrofit</a> at about $2M per MW, against $11M or more for new liquid-cooled capacity. The cost that bites is time: the hall earns nothing while it’s rebuilt.</p>
<p>The used market has priced this in. <a href="https://www.amcompute.com/appraisals/h100">In 2025 H100 rents fell while the median used H100 price rose</a>, because many data centers couldn’t yet host Blackwell; the old chips were worth more to operators whose halls fit nothing newer.</p>
<p>So swaps start in liquid-ready halls, and air-cooled Hopper halls keep air-cooled gear until the building itself is redeveloped. Over a few generations the building becomes a consumable, rebuilt around a durable interconnect. Resale value and retrofit cost shift the line at the margin: a liquid secondary market lowers the bar to evict, and a retrofit raises it.</p>
<h2>Where scale matters</h2>
<p><strong>Scale enters through the growth rate of powered capacity, not its size.</strong> If powered capacity grows by g a year and GPUs leave after L years, this year’s replacements are the deployments of L years ago, so replacement’s share of this year’s new megawatts is (1 + g)<sup>−L</sup>. It reaches half once growth slows to about 15–26% a year.</p>
<h2>Replacement share of new GPU megawatts</h2>
<figure><a href="https://06.ai/packs/gpu-decommission#replacement"><img src="https://06.ai/newsletter/gpu-decommission/fig-04.png" alt="Replacement share of new GPU megawatts" width="694" height="348"></a><figcaption>Share of each year’s new GPU megawatts that replace retired ones, (1 + g)<sup>−L</sup>, by annual growth in powered capacity g and chip life L. The marker is the average growth in Goldman Sachs’s US data-center power demand from 2025 to 2027.</figcaption></figure>
<p><a href="https://www.goldmansachs.com/insights/articles/us-data-center-power-demand-projected-to-double-by-2027">Goldman Sachs sees US data-center power demand</a> rising from 31 GW in 2025 to about 66 GW in 2027, an average of about 46% a year. At that pace replacement is 15–32% of deployments, depending on chip life. Power scarcity shortens that life, which pulls the crossover forward.</p>
<p>Site size cuts the other way. Frontier training needs contiguous multi-gigawatt campuses that no legacy site can provide, so the frontier stays greenfield. In-place refresh is an inference-fleet story.</p>
<p>And decommissioning rarely means scrap. The efficient outcome is a cascade: hyperscalers keep dense, power-scarce slots for the newest silicon and sell older GPUs to operators with air-cooled halls or stranded power. Aggregate power demand keeps rising; the market just tiers.</p>
<h2>What it means</h2>
<p><strong>The decommission wave starts at walled-off sites with liquid-ready halls, and two spreads set its timing, not the size of the build-out.</strong></p>
<ul><li><strong>Rent migrates to megawatts.</strong> Where power binds, a slot is worth the newest chip’s profit, so whoever owns the interconnect captures the scarcity rent. The power-wall line has risen with every B300 price increase this year.</li><li><strong>Nvidia’s demand grows a replacement leg.</strong> At 46% annual growth in powered capacity, replacement is 15–32% of deployments; at 20%, it’s 40–58%.</li><li><strong>Hopper’s book value is the exposure.</strong> <a href="https://www.sec.gov/Archives/edgar/data/1018724/000101872425000004/amzn-20241231.htm">Amazon cut the useful life of a subset of its servers and networking gear</a> from six years to five in January 2025, citing AI’s pace. <a href="https://www.fool.com/earnings/call-transcripts/2026/01/28/microsoft-msft-q2-2026-earnings-call-transcript/">Microsoft’s CFO says</a> about two-thirds of its capex goes to short-lived assets, mainly GPUs and CPUs, depreciated over six years; most of that capital is contracted for most of its useful life. A walled-site swap writes Hopper off early, so useful-life changes and impairments are the accounting tell.</li><li><strong>Used-GPU prices carry the cascade.</strong> It works only while buyers with power pay around $195,000 per H100 node, and residual forecasts narrow that window through 2028.</li></ul>
<h2>What to watch</h2>
<p><strong>The spread between the newest chip and Hopper,</strong> per kWh, against the newest chip’s $1.33 capital charge. Above it, walled sites should be swapping.</p>
<p><strong>The ratio between them,</strong> per kW, against 3.4×. Above it, Hopper loses even with power to spare.</p>
<p><strong>Hopper’s rent per kWh</strong> against $0.56. Below it, Hopper loses its slot anywhere.</p>
<p><strong>Turbine lead times,</strong> which decide how many sites sit behind a wall.</p>
<p><strong>Rubin’s first rents.</strong> When enough providers list it, the newest chip moves from the B300 to Rubin, and the power-wall and new-chips-idle lines reset with it. The rents behind all of these update every day in the <a href="https://06.ai/trackers/gpu-compute-price-index">GPU Compute Price Index</a>.</p>
<h2>What this is anchored on</h2>
<p>Figures are as reported. The lines, the power per GPU, the B300’s all-in cost and the replacement shares are ours, built from these.</p>
<ul><li><strong>GPU rents</strong> — Composite rental price by GPU and price basis at the Nov 5, 2026 16:00 UTC fixing; neocloud tier composite; term prices implied from providers’ own term-to-on-demand ratios · <a href="https://06.ai/trackers/gpu-compute-price-index">06.ai GPU Compute Price Index</a></li><li><strong>Weekly rents</strong> — Weekly median on-demand price by GPU, Nov 2025 to Nov 2026 (CC BY 4.0) · <a href="https://getdeploying.com/dataset/gpu-prices">GetDeploying GPU price dataset</a></li><li><strong>Data-center cost</strong> — A 1 GW AI data center: $38B upfront, $8.5B a year all-in, servers about $5.0B a year and operations $0.9B · <a href="https://epoch.ai/data-insights/ai-datacenter-cost-breakdown">Epoch AI</a></li><li><strong>Lease rates</strong> — Hyperscale about $100–150 per kW-month; wholesale colocation $150–250 · <a href="https://build.inc/insights/hyperscale-data-center-lease-terms-2026">Build</a></li><li><strong>Gas plants</strong> — New combined-cycle plants cost $2,157/kW in 2025 · <a href="https://www.bloomberg.com/news/articles/2026-04-23/cost-to-build-natural-gas-fired-power-plant-surges-66-bnef-says">Bloomberg, on BNEF</a></li><li><strong>Turbine prices</strong> — GE Vernova’s CEO expects full-year 2026 gas orders to price about 50% higher per kW than 2025’s (Q3 call, Oct 28, 2026) · <a href="https://247wallst.com/investing/2026/10/29/ai-data-centers-have-driven-gas-turbine-prices-up-50/">24/7 Wall St.</a></li><li><strong>Turbine backlog</strong> — Gas Power backlog and slot reservation agreements 122 GW at the end of Q3 2026; slots sold for 2032 delivery · <a href="https://www.sec.gov/Archives/edgar/data/0001996810/000199681026000162/gev3q2026form8-k10282026.pdf">GE Vernova Q3 2026 results, Form 8-K</a></li><li><strong>B300 servers</strong> — A new 8-GPU B300 server: $400,000–450,000 at volume · <a href="https://www.amcompute.com/gpu-prices">American Compute</a></li><li><strong>Used H100s</strong> — A used 8-GPU HGX H100 node about $195,000 (August 2026); conservative residuals 11–19% of list in 2027 and 5–7% in 2028; in 2025 rents fell while the median resale price rose · <a href="https://www.amcompute.com/appraisals/h100">American Compute</a></li><li><strong>Chips without power</strong> — Microsoft had chips in inventory it couldn’t plug in for want of powered shells (BG2 podcast, Oct 31, 2025) · <a href="https://www.techzine.eu/news/infrastructure/135966/satya-nadella-power-shortage-not-gpu-inventory-is-the-problem">Techzine</a></li><li><strong>Throughput per MW</strong> — GB300 NVL72 up to 15x the throughput per megawatt of H200 NVL8 on DeepSeek V4 Pro, in Nvidia’s own tests · <a href="https://developer.nvidia.com/blog/nvidia-vera-rubin-and-blackwell-set-a-new-standard-for-agentic-ai-performance-per-watt/">NVIDIA Technical Blog</a></li><li><strong>Tokens per MW</strong> — Vera Rubin NVL72 up to 10x the tokens per megawatt of GB200 NVL72 · <a href="https://blogs.nvidia.com/blog/vera-rubin/">NVIDIA</a></li><li><strong>Rack power</strong> — About 30–40 kW for Hopper, 120–140 kW for GB200 and GB300 NVL72, an estimated 190–230 kW for Vera Rubin NVL72, about 600 kW for Kyber in 2027 · <a href="https://www.moduledge.com/blog/nvidia-vera-rubin">ModuEdge</a></li><li><strong>Retrofits</strong> — A liquid-cooling retrofit about $2M per MW, against upwards of $11M for new liquid-cooled capacity · <a href="https://stlpartners.com/press/liquid-cooling-retrofits-can-cost-roughly-80-less/">STL Partners</a></li><li><strong>Power demand</strong> — US data-center power demand 31 GW in 2025, 41 GW in 2026, 66 GW in 2027 · <a href="https://www.goldmansachs.com/insights/articles/us-data-center-power-demand-projected-to-double-by-2027">Goldman Sachs</a></li><li><strong>Useful lives</strong> — Amazon cut a subset of servers and networking equipment from six years to five from Jan 1, 2025 · <a href="https://www.sec.gov/Archives/edgar/data/1018724/000101872425000004/amzn-20241231.htm">Amazon 2024 Form 10-K</a></li><li><strong>Useful lives</strong> — Microsoft: about two-thirds of capex on short-lived assets, mainly GPUs and CPUs, depreciated over six years · <a href="https://www.fool.com/earnings/call-transcripts/2026/01/28/microsoft-msft-q2-2026-earnings-call-transcript/">Microsoft FY26 Q2 call (Jan 28, 2026)</a></li><li><strong>System power</strong> — DGX A100 6.5 kW, DGX H100 10.2 kW, DGX B200 14.3 kW, DGX B300 about 15 kW maximum · <a href="https://docs.nvidia.com/dgx/dgxh100-user-guide/introduction-to-dgxh100.html">NVIDIA DGX user guides</a></li><li><strong>System power</strong> — 8-GPU MI300X server with 3+3 3 kW power supplies · <a href="https://www.supermicro.com/en/products/system/datasheet/as-8125gs-tnmr2">Supermicro AS-8125GS-TNMR2</a></li></ul>
<h2>Method</h2>
<p><strong>The lines.</strong> Every line is a rent per kWh of IT capacity: rent per GPU-hour over the GPU’s power in kW. With E the kWh a megawatt delivers in a year, an old GPU keeps its megawatt while its rent per kWh beats (slot + upkeep) / E if new power is available; the newest GPU’s rent per kWh less its capital charge over E if power is walled off; and the newest GPU’s whole rent per kWh if frontier chips are already bought and idle. If scarcity ends and the newest GPU rents at its full cost, an old GPU survives while the newest’s rent per kW stays below (capital + slot + upkeep) / (slot + upkeep) times its own: 3.4×.</p>
<p><strong>Inputs.</strong> Utilization 85% of hours, so E is 7.45 million kWh per MW-year. The slot is $3.0M a year: Epoch AI’s non-server capital cost for a 1 GW site, annualized ($2.6M per MW), plus about $0.4M for on-site gas generation. Power and upkeep are $1.2M a year: Epoch’s $0.9M of operations per MW, mostly energy, rounded up for on-site fuel and upkeep. The B300’s capital charge is $63,000 per GPU over 2.0 kW, recovered over 4 years at 10%: $9.9M per MW-year, or $1.33 per kWh. The three constants are $0.56 per kWh, $1.33 per kWh and 3.4×.</p>
<p><strong>Set high on purpose.</strong> The slot and upkeep together are $4.2M a year, above Epoch’s whole non-server cost of $3.5M, and the $3.0M slot alone is above the $1.2–1.8M a year that hyperscale leases imply. A lower figure only lowers the power-available line and raises the equilibrium ratio: at Epoch’s $3.5M they would be $0.47 per kWh and 3.8×.</p>
<p><strong>Power per GPU</strong> is the vendor’s 8-GPU system maximum divided by eight, plus about 10% for networking and storage: H100 1.4 kW, H200 1.4 kW, B200 1.9 kW, B300 2.0 kW, GB200 2.0 kW, A100 0.9 kW, MI300X 1.2 kW, L40S 0.6 kW, RTX PRO 6000 0.85 kW. The GB200 takes its share of a 132 kW rack, the MI300X its share of a 9 kW server; the L40S and RTX PRO 6000 are estimates.</p>
<p><strong>Rents.</strong> On-demand and spot are the index’s composites (the 20% trimmed mean of each provider’s own price), neocloud is its neocloud-tier composite, and the one- and three-year prices are the median of providers’ term-to-on-demand ratios applied to the on-demand composite, all at the Nov 5, 2026 16:00 UTC fixing. They are offered prices, not transactions, with providers weighted equally. The B300’s one-year ratio rests on two providers and its three-year ratio on one.</p>
<p><strong>The year of weekly rents</strong> is GetDeploying’s weekly median of each provider’s median on-demand price (CC BY 4.0), with the B300’s capital charge held at today’s $1.33 throughout.</p>
<p><strong>Replacement share</strong> is (1 + g)<sup>−L</sup> in megawatts: this year’s replacements are the deployments of L years ago. Goldman’s path is uneven: 31 GW in 2025, 41 GW in 2026 and 66 GW in 2027, a 32% rise and then a 61% one; 46% is the average.</p>
<p><strong>Left out.</strong> The B300’s all-in cost is an estimate: a 20% miss moves its capital charge by $0.27 per kWh. Retrofit cost, downtime, resale value, taxes and the option to wait for Rubin are outside the lines; they move them at the margin.</p>
<hr>
<p>Every chart above links to its interactive version on 06.ai, with the data behind it. The daily trackers are at <a href="https://06.ai/trackers">06.ai/trackers</a>.</p>
<p><em>Not investment advice. Estimates are labeled as estimates.</em></p>]]></content:encoded>
</item>
<item>
  <title>Where is the unit growth left in AI hardware?</title>
  <link>https://inthesingularity.ai/packs/unit-growth</link>
  <guid isPermaLink="true">https://inthesingularity.ai/packs/unit-growth</guid>
  <pubDate>Fri, 06 Nov 2026 12:23:00 GMT</pubDate>
  <description>From 2026 to 2029, Bloom’s fuel cells (+55% a year) and 800G-and-faster optics (+32%) ship the most extra units; hard-drive exabytes, server CPUs and memory bits (+21–25%) come next. GPUs (Nvidia alone +11%) and GE Vernova’s gas turbines grow least, about 15% a year, with price doing as much of the work. Across fourteen stocks, more unit growth comes with a higher P/E, about 31% more for each extra 10 points a year, but the link is weak (R² 0.18); among the eleven that mostly sell these products it is stronger (0.53). Price already in 2027 earnings matters more: memory trades at 3–4 times them, with revenue per bit about 4.6 times 2025’s by our estimate. Since Oct 6, P/Es have fallen a median 47% while 2027 estimates rose 5%, and the unit stories fell no further than the rest.</description>
  <category>AI hardware</category>
  <enclosure url="https://inthesingularity.ai/newsletter/unit-growth/cover.png" length="59735" type="image/png" />
  <media:content url="https://inthesingularity.ai/newsletter/unit-growth/cover.png" medium="image" type="image/png" />
  <content:encoded><![CDATA[<figure><a href="https://06.ai/packs/unit-growth"><img src="https://06.ai/newsletter/unit-growth/thumb.png" alt="Where is the unit growth left in AI hardware?" width="421" height="391"></a><figcaption><a href="https://06.ai/packs/unit-growth">Interactive version and data on 06.ai →</a></figcaption></figure>
<ul><li><strong>+55%</strong> — Most unit growth left, 2026–29. A year, fuel cells. Then optics <strong>+32%</strong> and hard-drive exabytes +25%</li><li><strong>+15%</strong> — Least unit growth left. A year, GPUs (Nvidia alone <strong>+11%</strong>) and gas turbines. Both are capped by supply; price does about as much of the work as units</li><li><strong>+31%</strong> — P/E per extra 10 points of unit growth. All fourteen, R² just 0.18. Among the eleven that mostly sell these products, <strong>+57%, R² 0.53</strong></li><li><strong>−47%</strong> — P/E since Oct 6, median of 14. Prices <strong>−43%</strong>, 2027 EPS estimates +5%</li></ul>
<h2>Units times price</h2>
<p>Every hardware supplier’s revenue is units shipped times the price of each unit. In the AI buildout the two have moved very differently. Memory makers sold roughly the bits they would have anyway, at contract prices that rose about sixfold for conventional DRAM from the third quarter of 2025 to the fourth of 2026 (the <a href="https://06.ai/packs/memory-prices">memory-price</a> pack). Optics makers are shipping far more lasers and modules than a year ago. The question for the next three years is which of the two is left, and whether the market pays for one more than the other.</p>
<p>For each of seven segments we counted the physical unit that best measures volume, took 2026 to 2029 paths for units and for price per unit from company guidance and published forecasts, and set them against the P/E on 2027 consensus earnings for the fourteen stocks most tied to them. A unit is a module for optics, a bit for memory, an exabyte for hard drives, a package for GPUs, a socket for CPUs, a gigawatt for gas turbines and fuel cells. Custom AI chips, from the <a href="https://06.ai/packs/ai-memory-balance">memory</a> pack, are drawn alongside for comparison.</p>
<h2>Unit growth left, 2026–29</h2>
<p>Base case, index 2026 = 100, log scale. Units are shipments; price is revenue per unit. Revenue is the two multiplied. The dashed line is custom AI chips (TPU, Trainium and others), which aren’t one of the seven; their price isn’t known, so they appear only under Units. Ranges are in the data.</p>
<p>By 2029 fuel cells ship about <strong>3.8 times</strong> 2026’s gigawatts and optics about <strong>2.3 times</strong> its modules. GPUs and gas turbines add about half again (153 and 153), the least of the seven.</p>
<figure><a href="https://06.ai/packs/unit-growth"><img src="https://06.ai/newsletter/unit-growth/fig-01.png" alt="Unit growth left, 2026–29" width="694" height="452"></a></figure>
<figure><a href="https://06.ai/packs/unit-growth"><img src="https://06.ai/newsletter/unit-growth/fig-02.png" alt="Unit growth left, 2026–29 — Price per unit" width="694" height="483"></a><figcaption><strong>Price per unit.</strong></figcaption></figure>
<figure><a href="https://06.ai/packs/unit-growth"><img src="https://06.ai/newsletter/unit-growth/fig-03.png" alt="Unit growth left, 2026–29 — Revenue" width="694" height="483"></a><figcaption><strong>Revenue.</strong></figcaption></figure>
<h2>Segment by segment</h2>
<ul><li><strong>Fuel cells: +55% a year, price flat.</strong> Bloom doesn’t report megawatts. Its guided 2026 revenue at about $3,500–4,000 a kW is roughly 1.2 GW, and Morgan Stanley has 2.1 GW in 2027 and 3.7 GW in 2028. Bloom is on track for 2 GW a year of factory capacity by the end of 2026 and is preparing to go beyond 5 GW, adding it as orders come rather than on a fixed schedule; Oracle alone has contracted 1.2 GW. This is the purest unit story here.</li><li><strong>Optics: +32% a year in modules, more in lasers.</strong> Forecasts are far apart. Nomura has 800G and 1.6T modules going from about 80M in 2026 to 217M in 2028, about +64% a year, and laser chips from 480M to over 1.5B. BofA has 800G flat in 2027 and 1.6T up at least half, about +20% in modules. Our base, +25% in 2027 and +40% in 2028, sits between them; our high case follows Nomura. Supply covers only 50–60% of 2026’s 1.6T demand (BofA), and shortages have slowed the usual price declines (Cignal AI), so price has been doing part of the work: Lumentum’s revenue rose 136.5% from a year earlier, while its EML units are on track to grow by just over half. As Coherent more than doubles its indium-phosphide laser output again by the end of 2027 and Lumentum’s Greensboro fab starts adding revenue in early 2028, units take over and price per module drifts down, about 4% a year in our base case.</li><li><strong>Hard drives: exabytes +25% a year, drives flat.</strong> Seagate plans about 25% a year more nearline exabytes, with supply allocated through 2028, and Western Digital sees demand growing 25% or more a year. Both are growing through denser drives, not more of them; Western Digital says it won’t add unit capacity. Its price per TB is up 28% from a year ago, and Seagate’s revenue per TB about 33% on our math. We have price rising another 15% in 2027 as locked build-to-order prices carry over, then easing.</li><li><strong>Memory: bits +21% a year, the price already spent.</strong> Bit growth is not slow: Micron expects industry DRAM bits up in the low 20s and NAND in the mid 20s in 2027 and 2028. Price is what moved. TrendForce has 2026 DRAM sales up 303% and NAND 281%; on bit growth of about a fifth, revenue per bit comes to more than three times 2025’s (our estimate), and 2027 adds another step as those rises carry over and HBM reprices. From 2028 new fabs arrive and the <a href="https://06.ai/packs/memory-prices">memory-price</a> pack’s plan path has price per bit falling about 10% a year; Bernstein has SK hynix’s DRAM price falling by more than half in 2028.</li><li><strong>Server CPUs: sockets +21% a year, price +9%.</strong> Agentic AI needs more general-purpose compute per accelerator. TrendForce has shipments up 31% in 2026 and over 40% in 2027; DIGITIMES Research has general-purpose and AI server CPUs together up about 33% in 2027. Our base takes +35% in 2027, slowing after. Price isn’t the unknown it looks: list prices rose 10–20% from March as Intel and AMD sold out, and the average socket costs more still as buyers move to chips with more cores. In our base case Arm-based chips grow about 45% a year as hyperscalers move to their own designs, and Intel, short of supply, about 7%. AMD expects its server revenue to double in 2027.</li><li><strong>GPUs: packages +15% a year, price +17%.</strong> Nvidia ships about 7M packages in 2026 on our estimate (Bernstein has 6.7M GPUs) and, in our base case built on Bernstein, Morgan Stanley and Omdia, about 9.6M in 2029: +11% a year. The <a href="https://06.ai/packs/ai-memory-balance">memory</a> pack, which starts from HBM supply, has 6.3M growing about 8% a year. Nvidia calls its outlook for fiscal 2028, about 70% revenue growth, supply-constrained, so price does more of the work: revenue per package rises about 35% in 2027 as Rubin racks, about $7.8M each with $2.0M of it memory (Morgan Stanley), replace Blackwell. AMD is the exception, growing on units from a small base; it expects data center AI revenue to more than double in 2027. The accelerator units are going to custom chips, about +29% a year in the <a href="https://06.ai/packs/ai-memory-balance">memory</a> pack.</li><li><strong>Gas turbines: +15% a year, realized price +12%.</strong> GE Vernova’s deliveries follow its factory: about 17 GW in 2026 on our estimate (it shipped 5 GW in the third quarter), output of 20 GW a year from the third quarter and 24 GW in 2028. Production is mostly sold out through 2030, and it is now taking reservations for 2032. Orders priced about 50% more per kW in 2026 than in 2025, and they ship over the next four years, so realized price keeps rising even if new-order prices flatten.</li></ul>
<h2>Does the P/E follow the units?</h2>
<p>More than it follows price growth, and less than it follows price already made. At the Nov 5 close, each extra 10 points a year of unit growth came with about a <strong>31% higher P/E</strong> on 2027 earnings. But across all fourteen, unit growth explains only 18% of the spread in P/Es (R² 0.18), and with so few stocks a link that weak could be chance: the 95% range runs from −9% to +90% per 10 points. Price growth still ahead has a similar slope but explains less (R² 0.04), because it varies less from stock to stock.</p>
<p>The link is clearer where these products are the business. For Intel, Arm and GE Vernova they are under half of revenue (about 45%, 15% and 40% on our estimates), and all three trade well above the line: Intel at 29.0x on depressed earnings, Arm at 58.5x on its royalties, GE Vernova at 29.8x. Among the other eleven, unit growth explains <strong>53% of the spread</strong>, and each extra 10 points a year comes with a 57% higher P/E. Without memory as well, it still explains 56% across the eight left. Dropping stocks after looking at them flatters any fit, so take these as the top of the range; the bottom is R² 0.10, all fourteen without Arm.</p>
<p>The clearest pattern is about price that has already happened. What separates the cheapest stocks from the rest is how much price has already gone into 2027’s earnings. Memory’s revenue per bit in 2027 is about 4.6× its 2025 level on our estimate, and Micron, SanDisk and EWY trade at 3.1x–4.0x, the three furthest below the unit-growth line. Hard drives (1.4×) and Nvidia (about 2.0×) come next, at 8.2x–11.8x. Across all fourteen, price already in explains 66% of the spread, and once it is allowed for, unit growth adds little. Nearly all of that is memory, though: without the three memory names, price already in explains 8%.</p>
<h2>P/E on 2027 earnings, against what drives the growth</h2>
<p>Each dot is a stock: P/E at the close on the date chosen, on calendar-2027 consensus EPS (log scale). Unit and price growth are for the datacenter products each company sells in this piece, weighted by revenue. The dashed line is the least-squares fit of log P/E across all fourteen. On the unit-growth view, the solid line fits the eleven whose products here are at least half of revenue; Intel, Arm and GE Vernova (faded) are left out of it. Lumentum (hollow) has traded off Marvell’s offer since talks were reported on Oct 29. EWY’s P/E is a look-through of its top ten holdings.</p>
<p>At the Nov 5 close, each extra 10 points a year of unit growth came with a <strong>31% higher P/E</strong> across all fourteen, but explains only 18% of the spread (R² 0.18). Leave out Intel, Arm and GE Vernova, which mostly sell other things, and it explains <strong>53%</strong>, with 57% more P/E per 10 points.</p>
<figure><a href="https://06.ai/packs/unit-growth"><img src="https://06.ai/newsletter/unit-growth/fig-04.png" alt="P/E on 2027 earnings, against what drives the growth" width="694" height="426"></a></figure>
<figure><a href="https://06.ai/packs/unit-growth"><img src="https://06.ai/newsletter/unit-growth/fig-05.png" alt="P/E on 2027 earnings, against what drives the growth — Price growth ahead" width="694" height="457"></a><figcaption><strong>Price growth ahead.</strong></figcaption></figure>
<figure><a href="https://06.ai/packs/unit-growth"><img src="https://06.ai/newsletter/unit-growth/fig-06.png" alt="P/E on 2027 earnings, against what drives the growth — Price already in" width="694" height="457"></a><figcaption><strong>Price already in.</strong></figcaption></figure>
<figure><a href="https://06.ai/packs/unit-growth"><img src="https://06.ai/newsletter/unit-growth/fig-07.png" alt="P/E on 2027 earnings, against what drives the growth — Oct 6" width="694" height="457"></a><figcaption><strong>Oct 6.</strong></figcaption></figure>
<figure><a href="https://06.ai/packs/unit-growth"><img src="https://06.ai/newsletter/unit-growth/fig-08.png" alt="P/E on 2027 earnings, against what drives the growth — Aug 7" width="694" height="457"></a><figcaption><strong>Aug 7.</strong></figcaption></figure>
<h2>What the selloff did</h2>
<p>The Nasdaq closed at a record on Oct 6. By Nov 5 the Nasdaq-100 (QQQ) was 15% lower, after the Nasdaq’s worst day since April 2025: down 5.5%, with the Philadelphia semiconductor index down more than 12%, as investors weighed what Democrats’ sweeping midterm wins could mean for AI regulation. Some Democrats had called for pauses on new data center projects (the <a href="https://06.ai/packs/texas-data-center-risk">Texas data center</a> pack has the Texas version). Our fourteen fell further. Their median price fell 43% in a month and their median P/E <strong>47%</strong>, while the median consensus for 2027 EPS rose 5%. Over three months, P/Es are down a median 40% with estimates up 10%.</p>
<p>The cut was about as deep whatever the unit growth. The five stocks with 30% a year or more of unit growth ahead (Lumentum, Coherent, AMD, Arm and Bloom) lost a median 48% of their P/Es in a month; the other nine lost 47%. SanDisk (−55%), Seagate (−50%) and Western Digital (−48%), all price stories, fell as far as Coherent (−50%) and Bloom (−48%). Unit growth explains 10% of the differences in how far P/Es fell. Nvidia (−25%), GE Vernova (−29%) and EWY (−30%) held up best.</p>
<p>Lumentum and Marvell fell most, partly because of their deal. Marvell is down 53% since Oct 6, most of it after talks to buy Lumentum were reported on Oct 29, and Lumentum, due 4.5 Marvell shares under the Nov 2 agreement, trades 15% below what the offer is worth. Because the cut was so even, the market’s ranking barely moved: each extra 10 points a year of unit growth came with a 33% higher P/E on Aug 7, 37% on Oct 6 and 31% now (58%, 65% and 57% across the eleven). What changed is the level. Estimates kept rising while multiples fell: Micron’s 2027 EPS is +17% in three months and Western Digital’s +36%. The market is paying less for higher earnings.</p>
<h2>Every multiple fell, the unit stories no more than the rest</h2>
<figure><a href="https://06.ai/packs/unit-growth"><img src="https://06.ai/newsletter/unit-growth/fig-09.png" alt="Every multiple fell, the unit stories no more than the rest" width="694" height="440"></a><figcaption>P/E on calendar-2027 consensus EPS at the Aug 7 (hollow), Oct 6 (grey) and Nov 5 (colored) closes, log scale. Stocks ranked by unit growth, 2026–29. Right: change in P/E since Oct 6.</figcaption></figure>
<h2>Where the unit growth is left</h2>
<p><strong>The most is in fuel cells and optics.</strong> Bloom’s gigawatts grow about 55% a year with price flat, and optics’ modules about 32%, with laser chips faster still. Little price is already in their 2027 earnings: about none for Bloom and Coherent (1.0× and 0.9×), a little for Marvell (1.1×), more for Lumentum (1.6×). They trade at 31.5x and 16.0x (the optics median) after losing about half their multiples in a month.</p>
<p><strong>Hard drives, server CPUs and memory still have plenty of units,</strong> about 21–25% a year. Hard drives and memory come with earnings the market appears to treat as peak: from 2025 to 2027 price per TB rises about 1.4× and revenue per bit about 4.6× on our estimates, and they trade at 8.5x and 3.5x. The average CPU socket’s price rises about as much as hard drives’ (1.4×, partly more cores per socket), yet the CPU stocks trade at a median 29.0x, on Intel’s depressed earnings and Arm’s royalties rather than on sockets.</p>
<p><strong>The least is in GPUs and gas turbines,</strong> about 15% a year each, both held back by supply. Nvidia calls its outlook for fiscal 2028 supply-constrained, and GE Vernova’s deliveries follow its factory. Price does about as much of the work as units. The market pays 11.8x 2027 earnings for Nvidia and 29.8x for GE Vernova, where gas turbines are about 40% of revenue on our estimate and the price rises are already booked in a backlog mostly sold out through 2030; Nvidia’s depend on the next rack. Among accelerators, the units have moved to custom chips: about 29% a year in the <a href="https://06.ai/packs/ai-memory-balance">memory</a> pack, against about 8% for Nvidia on the same basis.</p>
<h2>Units left, price already in, and the multiple</h2>
<p>Ranked by unit growth, 2026–29, base case with the range. P/E is the median of the segment’s stocks on calendar-2027 EPS. Optics’ price already in is the industry’s modules; its three stocks run from Coherent’s 0.9× to Lumentum’s 1.6×.</p>
<ul><li><strong>Fuel cells</strong> — Counted as: GW shipped · Units a year: <strong>55.4%</strong> · Range: 36–71% · Price a year: 0.0% · Revenue a year: 55.4% · Price 2027 vs 2025: 1.0× · P/E Nov 5: 31.5x · Oct 6: 60.3x · Stocks: BE</li><li><strong>Optics</strong> — Counted as: 800G-and-faster modules · Units a year: <strong>32.0%</strong> · Range: 15–50% · Price a year: −4.2% · Revenue a year: 26.5% · Price 2027 vs 2025: 1.0× · P/E Nov 5: 16.0x · Oct 6: 39.5x · Stocks: LITE, COHR, MRVL</li><li><strong>Custom AI chips</strong> — Counted as: packages · Units a year: <strong>28.9%</strong> · Range: 17–37% · Stocks: Not in the fourteen</li><li><strong>Hard drives</strong> — Counted as: exabytes · Units a year: <strong>24.9%</strong> · Range: 16–30% · Price a year: 4.2% · Revenue a year: 30.1% · Price 2027 vs 2025: 1.4× · P/E Nov 5: 8.5x · Oct 6: 16.8x · Stocks: WDC, STX</li><li><strong>Memory</strong> — Counted as: DRAM and NAND bits · Units a year: <strong>21.2%</strong> · Range: 15–26% · Price a year: 5.1% · Revenue a year: 27.4% · Price 2027 vs 2025: 4.6× · P/E Nov 5: 3.5x · Oct 6: 5.7x · Stocks: MU, SNDK, EWY</li><li><strong>Server CPUs</strong> — Counted as: sockets · Units a year: <strong>21.2%</strong> · Range: 8–33% · Price a year: 9.1% · Revenue a year: 32.3% · Price 2027 vs 2025: 1.4× · P/E Nov 5: 29.0x · Oct 6: 54.6x · Stocks: INTC, ARM, AMD</li><li><strong>GPUs</strong> — Counted as: Nvidia and AMD packages · Units a year: <strong>15.3%</strong> · Range: 7–24% · Price a year: 17.0% · Revenue a year: 35.0% · Price 2027 vs 2025: 2.0× · P/E Nov 5: 18.3x · Oct 6: 28.5x · Stocks: NVDA, AMD</li><li><strong>Gas turbines</strong> — Counted as: GW delivered · Units a year: <strong>15.2%</strong> · Range: 9–21% · Price a year: 12.0% · Revenue a year: 29.0% · Price 2027 vs 2025: 1.3× · P/E Nov 5: 29.8x · Oct 6: 41.9x · Stocks: GEV</li></ul>
<h2>What would change the ranking</h2>
<ul><li><strong>Which optics forecast is right.</strong> LightCounting expects slower growth after three years of 65–93% growth in Ethernet transceiver sales, with more conservative projections for 2027–31, and BofA has 800G flat in 2027. Our low case is 15% a year. On Nomura’s numbers, our high case, optics grows about 50% a year, close to fuel cells. Linear-drive and co-packaged optics also strip DSPs out of a growing share of ports, which cuts Marvell’s units even as lasers grow.</li><li><strong>Supply catching up.</strong> TSMC’s CoWoS shortfall was expected to narrow from about 20% of demand to about 10% by the end of 2026 (Economic Daily News via TrendForce). If Nvidia’s packages grow with supply rather than with demand for its pricing, GPU units move up the table and price down it.</li><li><strong>Bloom’s base year.</strong> Bloom doesn’t disclose megawatts. If 2026 is 1.0 GW rather than 1.2, the growth rate is about 65%; at 1.4 GW, about 48%.</li><li><strong>Hard drives and solid state.</strong> With drive makers’ gross margins near 60%, against about 30% historically, cloud buyers have reason to qualify high-capacity flash. TrendForce expects NAND supply to pass demand in the second half of 2027, though Micron says supply stays much tighter in 2027 and 2028 than in 2026. A looser NAND market would make switching cheaper.</li><li><strong>Memory’s floors.</strong> Micron’s 26 long-term agreements cover about 35% of its DRAM bits through 2030, mostly priced within floor and ceiling bands, and SK hynix’s cover more than half of its planned 2027 and 2028 output. If they hold, memory’s price behaves more like contracted volume, and the market may stop treating its earnings as peak.</li></ul>
<h2>What to watch</h2>
<p><strong>Nvidia, Nov 17.</strong> Whether supply still caps fiscal 2028 at about 70% revenue growth, and what 2027 units the guide implies. A bigger unit step would undercut the “price, not units” read.</p>
<p><strong>Marvell, Dec 1.</strong> Electro-optics growth against Lumentum’s, and anything on the deal’s shareholder votes and regulatory reviews. The discount on Lumentum prices the odds of closing, the time it takes and what Lumentum would be worth if the deal broke.</p>
<p><strong>Micron, late December.</strong> 2027 bit growth against price; whether customer inventory and falling NAND wafer spot prices, down 9% in the week of Oct 26 alone, start to show in contract prices.</p>
<p><strong>Seagate and Western Digital, late January.</strong> How 2028 allocations price. A third year of price-per-TB gains would make hard drives less of a units story still.</p>
<p><strong>GE Vernova and Bloom, January and February.</strong> Gigawatts under contract, now slowing at GE Vernova, and Bloom’s first 2027 guide, which would put a company number on its megawatts.</p>
<h2>What this is anchored on</h2>
<p>Figures are as reported. The unit and price paths, each company’s mix, the price-already-in estimates and the fits are ours, built from these.</p>
<ul><li><strong>Prices and EPS</strong> — Daily closes; consensus EPS for the current and next fiscal year and its trend 30 and 90 days back; EWY’s top ten holdings · <a href="https://finance.yahoo.com/quote/NVDA/analysis/">Yahoo Finance, fetched Nov 6 (data/unit-growth/raw)</a></li><li><strong>Selloff</strong> — Nov 5: Nasdaq −5.52%, its worst day since April 2025 and a correction from its Oct 6 record; S&amp;P 500 −4.36%; Philadelphia semiconductor index down more than 12%; Nvidia −9.26%; investors weighed what Democrats’ midterm wins could mean for AI regulation, with some Democrats calling for pauses on new data center projects · <a href="https://www.foxbusiness.com/markets/stocks-nov-5-2026">Fox Business, Nov 5</a></li><li><strong>Selloff</strong> — AMD and Micron down more than 17%, SanDisk more than 22%, Intel about 10% on Nov 5 · <a href="https://www.kucoin.com/news/flash/chip-and-ai-stocks-plunge-as-nasdaq-falls-5-5-in-worst-post-election-session-since-1896">KuCoin News, Nov 5</a></li><li><strong>Optics</strong> — 800G transceivers 55.0M in 2026 and 112.0M in 2028; 1.6T 25.1M and 104.5M; datacenter laser chips 480M (2026) to over 1.5B (2028); InP substrates 856K to 2.49M · <a href="https://longbridge.com/en/profiles/2071818324550963200">Nomura via Longbridge, Nov 1 and Nov 5</a></li><li><strong>Optics</strong> — 2026 demand: 800G 50–60M, 1.6T 30M; supply covers 70–80% and 50–60%. 2027: 800G about flat, 1.6T up at least 50% · <a href="https://longportapp.cn/topics/100000000796055">BofA via LongPort, Jul 14</a></li><li><strong>Optics</strong> — Ethernet transceiver sales +93% in 2024, +82% in 2025 and +65% in 2026, with more conservative projections for 2027–31; InP laser demand exceeds supply by about 30% · <a href="https://lightcounting.com/newsletter/en/april-2026-market-forecast-379">LightCounting, Apr 2026</a></li><li><strong>Optics</strong> — Datacom optical component sales $7.7B in 1Q26, more than double a year earlier; shortages slowing the usual price declines · <a href="https://cignal.ai/2026/06/datacom-optical-component-revenue-doubles-to-7-7-billion-in-1q26/">Cignal AI, Jun 18</a></li><li><strong>Optics</strong> — Lumentum FQ1 FY27: revenue $1.263B (+136.5%), non-GAAP gross margin 53.2%; Q2 guide $1.575–1.625B · <a href="https://investor.lumentum.com/financial-news-releases/news-details/2026/Lumentum-Announces-First-Quarter-of-Fiscal-Year-2027-Financial-Results/default.aspx">Lumentum, Nov 2</a></li><li><strong>Optics</strong> — Lumentum: EML units up more than 50% by the December 2026 quarter; Greensboro fab’s first revenue in early 2028 · <a href="https://www.moomoo.com/community/feed/lumentum-q4-fy2026-earnings-conference-call-117078865674245">Lumentum FQ4 FY26 call via Moomoo (AI transcript), Aug 11</a></li><li><strong>Optics</strong> — Marvell to buy Lumentum for 4.500 Marvell shares each, about $115B · <a href="https://lightreading.com/ai-machine-learning/marvell-strikes-115b-deal-for-lumentum">Light Reading, Nov 2</a></li><li><strong>Optics</strong> — Marvell Q2 FY27: revenue $2.74B, data center $2.17B; electro-optics about $1.4B (+97%) and custom about $425M (estimates) · <a href="https://www.nextplatform.com/connect/2026/09/02/optics-still-driving-marvells-ai-business-more-than-custom-chips/5294018">The Next Platform, Sep 2</a></li><li><strong>Optics</strong> — Coherent: Datacenter &amp; Communications 79% of revenue; InP laser output +80% y/y, doubling by September 2026 and more than doubling again by end-2027 · <a href="https://www.tradingkey.com/news/transcripts/262105524-tradingkey">Coherent Q4 FY26 call, Aug 12</a></li><li><strong>Memory</strong> — Units and prices: the <a href="https://06.ai/packs/memory-prices">memory-price</a> pack’s plan path (DRAM bits about +20% a year, NAND about +23%; revenue $889B in 2026 to $1.84T in 2029) · <a href="https://06.ai/packs/memory-prices">06.ai </a><a href="https://06.ai/packs/memory-prices">memory-price</a><a href="https://06.ai/packs/memory-prices"> pack</a></li><li><strong>Memory</strong> — 2026 DRAM market value $618.7B (+303%), NAND $270.6B (+280.7%) · <a href="https://www.itiger.com/news/1173429702">TrendForce via Tiger Brokers, May 29</a></li><li><strong>Memory</strong> — Industry DRAM bit shipments low 20s % in 2027 and 2028, NAND mid 20s %; supply and demand much tighter in 2027–28 than 2026 · <a href="https://www.trendforce.com/news/2026-10-01/news-micron-fy2026-revenue-surges-256-to-133-2b-locks-in-35-of-revenue-through-2030/">Micron FQ4 FY26 call via TrendForce, Oct 1</a></li><li><strong>Memory</strong> — Quarterly contract prices: 4Q26 DRAM +10–15% (with HBM +15–20%), NAND +15–20% · <a href="https://www.trendforce.com/presscenter/news/20260929-13255.html">TrendForce, Sep 29</a></li><li><strong>Memory</strong> — Micron: 26 strategic customer agreements through 2030 cover about 35% of DRAM bits and a little more of NAND; most priced within floor and ceiling bands · <a href="https://www.webull.com/news/15672802111996928">Micron FQ4 FY26 call via Webull, Sep 30</a></li><li><strong>Memory</strong> — SK hynix: long-term agreements with 14 customers cover more than 50% of planned 2027 and 2028 capacity · <a href="https://www.trendforce.com/news/?p=65818">TrendForce, Oct 28</a></li><li><strong>Memory</strong> — NAND supply grows faster than demand in 2027; the balance turns positive in the second half · <a href="https://www.trendforce.com/presscenter/news/20260721-13148.html">TrendForce, Jul 21</a></li><li><strong>Memory</strong> — 512Gb TLC NAND wafer spot price $16.88 on Oct 26, −9.2% on the week · <a href="https://www.trendforce.com/news/?p=65857">TrendForce, Oct 28</a></li><li><strong>Memory</strong> — SK hynix DRAM about $2.23/GB at the end of 2027, almost five times 2025; about $1.05 in 2028, more than a 50% drop · <a href="https://pcgamer.com/hardware/memory/analysts-think-sk-hynixs-dram-and-nand-prices-will-finally-drop-by-the-end-of-2028-but-not-before-a-huge-rise-over-the-next-year/">Bernstein via PC Gamer, Jul 7</a></li><li><strong>Memory</strong> — KOSPI 12-month forward P/E 4.41x on Oct 20; semiconductors below 3x, the rest 10.3x · <a href="https://www.mt.co.kr/en/stock/2026/10/20/2026102015013299958">Money Today (FnGuide), Oct 20</a></li><li><strong>Hard drives</strong> — Seagate FQ1 FY27: 220 EB (+21%), data center 198 EB; nearline exabyte supply about +25% a year from density, not more drives; build-to-order terms locked through 2027, supply allocated through 2028 · <a href="https://www.webull.com/news/15824229571375104">Seagate call via Webull, Oct 27</a></li><li><strong>Hard drives</strong> — Western Digital FQ1 FY27: 248 EB (+22%); revenue per exabyte +28% · <a href="https://www.blocksandfiles.com/disk/2026/10/23/wd-knocks-earnings-out-of-the-park-again/5304988">Blocks &amp; Files, Oct 23</a></li><li><strong>Hard drives</strong> — Western Digital: price per TB +28% y/y; exabyte demand “25% plus” a year; no added unit capacity, growth from areal density · <a href="https://bullfincher.io/companies/western-digital-corporation/earnings-call/fy2027-q1">Western Digital FQ1 FY27 call via Bullfincher</a></li><li><strong>Hard drives</strong> — Long-term model: price per TB stable to slightly up, against historical declines of mid-to-high single digits a year · <a href="https://fintool.com/news/wd-innovation-day-100tb-hamr-20-eps">WD Innovation Day via Fintool, Feb 3</a></li><li><strong>GPUs</strong> — Nvidia about 6.7M GPUs in 2026; 61K racks in 2026 and 88K in 2027; high-end AI servers +13–17% in 2027 and high single digits in 2028 · <a href="https://hao.cnyes.com/post/262306">Bernstein via Anue, Aug 5</a></li><li><strong>GPUs</strong> — Rubin and Rubin Ultra about 7M units in 2027; CoWoS demand about 2.69M wafers in 2027 (+93%) · <a href="https://www.kucoin.com/news/flash/morgan-stanley-report-cowos-demand-to-double-by-2027-amd-and-asics-drive-ai-chip-growth">Morgan Stanley via KuCoin, Jul 8</a></li><li><strong>GPUs</strong> — 2027: 23.3M accelerators, 10.9M GPUs and 12.5M ASICs · <a href="https://www.kucoin.com/news/flash/jpmorgan-ai-asic-shipments-to-surpass-gpus-by-2027">JPMorgan via KuCoin, Jun 22</a></li><li><strong>GPUs</strong> — About 10.1M GPUs and 10.5M ASICs in 2028 · <a href="https://omdia.tech.informa.com/blogs/2026/aug/2028-the-year-of-the-custom-asic-transition">Omdia, Aug 21</a></li><li><strong>GPUs</strong> — VR200 NVL72 rack about $7.8M, memory about $2.0M of it · <a href="https://pcgamer.com/hardware/a-single-nvidia-vera-rubin-rack-is-estimated-to-cost-usd7-803-148-with-over-usd2-million-of-that-figure-spent-on-memory-alone">Morgan Stanley via PC Gamer, May 22</a></li><li><strong>GPUs</strong> — Nvidia Q2 FY27: data center $89.0B; Q3 revenue guide $108B ±2% · <a href="https://www.storagenewsletter.com/2026/08/27/nvidia-fiscal-2q27-financial-results/">Nvidia via StorageNewsletter, Aug 27</a></li><li><strong>GPUs</strong> — Fiscal 2028 revenue up about 70%, “a supply-constrained outlook”; supply a bottleneck at least through fiscal 2028 · <a href="https://www.webull.com/news/15497781742806016">Nvidia Q2 FY27 call via Webull, Aug 26</a></li><li><strong>GPUs</strong> — TSMC CoWoS capacity 120–140K wafers a month in 2026; the shortfall seen narrowing from about 20% to about 10% by the end of 2026 · <a href="https://www.trendforce.com/news/2026/06/15/news-tsmc-cowos-supply-demand-gap-reportedly-seen-narrowing-from-20-to-10-by-end-2026-as-capacity-expands">Economic Daily News via TrendForce, Jun 15</a></li><li><strong>GPUs</strong> — AMD Q3 2026: data center $8.55B (+97%) · <a href="https://ir.amd.com/news-events/press-releases/detail/1303/amd-reports-third-quarter-2026-financial-results">AMD, Nov 3</a></li><li><strong>GPUs</strong> — MI455X shipping in volume; 2027 revenue up more than 80%, data center AI more than doubles, server revenue doubles; demand higher than supply · <a href="https://pro.stockalarm.io/quote/AMD/earnings-calls">AMD Q3 2026 call via StockAlarm, Nov 3</a></li><li><strong>Custom chips</strong> — TPU, Trainium and other custom accelerators: 7.0M packages in 2026 to 15.0M in 2029. On the same basis Nvidia 6.3M to 8.0M, AMD 0.55M to 1.6M · <a href="https://06.ai/packs/ai-memory-balance">06.ai </a><a href="https://06.ai/packs/ai-memory-balance">memory</a><a href="https://06.ai/packs/ai-memory-balance"> pack</a></li><li><strong>Server CPUs</strong> — Server CPUs 28.43M in 2025 and 48.57M in 2027. General-purpose 28.70M in 2026 and 35.28M in 2027; AI server CPUs 5.35M and 9.85M. Arm-based 15.8M in 2027 (32.5%) · <a href="https://udn.com/news/story/7240/9658461">DIGITIMES Research via UDN, Jul 29</a></li><li><strong>Server CPUs</strong> — Server CPU shipments +31% in 2026 and over 40% in 2027; Arm about 30% of the server CPU market in 2027 · <a href="https://www.trendforce.com/presscenter/news/20261022-13279.html">TrendForce, Oct 22</a></li><li><strong>Server CPUs</strong> — Server CPU market $61.4B in 2026 to $210.6B in 2030 · <a href="https://www.webull.com/news/15396229680931840">BofA via Webull, Aug 13</a></li><li><strong>Server CPUs</strong> — Server CPU prices up 10–20% since March; Intel raised prices twice in 2026 · <a href="https://www.techspot.com/news/112144-cpu-prices-jump-20-since-march-intel-amd.html">TechSpot (citing Nikkei)</a></li><li><strong>Server CPUs</strong> — Intel Q3 2026: DCAI $7.32B (+78%); supply short of demand through 2027 and into 2028 · <a href="https://www.intc.com/news-events/press-releases/detail/1786/intel-reports-third-quarter-2026-financial-results">Intel, Oct 2026</a></li><li><strong>Server CPUs</strong> — Arm Q2 FY27: royalties $670M (+8%); data center royalties doubled · <a href="https://www.sec.gov/Archives/edgar/data/0001973239/000197323926000141/exhibit992fye27q230-sepx26.htm">Arm 6-K, Nov 2026</a></li><li><strong>Gas turbines</strong> — Output 20 GW a year from Q3 2026, 24 GW in 2028, 30 GW in 2030; at least 30 GW to ship over six quarters; production mostly sold out through 2030 · <a href="https://www.gevernova.com/sites/default/files/gev_webcast_transcript_07222026.pdf">GE Vernova Q2 call transcript, Jul 22</a></li><li><strong>Gas turbines</strong> — Q3 2026: 5 GW shipped; 11 GW of new contracts; backlog plus slot reservations 122 GW · <a href="https://www.sec.gov/Archives/edgar/data/0001996810/000199681026000162/gev3q2026form8-k10282026.pdf">GE Vernova 8-K, Oct 28</a></li><li><strong>Gas turbines</strong> — Now taking reservations for 2032 delivery; orders and slot reservations slowing; Q3 pricing more than 20% above Q4 2025 · <a href="https://www.utilitydive.com/news/ge-vernova-in-9b-deal-to-buy-high-voltage-manufacturer-as-gas-turbine-orde/873223/">Utility Dive, Oct 29</a></li><li><strong>Gas turbines</strong> — 2026 orders priced about 50% more per kW than 2025’s · <a href="https://247wallst.com/investing/2026/10/29/ai-data-centers-have-driven-gas-turbine-prices-up-50/">24/7 Wall St, Oct 29</a></li><li><strong>Gas turbines</strong> — Turbine prices +195% since 2019, toward about $600/kW by end-2027 · <a href="https://www.woodmac.com/press-releases/gas-turbine-prices-soar-195-as-market-faces-supply-demand-crisis">Wood Mackenzie, Apr 1</a></li><li><strong>Gas turbines</strong> — Global gas turbine manufacturing capacity 67 GW a year in 2025 to 102 GW in 2030 · <a href="https://oenergetice.cz/en/technology/gas-turbine-queue-lengthens-as-manufacturing-capacity-set-to-rise-significantly-by-2030">BNEF via oEnergetice, Sep 18</a></li><li><strong>Fuel cells</strong> — Bloom Q3 2026: revenue $1.236B (+138%); 2026 guide raised to $4.4–4.6B · <a href="https://bloomenergy.com/news/bloom-energy-reports-record-third-quarter-2026-financial-results-and-raises-full-year-2026-guidance/">Bloom Energy, Oct 22</a></li><li><strong>Fuel cells</strong> — On track to double annual production capacity to 2 GW by the end of 2026 · <a href="https://www.utilitydive.com/news/bloom-energy-says-its-on-track-for-2-gw-annual-production-capacity/804291/">Utility Dive, Oct 30, 2025</a></li><li><strong>Fuel cells</strong> — Preparing capacity beyond 5 GW at Fremont, with no date; expansion to follow customer demand · <a href="https://www.zacks.com/stock/news/3017009/bloom-energy-q3-earnings-call-highlights-guidance-hike-&amp;-capacity-plan">Zacks on the Q3 call, Oct 28</a></li><li><strong>Fuel cells</strong> — Shipments 2.1 GW in 2027 and 3.7 GW in 2028 (raised from 1.8 and 3.1) · <a href="https://stocktwits.com/news-articles/markets/equity/be-stock-extends-selloff-after-q3-print-but-morgan-stanley-sees-a-production-ramp-ahead/cZxyKioRBxn">Morgan Stanley via Stocktwits, Oct 23</a></li><li><strong>Fuel cells</strong> — Oracle: up to 2.8 GW, 1.2 GW contracted and deploying into 2027 · <a href="https://www.businesswire.com/news/home/20260414697111/en">Bloom Energy via Business Wire, Apr 14</a></li></ul>
<h2>Method</h2>
<p><strong>P/E.</strong> Price at the close on Nov 5, Oct 6 (the Nasdaq’s record) and Aug 7, over consensus EPS for calendar 2027. Calendar 2027 blends the two fiscal years Yahoo Finance carries by the months each puts in 2027: half each for June years (Lumentum, Coherent, SanDisk, Western Digital, Seagate), 8 and 4 months for Micron, 3 and 9 for Arm, 1 and 11 for Nvidia and Marvell, and the second year alone for December years. Estimates for the earlier dates are Yahoo’s 30- and 90-day-old consensus for the same years. EWY is a look-through: the harmonic mean of its top ten holdings’ P/Es at their fund weights, with the rest of the fund at the non-memory names’ average. That gives 4.0x now, with its memory holdings (50% of the fund) at 2.5x and the rest at 10.3x, in line with FnGuide’s 10.3x for the KOSPI outside chips. The look-through uses Seoul’s Nov 5 closes, which came before that day’s New York selloff: EWY fell 13% in New York on Nov 5, against about 7% for its top ten in Seoul, so on the same timing it would be nearer 3.7x.</p>
<p><strong>Units and price.</strong> Each segment has a unit index and a price-per-unit index for 2026–29 in base, low and high cases, built from the forecasts and guidance in the sources. Memory is the <a href="https://06.ai/packs/memory-prices">memory-price</a> pack’s plan path, with DRAM including HBM; GPUs add Nvidia’s and AMD’s packages, with Nvidia about $43K of data center compute revenue per package in 2026 and AMD about $30K; custom chips are the <a href="https://06.ai/packs/ai-memory-balance">memory</a> pack’s base. Segments combine products at their 2026 revenue, and price is revenue over the unit index. For GPUs the unit is a package, so price includes the move to pricier parts; for memory, bits are weighted by 2026 revenue, so the DRAM–NAND mix sits in units while HBM sits in DRAM’s price per bit. For CPUs, price per socket includes more cores per socket.</p>
<p><strong>Each stock.</strong> Unit and price growth are its datacenter products in this piece weighted by their share of its 2027 revenue among them (see the data under the P/E chart). Products outside the seven (Nvidia’s networking, Intel’s PCs, most of Arm’s royalties, GE Vernova’s grid and wind businesses) are left out, and the share column shows how much that leaves.</p>
<p><strong>Price already in</strong> is price per unit in 2027 against 2025, our estimate: memory from TrendForce’s 2026 revenue and bit growth and the <a href="https://06.ai/packs/memory-prices">memory-price</a> pack’s 2027; hard drives from Western Digital’s revenue per exabyte; GPUs from Nvidia’s and AMD’s revenue per package; CPUs from BofA’s and DIGITIMES’s market size and units. For Lumentum, Coherent, Marvell, Arm and Bloom it is a rough judgment from the companies’ commentary.</p>
<p><strong>Fits.</strong> Least squares of log P/E on each measure, one at a time, across the fourteen, with the slope’s two-sided p-value. With so few stocks, one or two can move a fit a lot; the data under the chart repeats it without Intel, without Arm, without memory, and for the eleven whose products here are at least half of revenue, a cut we made after seeing the data. Lumentum is kept in, though its price has tracked Marvell’s offer since talks were reported on Oct 29; without it the fits barely change (unit growth R² 0.18).</p>
<hr>
<p>Every chart above links to its interactive version on 06.ai, with the data behind it. The daily trackers are at <a href="https://06.ai/trackers">06.ai/trackers</a>.</p>
<p><em>Not investment advice. Estimates are labeled as estimates.</em></p>]]></content:encoded>
</item>
<item>
  <title>What happens to memory prices if AI capex stops growing in 2028?</title>
  <link>https://inthesingularity.ai/packs/memory-prices</link>
  <guid isPermaLink="true">https://inthesingularity.ai/packs/memory-prices</guid>
  <pubDate>Wed, 04 Nov 2026 12:22:00 GMT</pubDate>
  <description>Memory makers have sold 2027 at about six times mid-2025’s prices, and the fabs they ordered add about a fifth more DRAM bits in 2028. If AI capex in 2028 only matches 2027, with Texas’s slower hookups trimming both years, the memory bought for AI in 2028 falls about a quarter short of plan. In 2019, the last time cloud capex stopped growing, DRAM’s price per bit fell by around half. Our model has DRAM contract prices averaging about half as much in 2028 as in 2027, and NAND more than 40% less.</description>
  <category>AI memory</category>
  <enclosure url="https://inthesingularity.ai/newsletter/memory-prices/cover.png" length="68799" type="image/png" />
  <media:content url="https://inthesingularity.ai/newsletter/memory-prices/cover.png" medium="image" type="image/png" />
  <content:encoded><![CDATA[<figure><a href="https://06.ai/packs/memory-prices"><img src="https://06.ai/newsletter/memory-prices/thumb.png" alt="What happens to memory prices if AI capex stops growing in 2028?" width="421" height="415"></a><figcaption><a href="https://06.ai/packs/memory-prices">Interactive version and data on 06.ai →</a></figcaption></figure>
<p>Memory makers have sold 2027 at prices about six times mid-2025’s, and the fabs they ordered add about a fifth more DRAM bits in 2028. If AI capex in 2028 only matches 2027, and Texas’s slower hookups trim both years, the memory bought for AI in 2028 falls about a quarter short of plan, partly because 2027’s contracted HBM has to be used up first. In 2019, the last time cloud capex stopped growing, DRAM’s price per bit fell by around half. Our model has <strong>DRAM contract prices averaging about 50% less in 2028 than in 2027, NAND about 43% less and 2028 HBM contracts about 35% lower.</strong> Half off a sixfold rise still leaves DRAM about three times its mid-2025 price.</p>
<ul><li><strong>−50%</strong> — DRAM contract price, 2028 vs 2027. Flat 2028 + blue wave, annual averages. On plan, <strong>−9%</strong>. Our tests range −35% to −62%</li><li><strong>−35%</strong> — HBM, 2028 contracts vs 2027. $30.8 to $20.1 per GB, if half the commodity move reaches deals struck in 2H27</li><li><strong>−43%</strong> — NAND contract price, 2028 vs 2027. On plan, <strong>−14%</strong>. NAND wafer spot fell 14% in the four weeks to November 2</li><li><strong>$0.95T</strong> — DRAM and NAND revenue, 2028. Against <strong>$1.48T</strong> in 2027 and $1.70T on plan</li></ul>
<h2>Where prices stand</h2>
<p>Conventional DRAM contract prices rose about sixfold from the third quarter of 2025 to the fourth quarter of 2026, chaining TrendForce’s quarterly estimates. NAND rose about 3.6 times over 2026. DDR4 8Gb, the old PC part, went from $1.35 in early 2025 to $27.50 in October, twenty times. The makers are earning margins memory has never seen: SK hynix an 80% operating margin in the third quarter, Micron an 87% gross margin, SanDisk 85%.</p>
<p>2027 is sold. SK hynix says its 2027 volumes and prices are final and its DRAM and HBM are effectively sold out through 2029, with long-term agreements covering more than half of its 2027–28 capacity. Samsung says its 2027 HBM is sold out, and Micron has agreed the vast majority of its 2027 HBM. The open question is 2028, and it hangs on AI capex: counting HBM at the wafer it takes, we estimate the HBM and CPU memory in AI accelerator servers is about 30% of DRAM demand in 2026 and 37% in 2028 (<a href="https://06.ai/packs/ai-memory-balance">the </a><a href="https://06.ai/packs/ai-memory-balance">memory</a><a href="https://06.ai/packs/ai-memory-balance"> pack</a>, which also counted CPU servers bought for agents, put 2027 near 43%). Counterpoint has servers of every kind at 48% of 2026 DRAM shipments, and HBM at another 9%.</p>
<p>The rises are slowing, and spot has stopped pulling contract up. DRAM spot has barely moved since early October and sits about 70% above October’s DDR4 contract price. NAND wafer spot fell 14% in the four weeks to November 2 as Chinese phone brands cut prices, and October’s MLC contract price was flat, ending a 21-month rise. Buyers are cutting back: IDC expects smartphone shipments to fall 16.7% this year and PCs 11.3%, and SK hynix says some customers are reviewing the timing of their projects.</p>
<h2>Contract price rises are slowing</h2>
<figure><a href="https://06.ai/packs/memory-prices"><img src="https://06.ai/newsletter/memory-prices/fig-01.png" alt="Contract price rises are slowing" width="694" height="305"></a><figcaption>Change in contract prices from the previous quarter, TrendForce, midpoint of each range. 4Q25 and 1Q26 DRAM are actuals; the rest are TrendForce’s forecasts for the quarter. TrendForce’s NAND figure for 4Q25 isn’t shown.</figcaption></figure>
<h2>The capex paths</h2>
<p><strong>Plan.</strong> <a href="https://06.ai/packs/ai-compute-balance">The </a><a href="https://06.ai/packs/ai-compute-balance">compute-balance</a><a href="https://06.ai/packs/ai-compute-balance"> pack</a> has worldwide AI capex rising from about $1.0T in 2026 to $1.40T in 2027 and $1.65T in 2028, up 18%. The Street has a little less growth in 2028: Morgan Stanley has hyperscalers and AI cloud providers up about 14%, to $1.8T, and Goldman has the large US hyperscalers at about $1.4T, some $100B above consensus. S&amp;P Global’s models generally assume capex flattens in 2028, and Counterpoint’s memory analyst Kang Kyung-soo has Big Tech’s capex growth approaching zero.</p>
<p><strong>Flat 2028.</strong> 2028 spends what 2027 planned to, $1.40T, and growth resumes at the plan’s rates in 2029.</p>
<p><strong>Blue wave.</strong> <a href="https://06.ai/packs/texas-data-center-risk">The </a><a href="https://06.ai/packs/texas-data-center-risk">Texas data center</a><a href="https://06.ai/packs/texas-data-center-risk"> pack</a> has Texas switching on 2.4 GW less data center load by the end of 2027 than on the path Gov. Greg Abbott had set, and 3.8 GW less by the end of 2028, with Gina Hinojosa as governor, Vikki Goodwin as lieutenant governor and a moratorium until new laws take effect in September 2027. At <a href="https://06.ai/packs/blue-wave-fcf">the </a><a href="https://06.ai/packs/blue-wave-fcf">blue-wave cash flow</a><a href="https://06.ai/packs/blue-wave-fcf"> pack</a>’s $50B a gigawatt, three-quarters of it leaving the year the switch-on moves, that takes $89B of AI capex out of 2027 and another $53B out of 2028, about 6% and 3% of plan. It counts every builder in Texas, not just the Big Four.</p>
<p><strong>Flat 2028 with the blue wave</strong>, our headline: $1.31T in 2027 and $1.35T in 2028, 18% below plan.</p>
<h2>Four paths for AI capex</h2>
<figure><a href="https://06.ai/packs/memory-prices"><img src="https://06.ai/newsletter/memory-prices/fig-02.png" alt="Four paths for AI capex" width="694" height="339"></a><figcaption>Worldwide AI capex, $T. Plan: the <a href="https://06.ai/packs/ai-compute-balance">compute-balance</a> pack’s base. Flat 2028: 2028 matches 2027’s plan. Blue wave: less the Texas capex the <a href="https://06.ai/packs/texas-data-center-risk">Texas data center</a> and <a href="https://06.ai/packs/blue-wave-fcf">blue-wave cash flow</a> packs move out of each year.</figcaption></figure>
<h2>Why 2028 takes the hit</h2>
<p><strong>Memory is bought before the chips ship.</strong> The <a href="https://06.ai/packs/ai-memory-balance">memory</a> pack’s model has about 35% of a year’s HBM bought the year before, as stacks move through CoWoS packaging, boards and buffers. 2027’s HBM is contracted and largely take-or-pay, so in the headline case it is bought on plan anyway, including the stacks meant for 2028’s chips; whatever 2027 buys beyond what the flat path needs, about 0.8 EB, is used up before 2028 orders anything new. Server orders placed late in 2027 for 2028’s builds go the same way. So 2028’s purchases fall harder than 2028’s capex: HBM bought in 2028 is 27% below plan, host DRAM in AI servers 23% and AI storage 23%, against 18% less capex.</p>
<p><strong>Buyers hold stock they bought in the shortage.</strong> Korea Invest Insights estimates customers built 2–9 weeks of DRAM stock in 2026 and that restocking makes up 8–13% of 2027 DRAM shipments. When prices turn, that stock gets used instead of bought. We assume 3 weeks of demand comes out of 2028 orders.</p>
<p><strong>Supply is already ordered.</strong> Micron expects industry DRAM bits to grow in the low 20s % in both 2027 and 2028, and NAND in the mid-20s %. SK hynix’s Yongin fab 1 opens its first cleanroom in February 2027; Samsung’s P5 takes tools from mid-2027 for mass production late in 2028; Micron’s Idaho fab ships from mid-2027, with a second Idaho fab and Japan late in 2028; CXMT goes from about 350,000 wafer starts a month at the end of 2026 to 500,000 by the end of 2028.</p>
<h2>2028 DRAM demand falls 13.7% short of plan, and price has to win back 9.6 points</h2>
<figure><a href="https://06.ai/packs/memory-prices"><img src="https://06.ai/newsletter/memory-prices/fig-03.png" alt="2028 DRAM demand falls 13.7% short of plan, and price has to win back 9.6 points" width="694" height="334"></a><figcaption>Headline case, 2028, DRAM demand as % of plan, with HBM counted at its wafer cost (about 2.9 conventional bits per HBM bit). Makers hold back 30% of the shortfall by delaying new tools; cheaper memory brings back the rest, and the price that does it is <strong>47% below plan</strong>.</figcaption></figure>
<h2>How far prices fall: 2019 is the test</h2>
<p>We have run this experiment before. Hyperscale capex rose 43% in 2018 and grew only modestly in 2019 (Synergy). In December 2018 Micron said customers had built higher-than-normal inventories, which it expected them to work down by the second half of 2019, and cut its capex plan by $1.25B. Contract prices fell nearly 30% in the first quarter of 2019 alone, and an 8 GB PC module went from $68 in the third quarter of 2018 to $25.50 by July 2019 (TrendForce). DRAM revenue fell 37.5% in 2019 (Gartner) while bit shipments kept growing, so the price per bit fell by around half.</p>
<p>Some of that was expected. In October and November 2018 TrendForce had 2019 bit output up about 22% and prices down 15–20%. Prices ended about a third below that path. The surprise was mostly stock: in January North American data centers held 5–6 weeks of server memory, about twice normal, and by March server and PC customers held more than seven weeks (TrendForce).</p>
<p>The 2022–23 downturn was slower but deeper: DRAM fell about 57% and NAND 55% from the third quarter of 2021 to the 2023 trough (Yole), as consumer demand fell first and cloud buyers cut last. Makers cut capex and wafer starts four to six quarters into the slide: SK hynix halved its 2023 capex, and Samsung cut output from April 2023. Contract prices bottomed that third quarter and rose in the fourth (TrendForce).</p>
<p>Our model fits how far demand bends to price to 2019’s surprise, not to the whole fall. About three weeks of excess stock and demand growth a few points short of what was expected, 8.6% of a year’s bits, with makers holding back 30% of it, clears 35% below the expected path. Then it adds what is new this time: HBM that vendors can put back on chips, AI budgets that buy more when memory is cheaper, and makers who say they will pace new tools to demand. A flat 2028 is the bigger shock: 13.7% of 2028’s DRAM demand goes missing against plan.</p>
<h2>Prices and revenue by path</h2>
<p>Annual averages. DRAM and NAND: contract price index, 2026 = 100. HBM: contract price, $ per GB. Revenue: DRAM including HBM, plus NAND. 2026 is actual or TrendForce’s current estimate; 2027–28 are modeled. 2029 is in the data.</p>
<p>In the headline case, conventional DRAM contract prices average <strong>50% less in 2028 than in 2027</strong>, against 9% less on plan. With a flat 2028 alone, 2028 averages 46% less than 2027; with the blue wave alone, 19% less. The 2028 average is still about <strong>3.1 times</strong> mid-2025’s.</p>
<figure><a href="https://06.ai/packs/memory-prices"><img src="https://06.ai/newsletter/memory-prices/fig-04.png" alt="Prices and revenue by path" width="694" height="348"></a></figure>
<h2>Reading the paths</h2>
<p><strong>On plan, 2028 is a mild correction.</strong> DRAM peaks in the second quarter of 2027 and gives back 3% in the second half, as Citi expects, then drifts lower in 2028 as new fabs arrive; Bernstein expects 2028 to normalize, more mildly than it first thought. NAND peaks earlier: TrendForce has supply catching up from the second half of 2027. Annual averages: DRAM −9% and NAND −14% in 2028.</p>
<p><strong>Flat 2028 is 2019 and then some.</strong> On its own it takes DRAM 46% below 2027 and NAND 44%. The blue wave adds to 2028’s fall and takes some of 2027’s rise away: HBM is contracted, but much conventional DRAM and NAND is still priced month to month or quarter to quarter, so 2027 averages 3% below plan for DRAM and 7% for NAND. On its own, with 2028 capex still growing, the blue wave leaves 2028 DRAM 14% below plan. We count every dollar of Texas capex that leaves 2027 as hardware nobody buys that year. Alphabet and Microsoft say demand exceeds what they can supply (<a href="https://06.ai/packs/blue-wave-fcf">the </a><a href="https://06.ai/packs/blue-wave-fcf">blue-wave cash flow</a><a href="https://06.ai/packs/blue-wave-fcf"> pack</a>), so some of those chips would go elsewhere; if all of them did, the headline would be −48% instead of −50%.</p>
<p><strong>HBM falls less.</strong> HBM and DDR5 come off the same wafers, and an HBM bit takes about three times the wafer of a DDR5 bit. When AI buys less HBM, makers move the spare wafers back to DDR5, so conventional DRAM takes most of the fall. 2028’s HBM deals are struck in the second half of 2027, before 2028 budgets are final, so we pass half of the pool’s move through to them: $20.1 per GB, against $30.8 in 2027. If all of it comes through, $14.6, about 2026’s price.</p>
<p><strong>Revenue falls back toward 2026’s level.</strong> $1.48T in 2027 becomes $0.95T in 2028, against $0.89T in 2026. The makers stay very profitable: the 2028 DRAM price in the headline case is still about 3.1 times its mid-2025 level.</p>
<p><strong>2029 is mostly about the makers.</strong> If they cut 2029 supply as hard as they did in 2023, holding back 60% of the shortfall, prices recover much of the way to plan in 2029; if they don’t, they stay down. The 2029 figures are in the data under the chart; treat them loosely.</p>
<h2>What moves the answer</h2>
<figure><a href="https://06.ai/packs/memory-prices"><img src="https://06.ai/newsletter/memory-prices/fig-05.png" alt="What moves the answer" width="694" height="248"></a><figcaption>Headline case, 2028 DRAM contract price against 2027, changing one input at a time. Pink: deeper fall; blue: shallower.</figcaption></figure>
<h2>Where we sit against the forecasters</h2>
<p>Counterpoint’s Kang, the one forecaster we found who ties 2028 to capex, expects prices to hold through 2027 and fall in 2028 as capex growth approaches zero, and put declines of 20–30% on the table. Bernstein’s July note had SK hynix’s DRAM price peaking at the end of 2027 and falling about 53% in 2028; in September it said 2028’s normalization would be milder than it had expected. Our headline is near Bernstein’s July case. It falls further than Counterpoint’s for three reasons: 2027’s contracted HBM that 2028 must work through, the stock buyers built in the shortage, and supply that grows about a fifth in 2028 because the tools are already ordered. If makers hold back 60% of the shortfall instead of 30%, we get −35%: closer to Counterpoint, though still a deeper fall.</p>
<h2>What would soften it</h2>
<ul><li><strong>Discipline.</strong> Samsung says a large share of its 2027 capex increase goes to space and cleanrooms, with equipment adjusted flexibly to demand; SK hynix says Yongin’s cleanrooms open in phases based on demand. Tools deferred in late 2027 come straight out of 2028’s bits.</li><li><strong>HBM going back on chips.</strong> Vendors cut memory per package because HBM was short and dear, and chipmakers are reportedly evaluating 8-Hi stacks for Rubin Ultra and next year’s custom chips (TrendForce). The <a href="https://06.ai/packs/ai-memory-balance">memory</a> pack’s announced specs need about 45% more HBM in 2028 than the trimmed ones. Cheap HBM gets put back, if 2028 parts can take taller stacks in time.</li><li><strong>Phones and PCs.</strong> IDC has smartphone shipments down 16.7% this year, with DRAM and NAND costs up more than 300%, and TrendForce expected low-end phones to go back to 4 GB base models in 2026. Lower prices bring back both units and memory per device, with a lag of a product cycle.</li><li><strong>Flat dollars, more chips.</strong> Memory is about 30% of hyperscaler spending this year (SemiAnalysis); on plan we have DRAM at 21% of 2028’s AI capex and NAND at 9%. With budgets fixed in dollars, cheaper memory buys more servers. At half strength, that brings back about 4% of AI DRAM demand in 2028.</li><li><strong>Floors.</strong> Three-quarters of the revenue in Micron’s strategic agreements has a defined pricing framework, most of it with floor and ceiling prices, and SanDisk’s agreements carry $93.9B of minimum revenue at floor prices. They protect makers’ revenue, not market prices.</li></ul>
<h2>What would make it worse</h2>
<ul><li><strong>China.</strong> CXMT adds about 150,000 wafer starts a month between the end of 2026 and the end of 2028, and Citi names Chinese capacity as the main risk to its call.</li><li><strong>Contracts that push the shortfall onto the open market.</strong> If buyers must take contracted volumes, the whole shortfall lands on what is left: contract and spot prices fall further, even as the makers’ own average price holds up.</li><li><strong>More stock than we assume.</strong> Six weeks of destocking instead of three takes the headline to −60%.</li><li><strong>A harder Texas.</strong> The <a href="https://06.ai/packs/blue-wave-fcf">blue-wave cash flow</a> pack’s stress case, with the Texas House flipped and a moratorium to the end of 2027, gives −58%.</li></ul>
<h2>What to watch</h2>
<p><strong>Spot against contract.</strong> Spot usually turns first: in October 2018, as 2019’s slide began, DDR4 spot was already below contract ($6.95 against $7.31, TrendForce). Today DRAM spot is flat at about 70% above contract, and NAND wafer spot is falling.</p>
<p><strong>Late January and early February 2027.</strong> Fourth-quarter results bring the first 2027 capex guides from Alphabet and Meta and Amazon’s “meaningfully more.” Any 2028 color, and whether capex keeps outrunning operating cash flow, matters more for memory than any governor.</p>
<p><strong>TrendForce’s first-quarter 2027 contract outlook,</strong> due around the turn of the year. A rise under 5% for DRAM would fit the peak-in-2027 camp.</p>
<p><strong>Mid-2027: tool-in at Yongin and P5,</strong> and 2028 HBM negotiations. 2027’s deals closed in September and October; where 2028’s land, and whether buyers hold back, is the first hard read on 2028.</p>
<p><strong>Inventory.</strong> Micron’s inventory days rose by 9 last quarter to 129. Rising supplier inventory alongside falling spot is how 2019 began.</p>
<h2>What this is anchored on</h2>
<p>Prices, guidance and estimates are as reported. The capex paths are the <a href="https://06.ai/packs/ai-compute-balance">compute-balance</a>, <a href="https://06.ai/packs/texas-data-center-risk">Texas data center</a> and <a href="https://06.ai/packs/blue-wave-fcf">blue-wave cash flow</a> packs’; units and HBM are the <a href="https://06.ai/packs/ai-memory-balance">memory</a> pack’s; the price response, the plan path after 2026 and the revenue are ours.</p>
<ul><li><strong>Prices</strong> — 4Q26 forecast: DRAM +10–15%, NAND +15–20%; enterprise SSD the only NAND category still accelerating · <a href="https://www.trendforce.com/presscenter/news/20260930-13258.html">TrendForce, Sep 30</a></li><li><strong>Prices</strong> — 1Q26 conventional DRAM contract prices +93–98% (actual, against a +90–95% forecast) · <a href="https://www.trendforce.com/presscenter/news/20260601-13070.html">TrendForce, Jun 1</a></li><li><strong>Prices</strong> — 2Q26 forecast: DRAM +58–63%, NAND +70–75% · <a href="https://www.trendforce.com/presscenter/news/20260331-12995.html">TrendForce, Mar 31</a></li><li><strong>Prices</strong> — 3Q26 forecast: DRAM +13–18%, NAND +10–15% · <a href="https://www.trendforce.com/presscenter/news/20260703-13134.html">TrendForce, Jul 3</a></li><li><strong>Prices</strong> — 1Q26 NAND +55–60%; 4Q25 DRAM +45–50% · <a href="https://evertiq.com/news/2026-02-02-drampocalypse-memory-price-projections-skyrocket-for-1q26">TrendForce via Evertiq, Feb 2</a></li><li><strong>Prices</strong> — DDR4 8Gb contract $27.50 in October, a record and the 19th monthly rise; 128Gb MLC flat, ending a 21-month rise · <a href="https://www.thelec.kr/news/articleView.html?idxno=64634">TheElec, Oct 30</a></li><li><strong>Prices</strong> — DDR4 8Gb contract $1.35 in January–March 2025 · <a href="https://siliconanalysts.com/market-data/dram-ddr4-price">Silicon Analysts, DRAMeXchange data</a></li><li><strong>Spot</strong> — DDR4 8Gb spot $46.32, unchanged on the week; 512Gb TLC wafer $18.83 · <a href="https://www.trendforce.com/news/2026/10/07/insights-memory-spot-price-update-dram-spot-largely-sideways-as-holidays-month-end-effects-weigh/">TrendForce, Oct 7</a></li><li><strong>Spot</strong> — DDR4 8Gb spot $46.82, −0.09% on the week and above suppliers’ quotes; 512Gb TLC wafer $16.15, −4.33% · <a href="https://www.trendforce.com/news/2026/11/04/insights-memory-spot-price-update-mainstream-ddr4-prices-dip-0-09-as-spot-premium-keeps-buyers-on-sidelines/">TrendForce, Nov 4</a></li><li><strong>Spot</strong> — 512Gb TLC wafer spot −9.19% in a week as Chinese phone brands discount · <a href="https://www.trendforce.com/news/2026/10/28/insights-memory-spot-price-update-nand-spot-price-cuts-accelerate-amid-china-phone-discounts-512gb-tlc-sink-9-19/">TrendForce, Oct 28</a></li><li><strong>HBM</strong> — 2027 blended HBM price +121%; chipmakers evaluating 8-Hi stacks, which cost 10–20% more per Gb than 12-Hi · <a href="https://www.trendforce.com/presscenter/news/20260929-13255.html">TrendForce, Sep 29</a></li><li><strong>HBM</strong> — HBM4 12-Hi 36 GB stack about $600 in 2026 and $1,300 in 2027 (industry estimates reported by Seoul Economic Daily) · <a href="https://siliconanalysts.com/market/trendforce-lifts-2027-hbm-asp-forecast-to-121-hbm4-stack-seen-doubling-to-1-300-2026-10-01">Silicon Analysts, Oct 1</a></li><li><strong>Revenue</strong> — 2026 DRAM $618.7B and NAND $270.6B; 2027 DRAM $903.3B and NAND $379.4B (May forecast) · <a href="https://www.trendforce.com/presscenter/news/20260529-13068.html">TrendForce, May 29</a></li><li><strong>Makers</strong> — SK hynix: 80% operating margin in 3Q; DRAM and HBM effectively sold out through 2029; long-term agreements over half of 2027–28 capacity · <a href="https://www.trendforce.com/news/2026/10/28/news-sk-hynix-3q26-opm-hits-record-80-says-dram-hbm-effectively-sold-out-to-2029-14-lta-customers-signed/">TrendForce, Oct 28</a></li><li><strong>Makers</strong> — SK hynix: 2027 volumes and prices finalized; customers reviewing project timing, some adjusting HBM specs; industry DRAM demand up mid-20s% in 2027 · <a href="https://www.fool.com/earnings/call-transcripts/2026/10/28/sk-hynix-skhy-q3-2026-earnings-call-transcript/">SK hynix 3Q26 call, Oct 27</a></li><li><strong>Makers</strong> — Samsung: 2027 HBM sold out at target prices · <a href="https://en.sedaily.com/finance/2026/10/29/samsung-says-2027-hbm-supply-already-sold-out">Seoul Economic Daily, Oct 29</a></li><li><strong>Makers</strong> — Micron: industry DRAM bit shipments up low-20s% in both 2027 and 2028, NAND mid-20s%; gross margin 87%; 26 strategic agreements, over 35% of revenue to 2030, three-quarters of it with a defined pricing framework, a majority with floor and ceiling prices; Idaho ID1 mid-2027, ID2 and Japan late 2028 · <a href="https://s25.q4cdn.com/621799436/files/doc_financials/2026/q4/Q4-FY26-Prepared-Remarks.pdf">Micron, Sep 30</a></li><li><strong>Makers</strong> — SanDisk: gross margin 85.1%; about half of FY27 bits and two-thirds of FY28 under long-term agreements; data center about half of NAND bits, data-center bit demand up about 40% in 2027 · <a href="https://www.fool.com/earnings/call-transcripts/2026/10/30/sandisk-sndk-q1-2027-earnings-call-transcript/">SanDisk call, Oct 29</a></li><li><strong>Makers</strong> — SanDisk agreements: $93.9B of minimum revenue at floor prices · <a href="https://investor.sandisk.com/static-files/4ca97b6e-7da4-4d52-b4ac-0e80f462716f">SanDisk investor materials, Oct 29</a></li><li><strong>Balance</strong> — DRAM 1–2% short in 2026 and wider in 2027 · <a href="https://www.trendforce.com/presscenter/news/20260730-13158.html">TrendForce, Jul 30</a></li><li><strong>Balance</strong> — NAND 4–5% short in 2026, turning to surplus in the second half of 2027; server over 40% of NAND bit demand · <a href="https://www.trendforce.com/presscenter/news/20260721-13148.html">TrendForce, Jul 21</a></li><li><strong>Balance</strong> — HBM 18%, 22% and 30% of the big three’s DRAM wafer starts in 2025–27, and 8%, 9% and 13% of bits · <a href="https://www.trendforce.com/presscenter/news/20260602-13074.html">TrendForce, Jun 2</a></li><li><strong>Balance</strong> — 2026 DRAM shipments: servers 48%, HBM 9%, mobile ~22%, PCs ~10% · <a href="https://www.techtimes.com/articles/319284/20260701/more-half-dram-now-goes-data-centers-counterpoint-says.htm">Counterpoint via TechTimes, Jul 1</a></li><li><strong>Forecasts</strong> — DRAM and NAND prices peak in 2Q27; 2H27 DRAM −3% and NAND −5% half on half; Chinese capacity the main risk · <a href="https://finance.yahoo.com/markets/stocks/articles/citi-cuts-micron-stock-target-124820288.html">Citi, Aug 7</a></li><li><strong>Forecasts</strong> — Counterpoint’s Kang Kyung-soo at iForum 2026: prices stable from 4Q26 through 2027, then a fall in 2028, with DRAM and NAND declines of 20–30% possible; Big Tech capex growth approaching 0% in 2028; capex over 80% of operating cash flow · <a href="https://finance.biggo.com/news/c3a3fef8-8dfc-4a8b-9acf-c3ae851ea8c8">BigGo Finance, Oct 22</a></li><li><strong>Forecasts</strong> — 2028 normalizes, more mildly than previously forecast; DRAM gross margin about 90% at the peak, high 70s% by end-2028 · <a href="https://www.itiger.com/news/1157176298">Bernstein via Tiger Brokers, Sep 29</a></li><li><strong>Forecasts</strong> — SK hynix DRAM price peaking at the end of 2027 and falling about 53% in 2028 · <a href="https://finance.yahoo.com/technology/articles/analysts-think-could-see-substantial-122435924.html">Bernstein via Yahoo Finance, Jul 7</a></li><li><strong>Forecasts</strong> — 2027 commodity DRAM +40%, NAND +32%, HBM +121% · <a href="https://en.fnnews.com/news/202611020851306461">KB Securities via FN News, Nov 2</a></li><li><strong>Capex</strong> — Large US hyperscalers about $1.2T in 2027 and $1.4T in 2028, against a consensus near $1.3T · <a href="https://www.techflowpost.com/en-US/article/34264">Goldman Sachs via TechFlow, Sep 25</a></li><li><strong>Capex</strong> — Hyperscalers and AI cloud providers $1.58T in 2027 and $1.8T in 2028 · <a href="https://fortune.com/2026/10/19/nvidia-concentration-hyperscaler-1-6-trillion-92-percent-exposure/">Morgan Stanley via Fortune, Oct 19</a></li><li><strong>Capex</strong> — S&amp;P Global’s models generally assume capex flattens in 2028 · <a href="https://finance.yahoo.com/technology/ai/articles/6-hyperscalers-driving-ai-revolution-104200428.html">Yahoo Finance, Sep 6</a></li><li><strong>Capex</strong> — Memory about 30% of hyperscaler spending in 2026, against about 8% in 2023–24 · <a href="https://www.tomshardware.com/tech-industry/memory-will-consume-30-percent-of-hyperscaler-spending-this-year">SemiAnalysis via Tom’s Hardware, Apr 3</a></li><li><strong>Capex</strong> — Worldwide AI capex plan; Texas GW paths; capex leaving 2027 · <a href="https://06.ai/packs/blue-wave-fcf">06.ai </a><a href="https://06.ai/packs/blue-wave-fcf">compute-balance</a><a href="https://06.ai/packs/blue-wave-fcf">, </a><a href="https://06.ai/packs/blue-wave-fcf">Texas data center</a><a href="https://06.ai/packs/blue-wave-fcf"> and </a><a href="https://06.ai/packs/blue-wave-fcf">blue-wave cash flow</a><a href="https://06.ai/packs/blue-wave-fcf"> packs</a></li><li><strong>AI memory</strong> — Accelerator units, HBM per package and HBM supply · <a href="https://06.ai/packs/ai-memory-balance">06.ai </a><a href="https://06.ai/packs/ai-memory-balance">memory</a><a href="https://06.ai/packs/ai-memory-balance"> pack</a></li><li><strong>AI memory</strong> — Nvidia reportedly halves Vera Rubin SOCAMM capacity; memory near 29% of the system bill of materials · <a href="https://www.trendforce.com/news/2026/07/28/news-nvidia-reportedly-halves-vera-rubin-socamm-capacity-as-memory-costs-near-29-of-system-bom/">TrendForce, Jul 28</a></li><li><strong>AI memory</strong> — Nvidia’s context-memory storage tier: up to 16 TB per GPU · <a href="https://blocksandfiles.com/2026/01/12/nvidias-basic-context-memory-extension-infrastructure/">Blocks &amp; Files, Jan 12</a></li><li><strong>Devices</strong> — Smartphone shipments −16.7% in 2026; DRAM and NAND costs up over 300% year on year · <a href="https://www.idc.com/resource-center/blog/smartphone-shipments-set-for-record-16-7-drop-in-2026-as-the-memory-crisis-hits-full-force/">IDC, Aug 26</a></li><li><strong>Devices</strong> — Low-end smartphones likely back to 4 GB base models in 2026 · <a href="https://www.trendforce.com/presscenter/news/20251211-12831.html">TrendForce, Dec 11, 2025</a></li><li><strong>Devices</strong> — PC shipments −11.3% in 2026 · <a href="https://www.idc.com/resource-center/blog/pc-market-enters-volatile-territory-as-memory-shortage-persists-through-2027/">IDC, Jun 2</a></li><li><strong>Inventory</strong> — Customers built 2–9 weeks of DRAM stock in 2026; restocking 8–13% of 2027 shipments · <a href="https://koreainvestinsights.substack.com/p/part-of-the-dram-shortage-is-inventory">Korea Invest Insights, Oct 30</a></li><li><strong>Supply</strong> — Samsung: a large share of the 2027 capex increase goes to space and cleanrooms; equipment adjusted flexibly to demand · <a href="https://www.businesspost.co.kr/BP?command=article_view&amp;num=450616">BusinessPost, Oct 29</a></li><li><strong>Supply</strong> — SK hynix Yongin fab 1 first cleanroom February 2027 · <a href="https://in.marketscreener.com/news/sk-hynix-speeds-up-new-chip-fab-opening-to-meet-memory-demand-executive-says-ce7e58d9db88f623">Reuters via MarketScreener, Jan 15</a></li><li><strong>Supply</strong> — Samsung P5 phase 1: tools from mid-2027, mass production late 2028 · <a href="https://www.techtimes.com/articles/326838/20260907/samsung-p5-cleanroom-piping-starts-contract-confirms-phase-one-under-active-construction.htm">TechTimes, Sep 7</a></li><li><strong>Supply</strong> — CXMT about 265k wafer starts a month at end-2025, 350k at end-2026, 420k at end-2027, 500k at end-2028 · <a href="https://xenospectrum.com/en/cxmt-2026-dram-wafer-capacity-micron-350k-wspm/">SemiAnalysis via XenoSpectrum, Jul 16</a></li><li><strong>2019</strong> — 2019 DRAM bit output up nearly 22%; prices down 15–20% · <a href="https://www.trendforce.com/presscenter/news/20181009-10059.html">TrendForce, Oct 9, 2018</a></li><li><strong>2019</strong> — Micron: customers carrying higher-than-normal inventory, worked down by 2H19; capex cut by $1.25B; DRAM bit output growth held to about 15% · <a href="https://www.electronicdesign.com/embedded-revolution/micron-cut-capital-expenses-next-year-125-billion">Electronic Design, Dec 2018</a></li><li><strong>2019</strong> — Hyperscale capex +43% in 2018 · <a href="https://www.srgresearch.com/articles/hyperscale-operator-capex-jumps-43-2018-while-telco-spending-remains-flat">Synergy Research, Mar 2019</a></li><li><strong>2019</strong> — Hyperscale capex growth only modest in 2019 · <a href="https://www.srgresearch.com/articles/hyperscale-operator-spending-data-centers-11-2019-despite-only-modest-capex-growth">Synergy Research, Mar 2020</a></li><li><strong>2019</strong> — 8 GB PC module $68 in 3Q18; 2019 DRAM prices down as much as about 20%; October DDR4 8Gb contract $7.31, spot $6.95 · <a href="https://www.trendforce.com/presscenter/news/20181105-10065.html">TrendForce, Nov 5, 2018</a></li><li><strong>2019</strong> — North American data centers holding 5–6 weeks of server DRAM, about double normal · <a href="https://www.trendforce.com/presscenter/news/20190121-10096.html">TrendForce, Jan 21, 2019</a></li><li><strong>2019</strong> — First-quarter contract prices down nearly 30%, the steepest fall since 2011 · <a href="https://techpowerup.com/253337/trendforce-recent-dram-pricing-decline-biggest-since-2011-nearly-30-decrease">TrendForce via TechPowerUp, Mar 6, 2019</a></li><li><strong>2019</strong> — Server and PC customers holding over 7 weeks of inventory, suppliers over 6 · <a href="https://www.trendforce.com/presscenter/news/20190325-10115.html">TrendForce, Mar 25, 2019</a></li><li><strong>2019</strong> — 8 GB module $25.5 by July 2019 · <a href="https://trendforce.com/presscenter/news/20190808-10172.html">TrendForce, Aug 8, 2019</a></li><li><strong>2019</strong> — DRAM revenue −37.5% in 2019 · <a href="https://www.computing.co.uk/news/3085045/dram-semiconductors-intel">Gartner via Computing, Jan 2020</a></li><li><strong>2022–23</strong> — DRAM −57% and NAND −55% from 3Q21 · <a href="https://www.storagenewsletter.com/2023/07/21/ww-2023-memory-market-declining-to-42-billion-for-dram-47-y-y-and-37-billion-for-nand-37-y-y/">Yole via StorageNewsletter, Jul 2023</a></li><li><strong>2022–23</strong> — 2022 DRAM sufficiency +7.9% (supply +18.7%, demand +11.9%) · <a href="https://securities.miraeasset.com/bbs/download/2107811.pdf?attachmentId=2107811">Mirae Asset, May 2023</a></li><li><strong>2022–23</strong> — SK hynix cuts 2023 capex by more than half · <a href="https://www.tomshardware.com/news/sk-hynix-to-halve-capex-amid-drop-of-memory-demand">Tom’s Hardware, Oct 2022</a></li><li><strong>2022–23</strong> — Samsung’s production cut, April 2023 · <a href="https://www.kedglobal.com/newsView/ked202304120020">KED Global, Apr 2023</a></li><li><strong>2022–23</strong> — Contract prices bottomed in 3Q23; 4Q23 forecast +13–18% as Samsung deepened its cuts · <a href="https://www.trendforce.com/presscenter/news/20231204-11942.html">TrendForce, Dec 4, 2023</a></li></ul>
<h2>Method</h2>
<p><strong>Capex.</strong> Worldwide AI capex on the <a href="https://06.ai/packs/ai-compute-balance">compute-balance</a> pack’s base: $1.0T in 2026, $1.40T in 2027, $1.65T in 2028, $1.85T in 2029 and $2.05T in 2030. Flat 2028 sets 2028 equal to 2027’s plan, then applies the plan’s growth rates. The blue wave takes the <a href="https://06.ai/packs/texas-data-center-risk">Texas data center</a> pack’s Texas paths (gigawatts switched on, Abbott’s path less the Democratic outcome) and prices them as the <a href="https://06.ai/packs/blue-wave-fcf">blue-wave cash flow</a> pack did: 2.4 GW by the end of 2027 at $50B a GW, 75% of it leaving the year, is $89B; the gap widening to 3.8 GW by the end of 2028 takes $53B more. Low and stress cases are the <a href="https://06.ai/packs/blue-wave-fcf">blue-wave cash flow</a> pack’s.</p>
<p><strong>Memory for AI.</strong> Accelerator packages are the <a href="https://06.ai/packs/ai-memory-balance">memory</a> pack’s base by family, scaled each year by the path’s AI capex against plan, so a cut falls evenly on chips and buildings (pack 019’s rule). HBM bought in a year is the <a href="https://06.ai/packs/ai-memory-balance">memory</a> pack’s K × [(1 − L) × C(y) + L × C(y + 1)], with L = 0.35, K = 1.06 and C the HBM in the year’s shipments at trimmed specs. Host DRAM per package: 240 GB per Blackwell GPU (Grace), 448 GB per Rubin GPU (Vera, between the reported cut and the full 1.5 TB per CPU), x86 hosts about 256 GB per GPU, TPUs and Trainium 160–256 and 128–224 GB, AMD 384–500 GB; a quarter is bought the year before. AI storage is 20–26 TB of NAND per package, including the share of training data and context-memory tiers, which puts AI at more than half of data-center NAND. 2026 is history. 2027’s HBM and the late-2027 orders for 2028’s builds are bought on plan; anything beyond the path’s need is used up first in 2028.</p>
<p><strong>Supply and the plan balance.</strong> DRAM is one wafer pool counted in conventional-equivalent bits, with an HBM bit worth 2.9 (TrendForce: HBM 30% of 2027 wafers for 13% of bits). DRAM output is about 43 EB in 2026, HBM 3.9 EB of it, growing 22% in 2027 and 21% in 2028 (Micron, low 20s %) and 18% in 2029. NAND is about 1,150 EB in 2026, growing 25% in 2027–28 and 20% after. Plan demand sits at the published balance: DRAM −1.5%, −3.5% and −2.0% in 2026–28 (supply over demand, less one); NAND −4.5%, 0.0% and 2.0%. Non-AI demand is plan demand less AI.</p>
<p><strong>Clearing.</strong> Supply is fixed by fabs already built or tooling. In each path and year we solve for the price against plan, p, at which demand equals plan demand times the path’s supply over plan supply. Each segment’s demand moves with p to the power −e: HBM 0.3 (vendors restoring stacks; HBM’s contract price moves as p to the power 0.5 in 2028), host DRAM 0.118, other DRAM 0.148, AI storage 0.222, other NAND 0.178. AI segments also gain from fixed budgets buying more, units × [1 / (1 − m + m c)]^0.5, where m is the memory’s share of AI capex at plan prices (DRAM 21% and NAND 9% in 2028) and c its cost against plan, with HBM at its own price (63% of the AI DRAM bill). In the year a flat path turns, buyers run down 3 weeks of demand and makers hold back 30% of the shortfall; in later years they hold back 60%, as capex cut in 2028 lands. HBM contracts are fixed for 2027 and take 50% of the pool’s move in 2028. The elasticities other than HBM’s keep fixed proportions and are scaled together to 2019’s surprise: TrendForce expected 2019 prices to fall 15–20%; with revenue down 37.5% and bit output growing in the teens, the price per bit fell by around 45–48%, about 35% below the expected path. We put the shortfall behind that at 8.6% of a year’s bits (three weeks of excess stock and demand growth 3 points short of what was expected), with makers holding back 30% of it.</p>
<p><strong>Prices.</strong> Quarterly contract prices to 4Q26 chain TrendForce’s midpoints. Chained, they put 4Q26 conventional DRAM at about six times 3Q25. The plan path then follows published views: DRAM +6% and +3% in the first two quarters of 2027 and −3% in each of the last two (−3.0% half on half, as Citi expects), then −2.5% a quarter in 2028 (ours); NAND +4%, 0%, −3% and −4% (−4.9% half on half), then −4% a quarter. 2026 average prices are set so revenue matches TrendForce’s 2026 figures; HBM is $1.74 per Gb in 2026 (our blend of HBM3E and HBM4), +121% in 2027 (TrendForce) and −10% a year on plan after. Each path’s annual average is the plan’s times p.</p>
<p><strong>Known limits.</strong> Every elasticity is a judgment fitted to one downturn’s surprise, and the tornado shows how much rides on it. Annual steps hide timing: in a flat 2028 we would expect spot to turn in late 2027, as 2028 budgets are set, and contract prices a quarter or two later. China’s domestic market is not modeled separately. The model doesn’t value the makers or their shares. This is analysis, not investment advice.</p>
<hr>
<p>Every chart above links to its interactive version on 06.ai, with the data behind it. The daily trackers are at <a href="https://06.ai/trackers">06.ai/trackers</a>.</p>
<p><em>Not investment advice. Estimates are labeled as estimates.</em></p>]]></content:encoded>
</item>
<item>
  <title>Will the blue wave fix Big Tech’s free cash flow?</title>
  <link>https://inthesingularity.ai/packs/blue-wave-fcf</link>
  <guid isPermaLink="true">https://inthesingularity.ai/packs/blue-wave-fcf</guid>
  <pubDate>Wed, 04 Nov 2026 12:21:00 GMT</pubDate>
  <description>Amazon, Alphabet and Meta spent more than they brought in last quarter. Getting back above zero in 2027 means pushing about a fifth to a third of their planned capex out of the year. Tuesday’s results push out about 3%, and most of it lands on Microsoft, which is already in the black.</description>
  <category>Capex &amp; politics</category>
  <enclosure url="https://inthesingularity.ai/newsletter/blue-wave-fcf/cover.png" length="29984" type="image/png" />
  <media:content url="https://inthesingularity.ai/newsletter/blue-wave-fcf/cover.png" medium="image" type="image/png" />
  <content:encoded><![CDATA[<p>Amazon, Alphabet and Meta all spent more on capex than their operations brought in during the third quarter. To get free cash flow back above zero in 2027 they would have to push <strong>22% to 32% of the capex the Street expects</strong> out of the year. Our base case for what Tuesday’s results push out is <strong>about $32B across Amazon, Alphabet, Meta and Microsoft, 3% of their 2027 capex</strong>, and 61% of it lands on Microsoft, the one of the four already in the black. A data center that switches on late is still built. At Street estimates, none of the three burners gets back to zero on politics alone, even in our stress case.</p>
<p><a href="https://06.ai/packs/blue-wave-fcf">Interactive version and data on 06.ai →</a></p>
<ul><li><strong>3 of 7</strong> — Spent more than they made, Q3. Amazon −$33.5B · Alphabet −$9.1B · Meta −$2.1B</li><li><strong>22%–32%</strong> — 2027 capex to push out to reach zero. Meta to Amazon: Street capex, operating cash flow up 20% a year</li><li><strong>$32B</strong> — What the vote pushes out of 2027. <strong>3%</strong> of the Big Four’s 2027 capex. Range $11B–$79B</li><li><strong>61%</strong> — Share that lands on Microsoft. Its Abilene expansion is the most exposed Big Four campus in Texas. Microsoft made $59.9B in the last 12 months</li></ul>
<h2>What changed on Tuesday</h2>
<p>Democrats won control of the House and the Senate, according to projections by the AP, NBC, CBS and CNN, flipping Senate seats in Iowa, Kansas, Ohio, Maine, North Carolina, South Carolina and Texas; Alaska and Mississippi were still uncalled Wednesday morning. They flipped the governor’s office in five states Trump carried, Texas, Georgia, Ohio, Iowa and Nevada, and held Arizona, Michigan, Wisconsin and Pennsylvania. Further down the ballot, Democrats won Texas’s lieutenant governorship and a majority on Georgia’s utility regulator, and claimed Pennsylvania’s state Senate; at least 14 Ohio towns passed data center bans or limits in unofficial results, including Sunbury, where the ban is aimed at a possible Amazon project.</p>
<p>Data center limits were not a one-party platform. In Ohio both nominees for governor proposed moratoriums; Michigan’s Republican Senate nominee backed a one-year pause; and in Texas it was Gov. Greg Abbott who froze data center grid connections in August and permits in September. What changed is who holds the levers in the biggest markets: Wells Fargo puts 27% of the planned US data center gigawatts in Texas, 7% in Georgia and 5% each in Ohio and Pennsylvania, and Democrats now hold all four governorships. Gina Hinojosa, who takes office in Texas on January 19, has pledged a moratorium on new data centers until the Legislature writes rules. Our <a href="https://06.ai/packs/texas-data-center-risk">Texas pack (020)</a> sizes that in gigawatts. This one asks what it does to the cash.</p>
<p>Markets read the night as a blow to the builders rather than the payers. Nasdaq-100 futures fell as much as 2.4% overnight before recovering. By 7:26 am ET, Reuters had Amazon, Meta and Microsoft trading higher and Nvidia up 0.5%, while SpaceX and Tesla were down about 4% and IREN, Hut 8, Applied Digital, CoreWeave and Nebius were all lower. In his first public comment after the results, Elon Musk wrote on X: “Data centers in space looking pretty appealing right now.”</p>
<h2>Only Amazon is negative over twelve months</h2>
<figure><a href="https://06.ai/packs/blue-wave-fcf"><img src="https://06.ai/newsletter/blue-wave-fcf/fig-01.png" alt="Only Amazon is negative over twelve months" width="694" height="269"></a><figcaption>Free cash flow over each company’s last twelve reported months, $B, on its own definition. Microsoft’s subtracts cash paid for property and equipment; subtracting its finance leases too gives $40.3B. Periods end September 30, 2026, except Apple (September 26) and Nvidia (July 26).</figcaption></figure>
<h2>The quarter is worse than the year</h2>
<p>Twelve-month figures still carry 2025’s lighter spending. In the third quarter alone, Amazon spent $75.9B on property and equipment against $42.4B of operating cash flow, a −$33.5B quarter. Alphabet ran −$9.1B, its second negative quarter in a row, and Meta −$2.1B. Microsoft made $18.6B in its September quarter, and its CFO says it stays free-cash-flow positive through fiscal 2027. Tesla made $1.5B after cutting its 2026 capex guide to more than $20B; Apple and Nvidia spend little on capex of their own.</p>
<p>The full-year guides set the fourth quarter. Alphabet’s $195–205B for 2026 leaves $63.0B–$73.0B for the fourth quarter, against $51.4B in the third; Meta’s $135–145B leaves $49.2B–$59.2B, against $34.9B. Amazon’s roughly $220B leaves only about $47.8B, against $75.9B in the third quarter: either its spending slows sharply in the fourth quarter or 2026 comes in above the guide. Microsoft’s “over $50 billion” is consistent with its $170B for the calendar year.</p>
<h2>Capex outran cash at three of the four</h2>
<figure><a href="https://06.ai/packs/blue-wave-fcf"><img src="https://06.ai/newsletter/blue-wave-fcf/fig-02.png" alt="Capex outran cash at three of the four" width="694" height="322"></a><figcaption>Third quarter of 2026, $B: operating cash flow, capex, and the fourth-quarter capex implied by each company’s 2026 guidance (guidance less the first nine months; midpoint of a range). Capex is on each company’s guidance basis: Amazon net of proceeds and incentives, Meta with finance-lease principal, Microsoft with finance leases.</figcaption></figure>
<h2>What it takes to get back above zero in 2027</h2>
<p>Every one of them plans to spend more next year. Amazon expects to spend “meaningfully more” than this year’s $220B, Alphabet to raise capex “significantly,” Meta to keep investing aggressively; the Street case we used in <a href="https://06.ai/packs/hyperscaler-capex-funding">the </a><a href="https://06.ai/packs/hyperscaler-capex-funding">hyperscaler</a><a href="https://06.ai/packs/hyperscaler-capex-funding"> pack</a> has Alphabet at $300B, Amazon $290B, Microsoft $215B and Meta $205B, and the highest published estimates run to $395B for Amazon (JPMorgan) and $357B for Alphabet (S&amp;P Global). Grow each company’s trailing operating cash flow 20% a year and 2027 free cash flow comes out at <strong>−$78.6B for Amazon, −$74.5B for Alphabet and −$37.8B for Meta</strong>; Microsoft stays at $31.0B even counting its finance leases. S&amp;P Global Ratings, which cut Meta’s outlook to negative on October 30, has Meta at −$71B in 2027 on its own measure.</p>
<p>To close those gaps by spending less, Amazon would have to push about $92B of its 2027 capex into 2028, Alphabet $87B and Meta $44B. The amounts are larger than the gaps because a server bought in 2028 instead of 2027 takes its tax deduction with it: under the 100% bonus depreciation made permanent last year, each dollar pushed out frees about 85 cents of cash. At $50B a gigawatt, three-quarters of it spent in the year the capacity switches on, that is roughly 2.5 GW of Amazon’s own capacity slipping a year, 2.3 GW of Alphabet’s and 1.2 GW of Meta’s.</p>
<h2>What has to slip, and what the vote slips</h2>
<p>Share of each company’s 2027 capex that would have to move into 2028 for 2027 free cash flow to reach zero (pink), against what Tuesday’s results push out (strong grey: base case, dashed to the stress case). Pick the capex estimate and how fast operating cash flow grows from its trailing level.</p>
<p><strong>Street now</strong> capex, operating cash flow up <strong>20% a year</strong>: Amazon needs 32%, Alphabet needs 29%, Meta needs 22%. Neither the base nor the stress case gets any of them to zero.</p>
<figure><a href="https://06.ai/packs/blue-wave-fcf"><img src="https://06.ai/newsletter/blue-wave-fcf/fig-03.png" alt="What has to slip, and what the vote slips" width="694" height="358"></a></figure>
<figure><a href="https://06.ai/packs/blue-wave-fcf"><img src="https://06.ai/newsletter/blue-wave-fcf/fig-04.png" alt="What has to slip, and what the vote slips — Consensus, early Aug" width="694" height="389"></a><figcaption><strong>Consensus, early Aug.</strong></figcaption></figure>
<figure><a href="https://06.ai/packs/blue-wave-fcf"><img src="https://06.ai/newsletter/blue-wave-fcf/fig-05.png" alt="What has to slip, and what the vote slips — Highest published" width="694" height="389"></a><figcaption><strong>Highest published.</strong></figcaption></figure>
<h2>Why the vote moves so little</h2>
<p>Start with Texas, where the vote matters most. The <a href="https://06.ai/packs/texas-data-center-risk">Texas data center</a> pack’s model has Texas switching on <strong>2.38 GW less</strong> new data center load by the end of 2027 than on the path Abbott had set, with Hinojosa and Goodwin in office and a moratorium that holds until new laws take effect in September (range 1.52–2.77 GW). At $50B a gigawatt, with three-quarters of each gigawatt’s capex spent in the year it switches on, that is <strong>$89B</strong> of capex leaving 2027.</p>
<p>Most of it is not the Big Four’s. The campuses that wait longest under a Democratic governor are the self-powered ones whose air permits are pending: Anthropic’s Hill County campus, OpenAI and Oracle’s Frontier, Fluidstack’s Cameron County site, and Microsoft’s Abilene expansion with Crusoe. Grid-tied Big Four campuses (Google’s Armstrong and Wilbarger, Amazon’s Comanche Peak, Meta’s Childress) wait in ERCOT’s queue with everyone else, and Amazon’s GW Ranch already holds its air permit. Allocated campus by campus, the Big Four’s share of the 2027 shortfall is <strong>0.84 GW, about $32B of capex</strong>, and 61% of that is Microsoft’s Abilene expansion, whose first building was due mid-2027. Amazon’s share is $3.9B, Alphabet’s $2.9B and Meta’s $5.6B. The rest of the $89B belongs to OpenAI and Oracle, the neoclouds, the former bitcoin miners and Anthropic, which fits premarket trading: the builders fell, the hyperscalers didn’t.</p>
<h2>From gigawatts to free cash flow</h2>
<figure><a href="https://06.ai/packs/blue-wave-fcf"><img src="https://06.ai/newsletter/blue-wave-fcf/fig-06.png" alt="From gigawatts to free cash flow" width="694" height="288"></a><figcaption>Base case, 2027, $B: Texas capex that leaves 2027 because data center load switches on later, how much of it is the Big Four’s, and what it adds to their 2027 free cash flow.</figcaption></figure>
<h2>At a bigger scale</h2>
<p>The stress case assumes everything breaks against the builders: Democrats also take the Texas House, the moratorium lasts to the end of 2027, Georgia and Ohio act with half Texas’s force, Big Four campuses waiting on ERCOT are first in line to be delayed, a gigawatt costs $60B and 90% of its capex leaves the year. That pushes out <strong>$79B, 8% of the Big Four’s 2027 capex</strong>: $18.9B of Meta’s, $13.4B of Amazon’s and $11.7B of Alphabet’s. Meta comes closest, to −$22.6B. Nobody gets to zero.</p>
<p>The only setting in the chart above where the vote flips anyone is Meta at August’s consensus of $186B and operating cash flow growing 30% a year, where Meta is within $0.9B of zero before the vote. Meta’s own 2027 guide, due with its fourth-quarter results in late January, matters more than any governor.</p>
<p>Washington can’t scale it up before 2029. A national moratorium, which Sen. Bernie Sanders and Rep. Alexandria Ocasio-Cortez have proposed, would need President Trump’s signature or a two-thirds override. The new majorities can hold hearings, and Sen. Ron Wyden’s plan for an excise tax on data centers that would also deny them bonus depreciation and the research credit would cut the other way if it ever passed: higher cash taxes lower operating cash flow without building a single gigawatt less. One forecaster, MOCA Economics, called a Democratic Congress a “nothingburger for policy” for the industry.</p>
<h2>A deferral isn’t free cash</h2>
<ul><li><strong>The tax deduction moves too.</strong> Servers and network gear are written off the year they’re bought. Push $32B into 2028 and roughly 15% of it comes back as 2027 cash taxes.</li><li><strong>So does the revenue.</strong> AWS grew 39% last quarter, Google Cloud’s backlog is $561B, and Alphabet and Microsoft both say demand exceeds what they can supply. A gigawatt that switches on in 2028 instead of late 2027 earns its cloud revenue (about $8–13B a gigawatt-year in the <a href="https://06.ai/packs/neocloud-capex-funding">neocloud</a> pack) later. In 2027 that costs little; in 2028 it costs growth.</li><li><strong>Each gigawatt gets dearer.</strong> The rules both parties now favor make data centers pay for their own power and grid upgrades, which the Big Four already promised in March’s Ratepayer Protection Pledge, and repeal sales tax exemptions worth $1.3B or more a year in Texas alone. Both add capex per gigawatt even as the gigawatts slow.</li><li><strong>The spending moves.</strong> Equipment that can’t switch on in Abilene can go to Louisiana, Wyoming, the Gulf or Scandinavia, and Nvidia, which says demand runs well ahead of the roughly 70% growth its supply allows next fiscal year, has other buyers for any GPU a hyperscaler defers. For the Magnificent Seven as a group the vote barely moves free cash flow, since one member’s capex is another’s revenue.</li></ul>
<h2>How the number could still turn positive</h2>
<p><strong>By choice.</strong> The premarket moves suggest some investors hope the vote gives management cover to spend less. If any of the three gets back above zero in 2027, it will be because it decided to, and the decisions arrive with fourth-quarter results in late January and early February, when Alphabet and Meta give their first 2027 numbers.</p>
<p><strong>By lease.</strong> Capex counts what a company buys. A data center a developer builds and leases out shows up later, as lease payments. Microsoft had $364.6B of leases signed but not yet started at September 30, Meta $355.7B, Amazon $164.9B and Alphabet $130.5B (<a href="https://06.ai/wire/w007">W.007</a>); each is several times the 2027 gap. If moratoriums push the building of data centers onto developers in friendlier states, reported capex falls and free cash flow rises without the obligation getting any smaller. Read any 2027 turn to positive free cash flow next to that line.</p>
<p><strong>By growth.</strong> At 30% a year operating cash flow growth and August’s consensus capex, Meta is roughly at zero and Amazon and Alphabet need 11% and 15%: cash flow, not politics, is the bigger lever.</p>
<h2>What to watch</h2>
<p><strong>December 10 and 17.</strong> ERCOT files its audit of Texas data centers and the utility commission acts on it, with Abbott still governor; what it approves then is hard for Hinojosa to undo.</p>
<p><strong>January 19.</strong> Whether Hinojosa’s order covers only new projects or anything not yet switched on, and whether Abbott-appointed agencies follow it. Microsoft’s Abilene expansion is the Big Four campus most exposed.</p>
<p><strong>Late January.</strong> Fourth-quarter results and the first 2027 capex guides from Meta and Alphabet, Amazon’s “meaningfully more,” and Microsoft’s second-half acceleration. A cut there is worth more to free cash flow than any moratorium.</p>
<p><strong>Georgia’s new commission majority and Ohio’s local bans.</strong> Who pays for grid upgrades for Georgia Power’s large-load pipeline, and whether Sunbury’s ban holds for Amazon.</p>
<p><strong>The lease line.</strong> Leases signed but not yet started, in each 10-Q and 10-K: if capex falls while it keeps rising, the cash flow is coming from the accounting, not the buildout.</p>
<h2>What this is anchored on</h2>
<p>Results, filings and estimates are as reported; the gigawatt paths are the <a href="https://06.ai/packs/texas-data-center-risk">Texas data center</a> pack’s; the allocation to companies, the dollars per gigawatt and the effects on free cash flow are ours.</p>
<ul><li><strong>Results</strong> — Democrats projected to win control of the House and the Senate; Trump’s approval at 38% in exit polls · <a href="https://www.cbsnews.com/live-updates/election-day-2026-midterms/">CBS News, Nov 4</a></li><li><strong>Results</strong> — Senate flips in Iowa, Kansas, Ohio, Maine, North Carolina, South Carolina and Texas; governorships flipped in Texas, Georgia, Nevada, Ohio and Iowa · <a href="https://www.nbcnews.com/politics/2026-election/live-blog/midterm-election-results-2026-news-trump-live-updates-rcna605763">NBC News, Nov 4</a></li><li><strong>Results</strong> — Alaska and Mississippi Senate races too early to call at 8:20 am ET; Democrats flipped at least 27 House seats so far · <a href="https://edition.cnn.com/2026/11/04/politics/live-news/midterm-election-congress">CNN, Nov 4</a></li><li><strong>Results</strong> — Democrats claimed Pennsylvania’s state Senate, a trifecta, with most contested races not yet called; were on track to flip the Texas House early Wednesday (not called) · <a href="https://stateline.org/2026/11/04/democrats-set-to-win-majority-of-governorships-for-the-first-time-in-16-years/">Stateline, Nov 4, 1:13 am</a></li><li><strong>Texas</strong> — Vikki Goodwin defeated Lt. Gov. Dan Patrick · <a href="https://www.cbsnews.com/texas/live-updates/texas-midterm-elections-2026/">CBS News Texas, Nov 4</a></li><li><strong>Texas</strong> — Hinojosa pledged a moratorium on new data center construction, stricter water and power rules and an end to tax breaks; Texas holds more than 25% of planned US capacity · <a href="https://finance.yahoo.com/economy/article/a-democrat-won-the-texas-governors-race-after-promising-to-halt-new-data-center-construction-044536389.html">Yahoo Finance, Nov 4</a></li><li><strong>Texas</strong> — Aug 3 audit and pause on ERCOT data center interconnections; Sep 21 halt on TCEQ data center permits until the audit ends · <a href="https://www.globalelr.com/2026/10/texas-pauses-certain-data-center-approvals-amid-grid-water-and-community-concerns/">Global Environmental &amp; Land Resources, Oct</a></li><li><strong>Texas</strong> — GW switched on later by the end of 2027 under each outcome; campuses and what each waits on · <a href="https://06.ai/packs/texas-data-center-risk">06.ai </a><a href="https://06.ai/packs/texas-data-center-risk">Texas data center</a><a href="https://06.ai/packs/texas-data-center-risk"> pack</a></li><li><strong>States</strong> — Share of planned US data center GW: Texas 27%, Georgia 7%, Ohio 5%, Pennsylvania 5% (Wells Fargo) · <a href="https://www.cnbc.com/2026/11/04/ai-stocks-texas-governor-data-centers.html">CNBC, Nov 4</a></li><li><strong>States</strong> — Bottoms backs a temporary statewide moratorium in Georgia and an end to incentives; Pennsylvania’s Republican nominee proposed a pause, Shapiro a stricter framework; Morgan Stanley: 284 data centers, 33 GW at risk in TX, PA and GA; Sanders–Ocasio-Cortez moratorium bill; MOCA Economics on a Democratic Congress · <a href="https://www.bisnow.com/news/national/data-center-community-relations/what-the-midterms-mean-for-americas-ai-data-center-boom">Bisnow, Nov 1</a></li><li><strong>States</strong> — Democrats won both Georgia Public Service Commission seats on the ballot, a majority for the first time in decades · <a href="https://georgiarecorder.com/2026/11/04/democrats-gain-control-of-georgia-board-that-regulates-utilities-for-first-time-in-decades/">Georgia Recorder, Nov 4</a></li><li><strong>States</strong> — At least 14 Ohio municipalities passed bans or restrictions in unofficial results, including Sunbury (aimed at a possible Amazon project); 41 measures in 10 states; both Ohio governor nominees proposed moratoriums · <a href="https://fortune.com/2026/11/04/data-centers-abdul-el-sayed-michigan-incumbents-data/">Fortune, Nov 4</a></li><li><strong>Washington</strong> — Wyden’s draft: an excise tax on data centers, no bonus depreciation and no research credit for them · <a href="https://www.bdo.com/insights/tax/top-democratic-senate-taxwriter-targets-data-centers">BDO</a></li><li><strong>Washington</strong> — Ratepayer Protection Pledge, Mar 4: Amazon, Google, Meta, Microsoft, OpenAI, Oracle and xAI to build, bring or buy their own power and pay for grid upgrades · <a href="https://www.publicpower.org/node/29601">American Public Power Association</a></li><li><strong>Markets</strong> — Nasdaq-100 futures fell as much as 2.4% before recovering; Amazon, Meta, Microsoft higher, Nvidia +0.5%, Alphabet −0.3%, SpaceX and Tesla about −4%; IREN, Hut 8, Applied Digital, CoreWeave, Nebius lower (7:26 am ET) · <a href="https://www.kitco.com/news/off-the-wire/2026-11-04/us-stock-futures-rise-investors-assess-democrats-midterm-sweep">Reuters via Kitco, Nov 4</a></li><li><strong>Amazon</strong> — TTM OCF $168.3B, purchases of P&amp;E net $210.7B, FCF −$42.4B; Q3 OCF $42.4B, purchases $77.3B less $1.4B proceeds; nine months $175.7B less $3.5B; 2026 cash capex about $220B · <a href="https://ir.aboutamazon.com/news-release/news-release-details/2026/Amazon-com-Announces-Third-Quarter-Results/">Amazon Q3 2026 release</a></li><li><strong>Amazon</strong> — Expects to spend “meaningfully more” in 2027 than the roughly $220B of 2026 · <a href="https://alcapitaladvisory.com/research/intelligence/ai-infrastructure.html">AL Capital Advisory, on the Q3 call</a></li><li><strong>Alphabet</strong> — Google Cloud revenue up 103% to $30.7B, backlog $561B · <a href="https://www.insiderfinance.io/news/alphabet-q3-2026-earnings-miss-estimates">InsiderFinance, Oct 28</a></li><li><strong>Capacity</strong> — Alphabet says it remains supply constrained; Microsoft that demand exceeds available supply · <a href="https://io-fund.com/ai-stocks/ai-capex-1-trillion-estimates-too-low">I/O Fund</a></li><li><strong>Alphabet</strong> — Quarterly OCF and FCF, Q4 2025 to Q3 2026; TTM OCF $179.6B, FCF $19.7B · <a href="https://www.sec.gov/Archives/edgar/data/0001652044/000165204426000080/googexhibit991q32026.htm">Alphabet Q3 2026 8-K</a></li><li><strong>Alphabet</strong> — Q3 capex $51.4B; 2026 guide $195–205B; 2027 to increase significantly, analysts above $300B · <a href="https://www.cnbc.com/2026/10/28/alphabet-googl-earnings-q3-2026.html">CNBC, Oct 28</a></li><li><strong>Alphabet</strong> — About 65% of Q3 capex servers and networking, 35% data centers · <a href="https://finance.yahoo.com/markets/stocks/articles/alphabet-inc-goog-q3-2026-050953011.html">Q3 call highlights (Yahoo Finance)</a></li><li><strong>Meta</strong> — Q4 2025 OCF $36.21B, FCF $14.08B; capex incl. finance-lease principal $22.14B · <a href="https://www.sec.gov/Archives/edgar/data/1326801/000162828026003832/meta-12312025xexhibit991.htm">Meta Q4 2025 8-K</a></li><li><strong>Meta</strong> — Q1 2026 OCF $32.23B, FCF $12.39B; Q2 OCF $31.86B, FCF $0.784B · <a href="https://www.sec.gov/Archives/edgar/data/0001326801/000162828026050596/meta-06302026xexhibit991.htm">Meta Q1 and Q2 2026 8-Ks</a></li><li><strong>Meta</strong> — Q3 2026 OCF $32.79B, capex $34.91B, FCF −$2.12B; 2026 capex $135–145B · <a href="https://www.stocktitan.net/articles/meta-q3-2026-earnings-results">Stock Titan, on the Q3 release</a></li><li><strong>Meta</strong> — S&amp;P Global Ratings: outlook to negative Oct 30; 2027 capex over $225B; free operating cash flow −$22B (2026), −$71B (2027), −$39B (2028) · <a href="https://finance.yahoo.com/markets/stocks/articles/p-cuts-meta-outlook-negative-213940742.html">Yahoo Finance</a></li><li><strong>Microsoft</strong> — Jul–Sep 2026: OCF $58.0B, cash PP&amp;E $39.4B, FCF $18.6B, capex incl. finance leases $45.2B; FCF positive in FY27 · <a href="https://www.microsoft.com/en-us/investor/events/fy-2027/earnings-fy-2027-q1">Microsoft FY27 Q1 call</a></li><li><strong>Microsoft</strong> — Apr–Jun 2026: OCF $55.4B, cash PP&amp;E $35.8B, FCF $19.6B, capex $41B · <a href="https://windowsforum.com/news/microsoft-fy27-q1-earnings-preview-azure-growth-50b-capex-and-copilot-cash-flow-test.449563/">Windows Forum, on the FY26 Q4 call</a></li><li><strong>Microsoft</strong> — Jan–Mar 2026: OCF $46.7B, FCF $15.8B, cash PP&amp;E $30.9B, capex $31.9B; Oct–Dec 2025: OCF $35.8B, FCF $5.9B, cash PP&amp;E $29.9B, capex $37.5B · <a href="https://www.marketbeat.com/instant-alerts/microsoft-q3-earnings-call-highlights-2026-04-29/">MarketBeat call highlights; AlphaSense</a></li><li><strong>Microsoft</strong> — Calendar 2026 capex about $170B, cut from $175B · <a href="https://www.cnbc.com/2026/10/29/microsoft-google-ai-capex-estimates.html">CNBC, Oct 29</a></li><li><strong>Apple</strong> — Fiscal 2026 OCF $145.1B, capex $9.0B · <a href="https://www.gurufocus.com/news/9163152/apple-inc-reports-strong-q4-fy2026-results-revenue-surges-to-1145-billion-and-eps-reaches-210-both-beating-analysts-estimates">GuruFocus, on the 8-K</a></li><li><strong>Nvidia</strong> — Free cash flow $22.1B, $34.9B, $48.6B and $21.3B in fiscal Q3 2026 to Q2 2027; hyperscale revenue $48.7B in Q2 · <a href="https://quartr.com/events/nvidia-corporation-nvda-q2-2027_33dRoFp6">Stock Analysis; NVIDIA Q2 FY27 (Quartr)</a></li><li><strong>Tesla</strong> — OCF, capex and FCF by quarter, Q3 2025 to Q3 2026 · <a href="https://www.sec.gov/Archives/edgar/data/0001318605/000162828026066738/exhibit9914.htm">Tesla Q3 2026 update</a></li><li><strong>Tesla</strong> — 2026 capex guide cut to more than $20B, from more than $25B · <a href="https://electrek.co/2026/10/22/tesla-cuts-2026-capex-guidance-q3-cash-flow/">Electrek, Oct 22</a></li><li><strong>2027 capex</strong> — Consensus, early August: Alphabet $284.8B, Amazon $256.5B, Microsoft $207.6B (may be its fiscal year), Meta $185.6B · <a href="https://www.yieldtheory.app/research/hyperscaler-capex-2027-forecast">Yield Theory, on MarketScreener consensus via I/O Fund</a></li><li><strong>2027 capex</strong> — Street case: Alphabet $300B, Amazon $290B, Microsoft $215B, Meta $205B (Citi, Morgan Stanley) · <a href="https://06.ai/packs/hyperscaler-capex-funding">06.ai </a><a href="https://06.ai/packs/hyperscaler-capex-funding">hyperscaler</a><a href="https://06.ai/packs/hyperscaler-capex-funding"> pack</a></li><li><strong>2027 capex</strong> — JPMorgan: Amazon about $395B · <a href="https://huggingnews.com/earnings/jpmorgan-estimates-amazon-2027-capex-will-jump-77percent-to-395b-as-stoc-89aa42f0">HuggingNews, Nov 2</a></li><li><strong>2027 capex</strong> — S&amp;P Global six-company model, early September: Alphabet $357B · <a href="https://www.useluminix.com/reports/industry-analysis/hyperscaler-capex-q3-2026">Luminix</a></li><li><strong>Cost per GW</strong> — A GW costs about $50–60B; demand well above the ~70% growth supply supports in fiscal 2028 · <a href="https://www.webull.com/news/15497781742806016">NVIDIA Q2 FY27 call (via </a><a href="https://www.webull.com/news/15497781742806016">06.ai </a><a href="https://www.webull.com/news/15497781742806016">compute-balance</a><a href="https://www.webull.com/news/15497781742806016"> pack</a><a href="https://www.webull.com/news/15497781742806016">)</a></li><li><strong>Cost per GW</strong> — $38B per GW of IT power for a GB200 NVL72 site · <a href="https://epoch.ai/data-insights/ai-datacenter-cost-breakdown">Epoch AI</a></li><li><strong>Tax</strong> — 100% bonus depreciation made permanent for property acquired after Jan 19, 2025 · <a href="https://www.bdo.com/insights/tax/one-big-beautiful-bill-act-expands-100-depreciation-expensing-opportunities">BDO</a></li><li><strong>Leases</strong> — Leases signed but not yet commenced at Sep 30: Microsoft $364.6B, Meta $355.7B, Amazon $164.9B, Alphabet $130.5B · <a href="https://06.ai/wire/w007">06.ai W.007</a></li></ul>
<h2>Method</h2>
<p><strong>Free cash flow</strong> is each company’s own measure: operating cash flow less capex as each defines it. Amazon subtracts purchases of property and equipment net of proceeds and incentives; Meta includes principal payments on finance leases; Microsoft subtracts cash paid for property and equipment, and its 2027 figures here subtract capex including finance leases, the stricter measure. Trailing twelve months are as reported (Amazon, Alphabet) or summed from quarterly releases (Meta, Microsoft, Tesla, Nvidia). Fourth-quarter capex implied by guidance is the 2026 guide less January–September capex on the same basis.</p>
<p><strong>2027.</strong> Operating cash flow is the twelve months to September 30, 2026 grown at 10%, 20% or 30% a year for 1.25 years, the gap between that period’s midpoint and mid-2027 (Amazon’s grew 29% over the last year). Capex is one of three sets for calendar 2027: consensus as of early August (MarketScreener, via I/O Fund; Microsoft’s figure may be its June fiscal year); the Street case in the <a href="https://06.ai/packs/hyperscaler-capex-funding">hyperscaler</a> pack (Citi and Morgan Stanley estimates); and the highest estimates we found published, JPMorgan for Amazon, S&amp;P Global’s September model for Alphabet and S&amp;P Global Ratings for Meta (we found none above the Street for Microsoft). The share that has to slip is the 2027 gap divided by capex times 0.853: we take 70% of deferred capex as equipment, deducted at once under 100% bonus depreciation, at a 21% federal rate. State taxes would lower it slightly.</p>
<p><strong>What the vote pushes out.</strong> The gigawatts come from the <a href="https://06.ai/packs/texas-data-center-risk">Texas data center</a> pack’s model of Texas, which we re-ran without changes: new data center load switched on by the end of 2027 on Abbott’s path less the Democratic outcome (low: Hinojosa alone, moratorium to June 2027; base: with Goodwin, to September; stress: with the Texas House too, to the end of 2027). Self-powered and outside-ERCOT campuses are modeled one by one there, so each Big Four campus (Microsoft’s Abilene expansion and Pecos, Amazon’s GW Ranch, Alphabet’s Meitner, Meta’s El Paso) takes its own delay. Campuses waiting on ERCOT (Alphabet’s Armstrong and Wilbarger, Amazon’s Comanche Peak and Black Pearl lease, Meta’s Childress) take a share of the delay to ERCOT’s cleared and base load in proportion to their GW (1.9, 1.3 and 1.0 GW of 33.1 GW after ERCOT’s 49.8% realization), doubled in the stress case on the view that they are further along than the average. Anthropic’s campus, whose debt Google has agreed to guarantee, and capacity neoclouds lease to hyperscalers are not counted as Big Four capex. Each GW is priced at $40B, $50B or $60B (NVIDIA puts a gigawatt at $50–60B; Epoch AI, $38B per GW of IT power), of which 60%, 75% or 90% leaves 2027 when the switch-on moves, mostly servers bought shortly before the power comes on. Only the stress case adds Georgia and Ohio: their combined 12% of planned US GW against Texas’s 27%, at half Texas’s delay, half of it the Big Four’s. Pennsylvania is left out; its Democratic governor’s framework continues, and his opponent proposed the pause.</p>
<p><strong>Free cash flow added</strong> is capex pushed out times 0.853, less the cloud cash a delayed gigawatt doesn’t earn in 2027: $10B of revenue a gigawatt-year at a 50% cash margin for half a year. All of these are judgments, and the conclusion doesn’t hinge on them: in the default setting the base case pushes out 3%–13% of what Amazon, Alphabet or Meta would need to reach zero, and the stress case 13%–43%.</p>
<hr>
<p>Every chart above links to its interactive version on 06.ai, with the data behind it. The daily trackers are at <a href="https://06.ai/trackers">06.ai/trackers</a>.</p>
<p><em>Not investment advice. Estimates are labeled as estimates.</em></p>]]></content:encoded>
</item>
<item>
  <title>What does a blue Texas mean for the data center boom?</title>
  <link>https://inthesingularity.ai/packs/texas-data-center-risk</link>
  <guid isPermaLink="true">https://inthesingularity.ai/packs/texas-data-center-risk</guid>
  <pubDate>Wed, 04 Nov 2026 12:20:00 GMT</pubDate>
  <description>Texas froze its own data center pipeline before Democrats won the governorship and a Senate seat. We size what Hinojosa adds to Abbott’s freeze, how many GW that delays or cancels, and whose campuses carry it.</description>
  <category>Power &amp; politics</category>
  <enclosure url="https://inthesingularity.ai/newsletter/texas-data-center-risk/cover.png" length="62244" type="image/png" />
  <media:content url="https://inthesingularity.ai/newsletter/texas-data-center-risk/cover.png" medium="image" type="image/png" />
  <content:encoded><![CDATA[<figure><a href="https://06.ai/packs/texas-data-center-risk"><img src="https://06.ai/newsletter/texas-data-center-risk/thumb.png" alt="What does a blue Texas mean for the data center boom?" width="421" height="391"></a><figcaption><a href="https://06.ai/packs/texas-data-center-risk">Interactive version and data on 06.ai →</a></figcaption></figure>
<p>Democrats won Texas’s races for governor and US Senate on Tuesday: the AP called both for Gina Hinojosa and James Talarico, and Gov. Greg Abbott conceded. The freeze came first: since August, Abbott’s audit has held every large data center (75 MW or more) waiting to switch on in ERCOT, and since September his order to the state’s environmental agency has been read to cover the permits for the gas plants that let campuses skip the queue. In our base case Hinojosa leaves Texas with <strong>about 2.2 GW less new data center load switched on at the end of 2027 than Abbott’s path would have, and 2.9 GW less at the end of 2028</strong> (2.0–5.9 GW, depending on the races still being counted and how long a moratorium lasts). That is on top of the roughly 2.2 GW Abbott’s freeze already costs at each date. About half the gap is grid load in ERCOT’s Batch Zero arriving later; per GW, the hardest hit are self-powered campuses for Microsoft, Anthropic and OpenAI whose air permits are pending or still to come.</p>
<ul><li><strong>474 GW</strong> — Requested from ERCOT. About 90% data centers. Batch Zero’s preliminary sort set aside 302 GW for a later round</li><li><strong>6.6 GW</strong> — Cleared, frozen since Aug 3. 17 loads done with every step but the switch-on. Batch Zero’s base and studied groups hold 194 GW</li><li><strong>2.9 GW</strong> — Less switched on, end of 2028. Hinojosa vs Abbott’s path, base case. Range 2.0–5.9 GW</li><li><strong>6.7 GW</strong> — Most exposed. Self-powered campuses whose air permits are pending or still to come: Microsoft, Anthropic, OpenAI and Oracle, Fluidstack. Plant and IT capacity as reported</li></ul>
<h2>The results so far</h2>
<p>As reported late on Nov 3. Calls are the AP’s unless noted; shares are from the time shown in each row and will move as counting finishes.</p>
<p>US Senate · James Talarico<em>✓ </em><strong>59.7%</strong> · Ken Paxton <strong>38.8%</strong> · Called for Talarico by the AP late Tuesday; CBS and NBC also projected him. Shares are the AP count at about 52%, 9:35 pm ET</p>
<p>Governor · Gina Hinojosa<em>✓ </em><strong>~57%</strong> · Greg Abbott <strong>~41%</strong> · Called for Hinojosa by the AP; Abbott conceded. Shares at 66% of precincts, 11:03 pm ET (KPRC)</p>
<p>Lieutenant governor · Vikki Goodwin <strong>~59%</strong> · Dan Patrick <strong>~40%</strong> · No call we could confirm as of 11:52 pm ET; Goodwin declared victory and Patrick had not conceded (Texas Tribune). Shares from early returns</p>
<p>Texas House · Control not called as of 10:23 pm ET; 88–62 Republican going in. The state Senate keeps a Republican majority (Texas Tribune)</p>
<h2>Texas froze its own boom</h2>
<p>Hinojosa inherits a pipeline her Republican predecessor had already stopped. On August 3, Gov. Greg Abbott told the Public Utility Commission and ERCOT to audit every data center seeking a grid connection and to deny any that failed. ERCOT stopped approving data centers of 75 MW or more to switch on. Seventeen loads, about 6.6 GW, had cleared every step but that one; none has been approved since. ERCOT slates a cohort of 17 loads, which we take to be the same ones, for the second quarter of 2027. ERCOT also halted Batch Zero, the first round of its new process for large loads, which holds 194 GW in its base and studied groups.</p>
<p>On September 21 the freeze reached the workaround. Abbott told the Texas Commission on Environmental Quality to stop issuing permits for data centers until the audits finish, and law firms read that to include the on-site gas plants that let a campus skip ERCOT’s queue, but not permits already issued. Enforcement has been leaky: watchdogs counted air notices and stormwater approvals after the order; TCEQ later said the permits issued since then had been suspended, and shut its automatic e-filing system. Abbott kept the pause after President Trump called it a mistake. ERCOT files its audit on December 10, and the commission takes it up on December 17, five weeks before Hinojosa is sworn in.</p>
<h2>The freeze came first</h2>
<h2>Most of the rulebook is already agreed</h2>
<p>Abbott’s June package, Paxton’s Senate plan and the Democrats’ plans overlap more than the campaigns suggest. Abbott, Paxton, Talarico and Hinojosa would all repeal the data center sales tax exemption, which state projections put at $1.3 billion or more in 2026 (the comptroller counts 121 data centers using it). Abbott, Paxton and Talarico would also make data centers pay for their own power and grid connections, and Hinojosa would protect households from rate increases driven by data centers. Senators of both parties questioned the exemption at a July hearing; whether existing certifications would be honored was left open. Both sides’ plans put those rules before the 2027 session.</p>
<p>What Democrats add is narrower: a moratorium until laws pass (Hinojosa by executive order, Goodwin through a special session), a local veto over projects (Talarico, Hinojosa; Goodwin wants residents consulted), union labor on construction (Hinojosa) and, from the Senate seat, federal minimum standards (Talarico).</p>
<h2>Both parties want most of the same rules</h2>
<p>What each statewide candidate proposed for data centers, or did in office, as reported through October.  proposed or did it ·  partly, or a narrower version.</p>
<ul><li><strong>Repeal the data center sales tax exemption Patrick ordered a Senate study of its cost; Goodwin attacks the incentives and wants a special session on costs. Neither has said “repeal” in coverage we found</strong></li><li><strong>Data centers pay for their own power and grid connection Hinojosa would protect households from rate increases driven by data centers; Goodwin opposes households paying for grid buildout</strong></li><li><strong>Closed-loop cooling or water priority for residents Hinojosa’s plan targets nondisclosure of water use rather than cooling design; Patrick ordered a study of data centers’ water demands</strong></li><li><strong>Disclose power and water use As attorney general, Paxton opened an inquiry into developers that skipped the state’s water survey. Talarico would publish records of data centers’ tax breaks and include transparency in federal minimum standards; his plan as reported has no power or water reporting</strong></li><li><strong>Pause new approvals Abbott’s pause runs until the audits finish; Hinojosa’s and Goodwin’s until the Legislature passes protections. Paxton backs Abbott’s state-level proposals but hasn’t addressed the pause in coverage we found</strong></li><li><strong>Local power to approve or reject projects Abbott requires setbacks; Paxton backs limits near rural homes and rules on noise and lighting; Goodwin wants nearby residents consulted. None of the three proposes a local veto</strong></li><li><strong>Licensed, union labor on construction</strong></li><li><strong>Federal minimum standards for data centers</strong></li><li><strong>Ban Chinese technology in data centers</strong></li></ul>
<h2>What Hinojosa can actually do</h2>
<p>Less than the campaign suggests in 2027, and more by 2028. A governor can’t write statute: the session that opens January 12 needs the Legislature, and the Texas House went into the election 88–62 Republican. Her levers are directives and appointments. Abbott’s freeze is a set of letters, which a successor can keep, rewrite or rescind, but the agencies’ commissioners decide how to act on them, and she can’t remove the ones Abbott appointed: only the governor who made an appointment can, with two-thirds of the state Senate.</p>
<p>The utility commission turns over faster than that suggests. A new governor can name its chair as soon as she takes office. Two of its five commissioners appear to be Abbott recess appointees who need two-thirds of the Senate by May 31 or lose their seats, a third appears to be serving past an expired term, and two terms end on September 1, 2027. Recess appointments after the session could give Hinojosa a majority of the commission by late 2027 without a Senate vote, on one reading of the constitution that hasn’t been tested. That matters less for the 2027 freeze than for the rules that follow it: Batch Zero’s allocations, the next batch and who pays for transmission. At TCEQ, one of three seats turns over in August 2027.</p>
<p>How far a letter goes is already in question. ERCOT and the commission complied with Abbott’s, yet a former ERCOT market monitor doubts its requirements can be enforced without new laws, and both agencies asked lawmakers for more authority. A successor’s order would meet the same doubt, from agencies she didn’t appoint, and probably a lawsuit. That is why our paths assume only part of a Democratic moratorium holds, and more of it the more offices Democrats win.</p>
<h2>Most of the queue was never coming</h2>
<figure><a href="https://06.ai/packs/texas-data-center-risk#queue"><img src="https://06.ai/newsletter/texas-data-center-risk/fig-01.png" alt="Most of the queue was never coming" width="694" height="290"></a><figcaption>ERCOT’s large-load pipeline by stage, from its September Batch Zero classification and monthly reports, with the named self-powered campuses whose permits are pending. ERCOT’s stages are separate counts on different dates, so they don’t add up to a total.</figcaption></figure>
<p>The 474 GW is a list of requests, not a pipeline: more than five times ERCOT’s record peak of about 91 GW. ERCOT’s 2025 forecasting method counted new data center requests at 49.8%, and only about 46% of the large load it has approved since 2022 showed up at July’s peak. So the freeze matters less for the headline number than for the few dozen GW with studies, tenants and money behind them. We model those: the cleared loads, Batch Zero’s base load at ERCOT’s 49.8% realization, and the named self-powered and out-of-ERCOT campuses. That puts about 17.2 GW of new data center load on before the end of 2028 had nothing been frozen, and 14.9 GW on Abbott’s path.</p>
<h2>What Hinojosa adds to the delay</h2>
<p>New data center load switched on in Texas after September 2026, GW, before the freeze, on the path Abbott had set and under Hinojosa. Pick which other races Democrats win and when a moratorium ends; the base case is Hinojosa alone, with a Republican lieutenant governor and House, and a moratorium that ends in September 2027.</p>
<p>About <strong>2.2 GW</strong> less switched on by the end of 2027 than on Abbott’s path and <strong>2.9 GW</strong> less by the end of 2028; about 1.0 GW never built by 2030. Grid approvals wait about 3 months longer, own-gas permits about 4 months.</p>
<figure><a href="https://06.ai/packs/texas-data-center-risk#paths"><img src="https://06.ai/newsletter/texas-data-center-risk/fig-02.png" alt="What Hinojosa adds to the delay" width="694" height="422"></a></figure>
<figure><a href="https://06.ai/packs/texas-data-center-risk#paths"><img src="https://06.ai/newsletter/texas-data-center-risk/fig-03.png" alt="What Hinojosa adds to the delay — June 2027" width="694" height="453"></a><figcaption><strong>June 2027</strong>: About 1.5 GW less switched on by the end of 2027 than on Abbott’s path and 2.0 GW less by the end of 2028; about 1.0 GW never built by 2030. Grid approvals wait about 2 months longer, own-gas permits about 3 months.</figcaption></figure>
<figure><a href="https://06.ai/packs/texas-data-center-risk#paths"><img src="https://06.ai/newsletter/texas-data-center-risk/fig-04.png" alt="What Hinojosa adds to the delay — End of 2027" width="694" height="453"></a><figcaption><strong>End of 2027</strong>: About 2.4 GW less switched on by the end of 2027 than on Abbott’s path and 4.0 GW less by the end of 2028; about 1.0 GW never built by 2030. Grid approvals wait about 5 months longer, own-gas permits about 7 months.</figcaption></figure>
<p>On Abbott’s path the commission acts on the audit in December, switch-ons and permits resume early in 2027, and the 2027 session writes his package into law. Texas adds about 5.8 GW of new data center load by the end of 2027 and 14.9 GW by the end of 2028. With Hinojosa sworn in on January 19, a moratorium on new data centers runs until the Legislature passes protections she signs. In the base case approvals wait about 3 months longer and own-gas permits about 4, which leaves Texas 2.2 GW short at the end of 2027 and 2.9 GW short at the end of 2028, with about 1.0 GW never built by 2030. If Goodwin’s lead for lieutenant governor holds, the gap at the end of 2028 is about 3.8 GW.</p>
<p>The range runs from 2.0 GW at the end of 2028, if the moratorium ends when the session does, to 5.9 GW if Democrats also win the lieutenant governor’s office and the Texas House and the moratorium lasts to the end of 2027. Most of the gap is time, not cancellations: the GW come later. The hardest hit are self-powered campuses, where the governor’s environmental agency is the gate: on Abbott’s path about 2.8 GW of them is running by the end of 2028; in the base case, 1.6 GW.</p>
<h2>Whose campuses carry the risk</h2>
<figure><a href="https://06.ai/packs/texas-data-center-risk#companies"><img src="https://06.ai/newsletter/texas-data-center-risk/fig-05.png" alt="Whose campuses carry the risk" width="694" height="538"></a><figcaption>Texas capacity by company, split by what it is waiting on; Batch Zero base load, whose required studies are approved, counts as approved. GW are near-term phases as each project reports them (IT, facility or on-site generation), so the bars compare exposure, not exact load. Hover or tap a bar for the projects.</figcaption></figure>
<h2>Self-powered campuses are the most exposed</h2>
<p>The irony of the freeze is that the campuses built to avoid the grid are the ones a governor reaches most directly. ERCOT’s pause runs through a utility commission whose members a new governor can replace only as their seats come open; TCEQ’s permits are where a governor’s letter has already worked. Five campuses were waiting on air permits, or will need them:</p>
<ul><li><strong>Anthropic’s Hill County campus</strong> (Nexus): a 1.6 GW gas plant, with about $15 billion of senior debt that Google agreed to guarantee, per Bloomberg; its major-source permit drew a hearing request in October.</li><li><strong>Microsoft’s</strong> 900 MW Abilene expansion with Crusoe, first building due mid-2027, where a pending permit for added generation appears to cover its plant (our reading of a TCEQ notice that names Crusoe), and its Pecos campus with Chevron, about 2 GW from 2028, where a minor-source authorization is in review and we found no filing yet for the main air permit.</li><li><strong>OpenAI and Oracle’s Frontier</strong> in Shackelford County, where VoltaGrid’s major-source air permit for the site is pending; its first engines were approved in 2025.</li><li><strong>Fluidstack’s Cameron County</strong> site, up to 1.5 GW with on-site generation that will need permits, in the county with the most Batch Zero base and studied load; its tenant isn’t named.</li></ul>
<p>Grid-tied campuses carry timing risk, not existential risk. Some hold ERCOT approvals or base-load status: Galaxy’s Helios for CoreWeave, IREN’s Sweetwater, Cipher’s Colchis and Crusoe’s Childress site for Meta. Others’ ERCOT status isn’t public: Cipher’s Barber Lake, Google’s Armstrong and Wilbarger campuses and Amazon’s Comanche Peak site, which Vistra says the pause doesn’t affect. Most of their dates hinge on December 17. Amazon’s GW Ranch already holds its 7.65 GW air permit, which makes it one of the least exposed large projects in the state. Fermi’s risk is its own: its only announced lease isn’t in force.</p>
<p>A longer freeze also has winners. Every megawatt already energized in Texas becomes scarcer: IREN’s Childress site, Cipher’s Black Pearl for AWS and Core Scientific’s Denton campus could command better terms on new or renewed contracts as new supply slips, and capacity outside Texas gets a political-risk premium it didn’t have a year ago.</p>
<h2>The projects</h2>
<p>Named Texas campuses of about 200 MW or more, what each is waiting on and our read of its exposure to Hinojosa relative to the path Abbott had set. Capacity is as each project reports it.</p>
<ul><li><strong>Nexus Hubbard Hill County</strong> — Nexus; Anthropic tenant; Google agreed to guarantee about $15B of senior debt (Bloomberg) GOOGL · 1.6 GW gas plant (build-out 6–8 GW) · ERCOT area, own gas plant · Major-source air permit: comments closed Oct 23, hearing requested; paused · Exposure: High</li><li><strong>Frontier (Stargate) Shackelford County</strong> — Vantage, VoltaGrid; Oracle for OpenAI ORCL · 1.4 GW IT · Off-grid gas · VoltaGrid major-source air permit (notice Sep 22) paused; first engines approved in 2025 · Exposure: High</li><li><strong>Abilene expansion (Microsoft) Taylor County</strong> — Crusoe; Microsoft MSFT · 0.9 GW (672 MW IT) · Own 900 MW plant · PSD air permit for added generation under review, paused (our attribution) · Exposure: High</li><li><strong>Pecos campus and Project Kilby Reeves County</strong> — Chevron, Engine No. 1; Microsoft 20-year PPA MSFT, CVX · ~2 GW DC (2.67 GW gas) · Islanded gas · Final investment decision by end-2026; a minor-source authorization in review, no main air permit filing found · Exposure: High</li><li><strong>Cameron County campus Cameron County</strong> — Fluidstack; tenant not named · Up to 1.5 GW · ERCOT plus own generation · ERCOT switch-on and permits for on-site generation; no permit filing found · Exposure: Medium</li><li><strong>Milam County (Stargate) Milam County</strong> — SB Energy (SoftBank); OpenAI 1.2 GW lease 9984.T · 1.2 GW · ERCOT plus new generation · ERCOT switch-on (inferred); building 1 was due Oct 2026 · Exposure: Medium</li><li><strong>Childress “Project Hyper” Childress County</strong> — Crusoe, Lancium; Meta (reported) META · 1.0 GW · ERCOT · 1 GW interconnection approved; switch-ons after the audit · Exposure: Medium</li><li><strong>Armstrong County (Goodnight) Armstrong County</strong> — Crusoe; Google GOOGL · ~1.07 GW (secondary) · ERCOT plus co-located wind and on-site gas · Switch-ons for buildings due 2027–28; a 525 MW curtailment obligation behind one wind farm; 933 MW of on-site gas turbines reported under construction (secondary) · Exposure: Medium</li><li><strong>Wilbarger County “Project Rise” Wilbarger County</strong> — Google GOOGL · ~0.85 GW (secondary) · ERCOT plus AES generation · Switch-on (likely in the paused group) · Exposure: Medium</li><li><strong>Comanche Peak campus Somervell County</strong> — Amazon; Vistra 20-year PPA AMZN, VST · 1.2 GW PPA · ERCOT, nuclear PPA · Energize 2027; Vistra says the pause doesn’t affect it · Exposure: Medium</li><li><strong>Helios Dickens County</strong> — Galaxy; CoreWeave leases Phases I–III GLXY, CRWV · 1.6 GW approved; 2.6 GW more in study · ERCOT · Phase II (260 MW IT) from Q2 2027 and Helios II (830 MW, no tenant) need switch-ons; 2.6 GW in study · Exposure: Medium</li><li><strong>Barber Lake Mitchell County</strong> — Cipher; Fluidstack with a Google backstop, then an AI lab CIFR, GOOGL · 0.3 GW (207 MW IT) · ERCOT · Halls due Q4 2026–Q1 2027 need switch-ons · Exposure: Medium</li><li><strong>Colchis and Cipher’s other sites West Texas</strong> — Cipher; no tenant named CIFR · 1 GW base + 1.6 GW in study · ERCOT (AEP) · Batch Zero base load (Colchis, 2028) and studied load; plans for up to 2.5 GW of on-site gas would need TCEQ · Exposure: Medium</li><li><strong>Sweetwater Nolan County</strong> — IREN; no tenant named IREN · 2 GW (1.4 GW substation energized) · ERCOT · Batch Zero base load awaiting verification; first 300 MW due Q4 2027 · Exposure: Medium</li><li><strong>Beacon Point Nueces County</strong> — Hut 8; investment-grade tenant not named (704 MW IT leased) HUT · 1 GW campus · ERCOT · Batch Zero base load (secondary source); phase 1 timing undisclosed, phase 2’s first hall Q2 2028 · Exposure: Medium</li><li><strong>El Paso campus El Paso County</strong> — Meta; El Paso Electric’s 366 MW McCloud plant META · ~1 GW · WECC (El Paso Electric) · PUCT vote on McCloud; judges back it only if Meta bears the cost · Exposure: Medium</li><li><strong>San Antonio off-grid pair Bexar County</strong> — Vantage, VoltaGrid; tenant not named · 0.42 GW gas · Off-grid gas · Minor-source permits issued; notice of intent to sue over source splitting · Exposure: Medium</li><li><strong>CloudBurst San Marcos Guadalupe County</strong> — CloudBurst; tenant not named ET · 1.2 GW · ERCOT area, Energy Transfer gas · Grid or own-gas approvals, mix not disclosed · Exposure: Medium</li><li><strong>Caldwell Valley Caldwell County</strong> — Tract; tenant not named · Up to 4 GW; 360 MW first · ERCOT · Later ERCOT batches; a gas plant would need TCEQ · Exposure: Medium</li><li><strong>Stargate Abilene Taylor County</strong> — Crusoe, Lancium; Oracle for OpenAI ORCL · 1.2 GW · ERCOT · Last buildings due by end-2026; a turbine amendment is paused · Exposure: Low</li><li><strong>GW Ranch Pecos County</strong> — Amazon; Pacifico Energy AMZN · 1 GW phase 1 (7.65 GW gas permitted) · Off-grid gas, grid later · Air permit issued Jan 2026; a later grid tie would be exposed · Exposure: Low</li><li><strong>Childress (Horizon) Childress County</strong> — IREN; Microsoft $9.7B contract, Nvidia $3.4B IREN, MSFT · 0.75 GW grid connection · ERCOT · Energized site; AI halls through 2027 · Exposure: Low</li><li><strong>Black Pearl and Stingray Winkler and Andrews counties</strong> — Cipher; AWS leases CIFR, AMZN · 0.3 GW + 0.1 GW · ERCOT · Black Pearl energized and ramping; Stingray’s shell due in Q4 (Cipher) · Exposure: Low</li><li><strong>Rockdale Milam County</strong> — Riot; Anthropic (per Bloomberg) 20-year lease RIOT · 191 MW IT · ERCOT · Built on an energized mining site; 96 MW Dec 2027, 95 MW Jun 2028 (secondary) · Exposure: Low</li><li><strong>Denton Denton County</strong> — Core Scientific; CoreWeave CORZ, CRWV · ~0.36 GW · ERCOT · Operating · Exposure: Low</li><li><strong>Meitner Gray and Roberts counties</strong> — Google, Intersect GOOGL · ~0.84 GW (secondary) · Panhandle, own gas firming · Gas permit appears issued in May · Exposure: Low</li><li><strong>Project Matador Carson County</strong> — Fermi; TensorWave lease not yet effective FRMI, AMD · 2 GW target by end-2027 (11 GW plan) · SPP plus own gas · Lease financing; a further 5 GW air permit was filed in March, status not found · Exposure: Company risk</li><li><strong>Project Horizon Pecos County</strong> — Poolside; CoreWeave exited CRWV · 2 GW · Own gas (planned) · No anchor tenant since April · Exposure: Stalled</li><li><strong>Comanche Circle Hood County</strong> — Sailfish; no tenant · Up to 5 GW · ERCOT plus planned gas · County rejected its settlement Oct 13; NRG withdrew the gas plant · Exposure: Stalled</li></ul>
<h2>The Senate seat matters less, for now</h2>
<p>Democrats needed a net gain of four seats for a Senate majority, and NBC and CBS project they got it, with Talarico’s seat among the gains; NBC also projects a Democratic House. Even with both chambers, bills face a Republican president’s veto until 2029, so Talarico’s federal minimum standards for data centers are a long shot. What changes sooner is confirmations, including FERC and NRC commissioners, and the ability to block a federal bill preempting state AI or data center rules.</p>
<p>ERCOT sits largely outside FERC’s wholesale jurisdiction, so Washington’s reach into the Texas buildout runs mainly through much of the Panhandle and El Paso, which belong to the SPP and WECC grids. For the campuses in this pack the governor matters far more than the senator.</p>
<h2>Democratic governors have gone both ways</h2>
<p>Facing the same backlash, two Democratic governors made opposite calls this year. In July, New York’s Kathy Hochul ordered a pause on new permits for data centers of 50 MW or more, for about a year while the state writes an environmental review process. In April, Maine’s Janet Mills vetoed a moratorium her Legislature had passed, saying she would sign a version that exempted one project. Hinojosa will be choosing between those two models, with a Republican president who has already criticized Abbott’s pause.</p>
<h2>What to watch</h2>
<p><strong>The rest of the count.</strong> The lieutenant governor’s race decides who runs the state Senate’s agenda, though the chamber stays Republican; the Texas House, 88–62 Republican going in, decides whether a local veto or labor rules can pass.</p>
<p><strong>December 10 and 17.</strong> ERCOT’s audit reports, then the commission’s open meeting on them, with Abbott still governor after conceding. Approvals granted before January 19 are hard to reverse, so what the commission clears in December is largely out of a new governor’s reach.</p>
<p><strong>January 19.</strong> Whether Hinojosa signs a moratorium order, what it covers (new projects, or anything not yet switched on) and whether TCEQ and the commission follow it.</p>
<p><strong>December 31.</strong> The commission’s deadline for its transmission-cost rule, which as proposed bills large loads on 12 monthly peaks instead of 4 and sets a 20-year minimum bill. The current commission writes it, before Hinojosa takes office.</p>
<p><strong>The exemption bill.</strong> Abbott, Hinojosa and both Senate candidates back repeal; grandfathering decides whether it changes the economics of campuses already certified.</p>
<p><strong>Company calls.</strong> Oncor and IREN report on November 5, Fermi next week. Watch for Batch Zero commitments, collateral withdrawals and anything self-powered campuses say about permits.</p>
<h2>What this is anchored on</h2>
<p>Results, filings, permits and statements are as reported; the paths, exposure ratings and the moratorium’s effect are ours.</p>
<ul><li><strong>Results</strong> — The AP called the Senate race for Talarico late Tuesday · <a href="https://time.com/article/2026/11/04/talarico-wins-texas-senate-race-first-democrat-to-win-statewide-in-more-than-30-years/">TIME, Nov 4</a></li><li><strong>Results</strong> — CBS News projects Talarico the winner · <a href="https://www.cbsnews.com/news/texas-2026-senate-james-talarico-ken-paxton/">CBS News, Nov 3</a></li><li><strong>Results</strong> — Talarico 59.7%, Paxton 38.8% with about 52% of the expected vote counted (AP count, 9:35 pm ET) · <a href="https://victoriaadvocate.com/2026/11/03/talarico-overtakes-paxton-in-texas-senate-race/">Victoria Advocate (AP data via the Texas Tribune), Nov 3</a></li><li><strong>Results</strong> — Hinojosa wins per the AP race call; 57% to 41% with 66.1% of precincts reporting · <a href="https://www.click2houston.com/decision-2026/2026/11/04/gina-hinojosa-wins-texas-governors-race-first-democrat-in-36-years/">KPRC Click2Houston, Nov 3</a></li><li><strong>Results</strong> — Abbott conceded the governor’s race to Hinojosa · <a href="https://abc7amarillo.com/news/local/gov-greg-abbott-to-watch-election-day-results-in-austin">ABC 7 Amarillo, Nov 3</a></li><li><strong>Results</strong> — Goodwin declared victory; Patrick had not conceded; the state Senate keeps a Republican majority · <a href="https://www.texastribune.org/2026/11/03/texas-dan-patrick-vikki-goodwin-lt-governor-election-2026/">Texas Tribune, Nov 3, 10:52 pm CT</a></li><li><strong>Results</strong> — Early lieutenant governor shares (AP) · <a href="https://www.kut.org/texas/2026-11-03/texas-nov-midterm-statewide-election-results">KUT, Nov 3</a></li><li><strong>Results</strong> — None of the most contested Texas House races called; one Democratic pickup in Corpus Christi · <a href="https://www.texastribune.org/2026/11/03/texas-house-legislature-outcomes-control-republicans-democrats-2026/">Texas Tribune, Nov 3, 9:23 pm CT</a></li><li><strong>Results</strong> — Democrats win the Senate majority, CBS News projects · <a href="https://www.cbsnews.com/news/senate-control-democrats-republicans-2026-midterm-elections/">CBS News, Nov 3</a></li><li><strong>Results</strong> — Democrats win control of the House and Senate, NBC News projects · <a href="https://www.nbcnews.com/politics/2026-election/live-blog/midterm-election-results-2026-news-trump-live-updates-rcna605763">NBC News live blog, Nov 3</a></li><li><strong>Abbott</strong> — June 10 package for 2027: repeal the sales tax exemption; data centers add generating capacity, pay for interconnection, use water-efficient cooling such as closed loops, report power and water · <a href="https://www.texastribune.org/2026/06/10/texas-greg-abbott-data-centers-regulation-sales-tax/">Texas Tribune, Jun 10</a></li><li><strong>Abbott</strong> — Aug 3 directive: PUCT and ERCOT audit every data center in the ERCOT queue; failures denied; more than 474 GW requested, about 90% data centers · <a href="https://gov.texas.gov/news/post/governor-abbott-directs-comprehensive-data-center-audit">Office of the Governor, Aug 3</a></li><li><strong>Abbott</strong> — Sep 21 directive: TCEQ to halt data center permits until the audits finish; conditions include covering grid costs, lower residential bills, water and setbacks · <a href="https://gov.texas.gov/news/post/governor-abbott-directs-tceq-to-halt-data-center-permits">Office of the Governor, Sep 21</a></li><li><strong>TCEQ</strong> — Paused permits for infrastructure directly supporting data centers; screening pending applications · <a href="https://www.tceq.texas.gov/downloads/response/tceq-response-gov-abbott-dc-directive.pdf">TCEQ letter to the governor, Oct 19</a></li><li><strong>TCEQ</strong> — Air notices, stormwater coverage and generator reviews kept moving after Sep 21 · <a href="https://fractracker.org/2026/10/texas-data-center-permits-after-governor-halt/">FracTracker Alliance, Oct</a></li><li><strong>TCEQ</strong> — TCEQ said permits issued after the order had been suspended, and shut its automatic e-filing system · <a href="https://www.tpr.org/text/economy-and-labor/2026-10-22/despite-abbott-moratorium-texas-kept-on-issuing-data-center-permits-now-the-state-agency-is-shutting-down-its-automatic-filing-system">Texas Public Radio (San Antonio Report), Oct 22</a></li><li><strong>TCEQ</strong> — Reads the directive to reach co-located and behind-the-meter generation serving data centers; implies issued permits stand · <a href="https://www.bracewell.com/resources/governor-abbott-orders-tceq-to-pause-all-permit-issuance-for-data-center-projects/">Bracewell, Sep 23</a></li><li><strong>Abbott, Paxton</strong> — Sep 14: Abbott directs the Water Development Board to enforce its water-use survey and refer violators to prosecutors; Sep 24: the attorney general opens an investigation of developers in 18 counties · <a href="https://www.willkie.com/publications/2026/10/texas-governor-halts-data-center-permits-pending-ercot-audit">Willkie Farr &amp; Gallagher, Oct</a></li><li><strong>Trump</strong> — Called Abbott’s moratorium a mistake; Abbott kept it · <a href="https://cbsaustin.com/news/local/trump-criticizes-abbott-moratorium-on-texas-data-centers-connecting-to-power-grid">CBS Austin, Aug 10</a></li><li><strong>Talarico</strong> — Plan: repeal the exemption, large users pay their own grid costs, closed-loop cooling, local approval or rejection, federal minimum standards · <a href="https://www.texastribune.org/2026/07/22/texas-james-talarico-data-centers-regulation-senate-2026/">Texas Tribune, Jul 22</a></li><li><strong>Paxton</strong> — “Texas First” plan: backs Abbott’s state-level proposals and repealing the exemption; own power, closed-loop cooling, limits near rural homes, China ban · <a href="https://www.fox7austin.com/election/james-talarcio-ken-paxton-texas-data-center-plan">Fox 7 Austin, Sep 30</a></li><li><strong>Hinojosa</strong> — Executive order halting new data centers until the Legislature passes protections; special session to end the tax breaks · <a href="https://ginafortexas.com/priority/data-centers/">Hinojosa campaign</a></li><li><strong>Hinojosa</strong> — July 23 call: union labor, local power to reject projects, ban on NDAs hiding water and power use, ratepayer protection, transparency · <a href="https://ginafortexas.com/2026/07/in-the-news-hinojosa-calls-for-data-center-moratorium-until-abbott-convenes-legislature/">Hinojosa campaign, citing press</a></li><li><strong>Goodwin</strong> — Pause new approvals until safeguards; water priority for families and farmers; disclosure; residents’ role in decisions · <a href="https://www.galvnews.com/news_reuters/texas/democratic-candidates-seek-pause-on-texas-data-center-growth/article_5600b5fd-624a-5fb6-ab8b-63057910da8b.html">Galveston County Daily News, Jul 21</a></li><li><strong>Patrick</strong> — Interim charges: data centers’ water and power demands and the cost of the sales tax exemption · <a href="https://www.texastribune.org/2026/03/27/dan-patrick-texas-senate-priorities-data-centers-thc-prediction-markets/">Texas Tribune, Mar 27</a></li><li><strong>Senate</strong> — Democrats need to net four seats to take control of the Senate · <a href="https://www.cbsnews.com/live-updates/senate-election-results-2026/">CBS News live results, Nov 3</a></li><li><strong>Exemption</strong> — At least $1.3B in 2026 (state projections); about $3.2B over the next two years, for 121 data centers (comptroller) · <a href="https://www.texastribune.org/2026/04/08/texas-data-centers-sales-tax-break-billion-dollars/">Texas Tribune, Apr 8</a></li><li><strong>PUCT</strong> — Johnson appointed Oct 23, 2025 to a term ending Sep 1, 2031 · <a href="https://gov.texas.gov/news/post/governor-abbott-appoints-johnson-to-public-utility-commission-of-texas">Office of the Governor, Oct 23, 2025</a></li><li><strong>PUCT</strong> — Rhode appointed Mar 12, 2026 to a term ending Sep 1, 2027 · <a href="https://gov.texas.gov/news/post/governor-abbott-appoints-rhode-to-public-utility-commission-of-texas">Office of the Governor, Mar 12</a></li><li><strong>PUCT, TCEQ</strong> — Senate nominations of May 21, 2025: Gleeson to 2029, Hjaltman to Sep 1, 2025; TCEQ’s Paup to Aug 31, 2027 and Gonzales to 2029 · <a href="https://capitol.texas.gov/tlodocs/89R/calendars/pdf/SA20250521.pdf">Texas Senate agenda, May 21, 2025</a></li><li><strong>Constitution</strong> — Appointments need two-thirds of senators present; unconfirmed recess appointees are treated as rejected at the end of a regular session · <a href="https://texaslegalguide.com/Texas_Constitution:Article_IV,_Section_12">Tex. Const. art. IV §12 (Texas Legal Guide)</a></li><li><strong>Constitution</strong> — Only the governor who appointed an officer may remove them, with two-thirds of the Senate · <a href="https://texaslegalguide.com/Texas_Constitution:Article_XV,_Section_9">Tex. Const. art. XV §9 (Texas Legal Guide)</a></li><li><strong>PUCT</strong> — Proposed transmission-cost rule: 12CP replaces 4CP; large loads pay a 20-year minimum billing demand; SB 6 deadline Dec 31 · <a href="https://www.ercot.com/files/docs/2026/07/23/PUCT-Project-58000-pfp_adopted-9-July-2026.pdf">PUCT Project 58000, Jul 9</a></li><li><strong>PUCT</strong> — ERCOT and the commission asked the Legislature for direct curtailment and registration authority over large loads · <a href="https://kvia.com/wp-content/uploads/2026/07/58317_18_1666954.pdf">PUCT response to the governor, Jul 17</a></li><li><strong>Elsewhere</strong> — New York’s EO 62 pauses permits for data centers of 50 MW or more for about a year; Maine’s governor vetoed a moratorium in April · <a href="https://builtin.com/articles/state-data-center-moratoriums">Built In, Aug 4</a></li><li><strong>Exemption</strong> — Senators of both parties questioned it at a Jul 27 hearing; honoring existing certifications left open · <a href="https://www.hklaw.com/en/insights/publications/2026/07/texas-data-center-tax-exemption-faces-serious-legislative-risk">Holland &amp; Knight, Jul 30</a></li><li><strong>Enforceability</strong> — Former ERCOT market monitor Beth Garza doubts the directive can be enforced without new laws · <a href="https://www.tpr.org/government-politics/2026-08-05/texas-governors-data-center-moratorium-prompts-calls-for-special-legislative-session">Texas Public Radio, Aug 5</a></li><li><strong>ERCOT</strong> — 17 data center and crypto loads, 6,608 MW, done with every step but approval to energize; a Nov 1 cohort of 17 loads, ~6,874 MW, for Q2 2027; reports Dec 10, PUCT open meeting Dec 17; loads of 25–75 MW not paused · <a href="https://www.ercot.com/files/docs/2026/08/20/Batch-Zero-Verification-and-Audit-PUCT-Presentation.pdf">ERCOT presentation to the PUCT, Aug 20</a></li><li><strong>ERCOT</strong> — Batch Zero conditional classes: base 66.4 GW (204 projects), studied 125.4–127.9 GW (158), excluded 302.2 GW (373) · <a href="https://www.ercot.com/files/docs/2026/09/11/14-Batch-Zero-Update.pdf">ERCOT Board, Item 14, Sep 11</a></li><li><strong>ERCOT</strong> — Batch Zero study restarts in December or January; the adjusted load forecast moves to December · <a href="https://www.ercot.com/files/docs/2026/09/11/13-CEO-Update.pdf">ERCOT Board, CEO update, Sep 11</a></li><li><strong>ERCOT</strong> — Large loads of all kinds: 9,456 MW approved to energize since 2022; 4,370 MW drawn at July’s peak · <a href="https://www.ercot.com/files/docs/2026/08/17/ERCOT-Monthly-Operational-Overview-July-2026.pdf">ERCOT monthly operational overview, July (Aug 17)</a></li><li><strong>ERCOT</strong> — Approved large load still 9,456 MW in September · <a href="https://www.ercot.com/files/docs/2026/10/16/ERCOT-Monthly-Operational-Overview-September.pdf">ERCOT monthly operational overview, September (Oct 16)</a></li><li><strong>ERCOT</strong> — Batch One, for requests outside Batch Zero, targeted to launch in June 2027 · <a href="https://www.datacenterfrontier.com/energy/article/55401760/ercot-puts-texas-ai-megawatts-to-the-test">Data Center Frontier, Aug 31</a></li><li><strong>ERCOT</strong> — The 2025 forecast method counted new data center requests at 49.8%; average in-service slip about 220 days · <a href="https://www.ercot.com/files/docs/2025/04/07/8.1-Long-Term-Load-Forecast-Update-2025-2031-and-Methodology-Changes.pdf">ERCOT Board, Item 8.1, Apr 7, 2025</a></li><li><strong>ERCOT</strong> — Preliminary 2026 forecast: 39.7 GW of new data center load in 2027 rising to 211.6 GW in 2031, from utility-submitted requests; ERCOT calls it preliminary · <a href="https://www.ercot.com/files/docs/2026/04/15/12-CEO-Update.pdf">ERCOT Board, CEO update, Apr 15</a></li><li><strong>ERCOT</strong> — Preliminary all-time peak demand of about 91.1 GW on Jul 22, 2026 · <a href="https://www.forbes.com/sites/daraabasiita/2026/10/13/texas-orders-data-center-audit-as-grid-requests-reach-474-gigawatts/">Forbes, Oct 13</a></li><li><strong>PUCT</strong> — 16 TAC §25.194 adopted Sep 18: $50,000 per MW of security, a $100,000 initial fee plus tiered study charges, for loads of 75 MW or more · <a href="https://www.whitecase.com/insight-alert/puct-sets-financial-commitments-data-center-interconnection-governor-pauses-permit">White &amp; Case, Sep 25</a></li><li><strong>Outside view</strong> — BNEF: the pause could delay 49.8 GW of data center load; about 8.25 GW of additions in ERCOT through 2030 · <a href="https://www.powermag.com/texas-audit-could-delay-49-8-gw-of-data-center-load-cost-projects-up-to-15-billion-bnef-warns/">POWER, Aug 6</a></li><li><strong>Outside view</strong> — CenterPoint expects 3, 8, 11 and 12 GW of Batch Zero load energized in 2027–30; about 2 GW of studied load withdrawn · <a href="https://www.spglobal.com/energy/en/news-research/latest-news/electric-power/102726-centerpoint-energy-boosts-capital-spending-plan-to-accommodate-large-load-growth">S&amp;P Global, Oct 27</a></li><li><strong>Outside view</strong> — 92 GW of advanced-planning data center load in ERCOT: 11 operating or expanding, 12 under construction, 15 committed · <a href="https://modoenergy.com/research/en/ercot-data-center-buildout-forecast-october-2026">Modo Energy, Oct 21</a></li><li><strong>Curtailment</strong> — 525.5 MW of uncompensated emergency curtailment on two Crusoe data centers behind one wind farm (Docket 59220); 933 MW of off-grid gas turbines under construction at the Armstrong County site (no source given) · <a href="https://www.globaldatacenterhub.com/p/ercot-data-center-interconnection-pause-docket-59220">Global Data Center Hub, Aug 26</a></li><li><strong>Nexus</strong> — Anthropic campus, 1.6 GW on-site gas plant; about $15B of financing in talks, with Google guaranteeing Anthropic’s lease and power payments (WSJ’s account) · <a href="https://www.globaldatacenterhub.com/p/morgan-stanley-leads-15-billion-financing">Global Data Center Hub (WSJ), Aug 3</a></li><li><strong>Nexus</strong> — Google agreed to guarantee the senior debt; $1.3B mezzanine loan closed; final investment decision reached · <a href="https://thenextweb.com/news/eagle-point-nexus-anthropic-texas-data-center-loan">The Next Web (Bloomberg), Aug 20</a></li><li><strong>Nexus, Frontier</strong> — Major-source air permit notices mailed Sep 22 for Nexus Hubbard and VoltaGrid’s Frontier plant · <a href="https://www.citizen.org/wp-content/uploads/Summary_-Data-Center-Permits-Issued-by-TCEQ-After-9_21_2026.pdf">Public Citizen permit summary</a></li><li><strong>Nexus</strong> — State building filings: HB1 to Oct 2027, HB2 to Jan 2028 · <a href="https://www.datacenterdynamics.com/en/news/nexus-dc-files-to-develop-second-data-center-outside-dallas-texas/">DCD</a></li><li><strong>Frontier</strong> — 1.4 GW of critical IT in 10 buildings; first building 2H 2026 · <a href="https://vantage-dc.com/data-center-locations/north-america/shackelford-county-tx">Vantage Data Centers</a></li><li><strong>Frontier</strong> — About 700 MW from 210 gas engines, approved by TCEQ per filings · <a href="https://www.constructionowners.com/news/oracle-and-openai-to-power-new-texas-data-center-off-the-grid">Construction Owners (Business Insider), Oct 24, 2025</a></li><li><strong>Frontier</strong> — Projected building dates Aug 2027 to Nov 2028 (estimate) · <a href="https://epoch.ai/data/ai-data-centers/directory/openai-stargate-shackelford">Epoch AI</a></li><li><strong>Abilene</strong> — 900 MW campus for Microsoft with its own power plant; first building mid-2027; the 1.2 GW Stargate campus complete by end-2026 · <a href="https://www.crusoe.ai/resources/newsroom/crusoe-announces-new-900-mw-ai-factory-campus-in-abilene-texas-to-support-microsoft-ai-infrastructure">Crusoe, Mar 27</a></li><li><strong>Abilene</strong> — PSD application by Crusoe for added generation at the Longhorn data center, under technical review; the link to Microsoft’s expansion is ours · <a href="https://www.tceq.texas.gov/downloads/permitting/air/bilingual/pending-permit-notices/182126-pls-english.pdf">TCEQ notice 182126</a></li><li><strong>Kilby</strong> — 20-year power agreement with Microsoft; about 2.67 GW of gas; FID by end-2026; first power 2028 · <a href="https://chevroncorp.gcs-web.com/node/38451">Chevron, Jun 22</a></li><li><strong>Kilby</strong> — Islanded campus, about 2 GW of data centers over 5–7 years · <a href="https://www.globaldatacenterhub.com/p/chevron-and-microsoft-sign-9b-west">Global Data Center Hub, Jun 29</a></li><li><strong>Kilby</strong> — Only a minor-source permit-by-rule application found at TCEQ · <a href="https://www.citizen.org/wp-content/uploads/Summary_-Data-Center-Permits-Issued-by-TCEQ-After-9_21_2026.pdf">Public Citizen permit summary</a></li><li><strong>Cameron</strong> — Fluidstack, up to 1.5 GW with on-site generation; phase 1 mid-2028 · <a href="https://www.theenergymag.com/news/2026-09-22/fluidstack-texas-ai-construction-cameron-county">The Energy Mag, Sep 22</a></li><li><strong>Milam</strong> — OpenAI 1.2 GW lease; SB Energy builds new generation · <a href="https://openai.com/index/stargate-sb-energy-partnership/">OpenAI, Jan 9</a></li><li><strong>Childress</strong> — Crusoe and Lancium 1 GW campus; construction Q3 2026 · <a href="https://www.crusoe.ai/resources/newsroom/crusoe-and-lancium-announce-1-gigawatt-ai-data-center-campus-in-childress-texas">Crusoe, Jul 15</a></li><li><strong>Childress</strong> — Meta contracts about 1.6 GW from Crusoe across Childress and Missouri (reported) · <a href="https://www.datacenterdynamics.com/en/news/meta-signs-16gw-capacity-agreement-with-crusoe-report/">DCD (Bloomberg)</a></li><li><strong>Google, Amazon</strong> — Campus power capacities (Armstrong ~1,065 MW, Wilbarger ~850 MW, Meitner ~840 MW); secondary · <a href="https://cleanview.co/data-centers/texas/1976/aws-comanche-peak-campus">Cleanview</a></li><li><strong>Comanche Peak</strong> — 1.2 GW nuclear PPA for Amazon; energize 2027; Vistra says the pause doesn’t affect it · <a href="https://www.utilitydive.com/news/vistra-texas-data-center-pause-comanche-amazon-earnings/827434/">Utility Dive, Aug 10</a></li><li><strong>GW Ranch</strong> — 7.65 GW gas air permit issued; phase 1 of 1 GW with first power H1 2027 · <a href="https://www.businesswire.com/news/home/20260126236053/en/Pacifico-Energy-Secures-7.65-GW-Power-Generation-Permit-for-GW-Ranch-Project">Pacifico Energy (BusinessWire), Jan 26</a></li><li><strong>GW Ranch</strong> — Amazon acquired the site in August (secondary) · <a href="https://letsdatascience.com/news/amazon-backs-765-gw-texas-gas-plant-5196d05e">Let’s Data Science, Aug 8</a></li><li><strong>El Paso</strong> — El Paso Electric’s 366 MW McCloud plant for Meta; judges recommend approval only if Meta bears the cost · <a href="https://insideclimatenews.org/news/24092026/el-paso-electric-only-considered-gas-to-power-meta-data-center/">Inside Climate News, Sep 24</a></li><li><strong>Helios</strong> — Base and studied classifications for Helios I–II and 2.6 GW of other sites · <a href="https://investingnews.com/galaxy-provides-update-on-ercot-batch-zero-large-load-classifications">Galaxy, Sep 8</a></li><li><strong>Helios</strong> — Phase II deliveries from Q2 2027; Helios II without a tenant; TCEQ freeze not expected to affect timelines · <a href="https://finance.biggo.com/news/US_GLXY_2026-10-27">BigGo Finance (Galaxy Q3 call), Oct 27</a></li><li><strong>Cipher</strong> — Colchis 1 GW base load; 1.6 GW studied; laterals for up to 2.5 GW of on-site generation · <a href="https://www.globenewswire.com/news-release/2026/11/02/3393000/0/en/cipher-digital-provides-third-quarter-2026-business-update.html">Cipher Digital, Nov 2</a></li><li><strong>Cipher</strong> — Barber Lake lease extended to 20 years, about $9.1B · <a href="https://www.globenewswire.com/news-release/2026/11/02/3393000/0/en/cipher-digital-provides-third-quarter-2026-business-update.html">Cipher Digital, Nov 2</a></li><li><strong>IREN</strong> — 2 GW Sweetwater hub included as Batch Zero base load · <a href="https://www.globenewswire.com/news-release/2026/09/08/3357498/0/en/iren-s-2gw-sweetwater-hub-included-as-base-load-in-ercot-batch-zero.html">IREN, Sep 8</a></li><li><strong>IREN</strong> — Horizon 1 delivered to Microsoft at Childress · <a href="https://www.globenewswire.com/news-release/2026/08/27/3352401/0/en/iren-reports-fy26-results.html">IREN FY26 results, Aug 27</a></li><li><strong>Riot, Hut 8</strong> — Rockdale 191 MW IT for an AI lab; Beacon Point 704 MW IT leased (secondary ledger citing company releases) · <a href="https://www.regardsofwallstreet.com/data/ai-deals">Regards of Wall Street</a></li><li><strong>Fermi</strong> — TensorWave lease not effective after the Oct 31 closing date passed · <a href="https://www.stocktitan.net/news/FRMI/fermi-provides-update-on-lease-agreement-as-parties-work-toward-s55mo4hurwnb.html">Fermi (via StockTitan), Nov 3</a></li><li><strong>Fermi</strong> — About 2 GW targeted by end-2027; a further ~5 GW air permit filed Mar 27 · <a href="https://convergedigest.com/fermi-america-resets-timeline-for-project-matador/?amp=1">Converge Digest (Fermi 10-K), Mar 30</a></li><li><strong>Horizon</strong> — CoreWeave’s deal with Poolside fell apart · <a href="https://baxtel.com/news/poolside-pursues-new-partners-for-texas-data-center-following-coreweave-deal-breakdown">Baxtel (Financial Times), Apr 4</a></li><li><strong>Sailfish</strong> — Hood County rejected the Comanche Circle settlement and permit appeal · <a href="https://www.tpr.org/news/2026-10-13/hood-county-rejects-comanche-circle-data-center-settlement-permit-appeal">Texas Public Radio, Oct 13</a></li><li><strong>San Antonio</strong> — Notice of intent to sue Vantage and VoltaGrid over off-grid permits · <a href="https://baxtel.com/news/lawsuit-targets-vantage-voltagrid-over-san-antonio-off-grid-data-centre-project">Baxtel, Jul 24</a></li></ul>
<h2>Method</h2>
<p><strong>What the paths count.</strong> New data center load switched on in Texas after September 2026, at quarter-ends through 2030, in five parts. ERCOT’s 17 cleared loads enter at 6.9 GW, the size of its November cohort of 17 loads, which we take to be the same ones, at their full requested peak; ERCOT gives that peak over five years, and we ramp it evenly. Batch Zero’s base load, 66.4 GW, is cut to the 49.8% realization rate of ERCOT’s 2025 forecast method, less the cleared loads, which we assume are part of it (26.2 GW), and ramps evenly from July 2027 to December 2032. Named self-powered campuses with permits pending (Nexus, Frontier’s second half, Abilene for Microsoft, Pecos) and with permits issued (Frontier’s first half, GW Ranch’s first phase, Meitner, the San Antonio pair) follow their announced or estimated dates, and the campuses outside ERCOT (El Paso, Abernathy) theirs. Batch Zero’s studied load (allocations from 2028, pending a study that hasn’t restarted) and the excluded 302 GW (waiting for Batch One) aren’t in the paths; we judge neither likely to arrive in volume before 2029. Fluidstack’s Cameron County site isn’t modeled separately: its grid request is in ERCOT’s queue, probably mostly in studied load.</p>
<p><strong>Before the freeze.</strong> The cleared loads start in December 2026, base load in July 2027, and every named campus on its own date. <strong>Abbott’s path</strong> moves the cleared loads to Q2 2027, when ERCOT slates its November cohort, base load by three months, and permits pending on September 21 by the 4.3 months to the end of January.</p>
<p><strong>Hinojosa’s path.</strong> A moratorium on new data centers runs from inauguration to the date picked (June 20, September 1 or December 31, 2027). It reaches 75% of base load, all pending own-gas permits and a quarter of the issued-permit and out-of-ERCOT campuses; it reaches the cleared loads only if the commission hasn’t approved them before January 19, which we put at even odds, and then half of them. Because the agencies are run by Abbott appointees and an order can be challenged, only part of the moratorium holds: 60% with Hinojosa alone, 80% with the lieutenant governor too and 90% with the Texas House as well. Each part’s schedule moves later by the share that holds times the months between the end of January and the moratorium’s end. Cancellations take a share of what each part adds after inauguration, across the three outcomes: 3–8% of base load, 10–20% of pending own-gas campuses, 3–8% of issued-permit campuses and 2–5% of those outside ERCOT, for local vetoes, a tax repeal without grandfathering and developers moving elsewhere. All of these are judgments, not data. The pace matters most: at a 25% realization of base load the base-case gap at the end of 2028 is 2.0 GW, at 100% it is 4.6 GW. Our Abbott path, about 30 GW of new data center load by 2030, is well above BNEF’s outlook of about 8 GW of additions in ERCOT through 2030; CenterPoint alone expects 12 GW of Batch Zero load of all kinds switched on in its territory by 2030.</p>
<p><strong>Exposure ratings</strong> are ours: high where a tenant-backed campus waits on a TCEQ air permit for power due in 2027–28; medium where it waits on an ERCOT switch-on, a commission vote or a contested permit; low where it is switched on or holds its approvals; company risk where its own financing or tenant, not the state, sets its timing. <strong>Results</strong> are unofficial counts and race calls as reported, with the time of each.</p>
<hr>
<p>Every chart above links to its interactive version on 06.ai, with the data behind it. The daily trackers are at <a href="https://06.ai/trackers">06.ai/trackers</a>.</p>
<p><em>Not investment advice. Estimates are labeled as estimates.</em></p>]]></content:encoded>
</item>
<item>
  <title>How many Cybercabs will be on US roads by 2027?</title>
  <link>https://inthesingularity.ai/packs/cybercab-rollout</link>
  <guid isPermaLink="true">https://inthesingularity.ai/packs/cybercab-rollout</guid>
  <pubDate>Mon, 02 Nov 2026 12:18:00 GMT</pubDate>
  <description>Texas lists every robotaxi it authorizes, car by car. We rebuilt the Cybercab’s first two months from that roster, mapped where state law already allows a car with no driver, and projected the US fleet month by month through 2027.</description>
  <category>Autonomy</category>
  <enclosure url="https://inthesingularity.ai/newsletter/cybercab-rollout/cover.png" length="64503" type="image/png" />
  <media:content url="https://inthesingularity.ai/newsletter/cybercab-rollout/cover.png" medium="image" type="image/png" />
  <content:encoded><![CDATA[<figure><a href="https://06.ai/packs/cybercab-rollout"><img src="https://06.ai/newsletter/cybercab-rollout/thumb.png" alt="How many Cybercabs will be on US roads by 2027?" width="421" height="391"></a><figcaption><a href="https://06.ai/packs/cybercab-rollout">Interactive version and data on 06.ai →</a></figcaption></figure>
<p>Tesla went from 45 Cybercabs on Texas’s robotaxi roster on August 31 to 573 on October 28, doubling about every two weeks. Our base case has <strong>about 2,000 on US roads at the end of 2026</strong> and <strong>about 29,000 at the end of 2027</strong>, in 15 metro areas. If NHTSA rejects the car’s self-certification, the end-2027 figure is nearer 4,900; if Tesla opens cities as fast as it can build cars, about 134,000.</p>
<ul><li><strong>573</strong> — On Texas’s roster, Nov 1. From 45 on Aug 31. Model Y: 671. Waymo: 1,259</li><li><strong>2,000</strong> — US fleet, end of 2026. Base case. Range 1,000–4,500</li><li><strong>29,000</strong> — US fleet, end of 2027. Base case. Range 4,900–134,000</li><li><strong>Apr 2027</strong> — 10,000th Cybercab. Base case. 50,000: Jul 2028. 100,000: Jun 2029</li></ul>
<h2>Texas counts every car</h2>
<figure><a href="https://06.ai/packs/cybercab-rollout#roster"><img src="https://06.ai/newsletter/cybercab-rollout/fig-01.png" alt="Texas counts every car" width="694" height="412"></a><figcaption>Vehicles on the Texas DMV’s automated-vehicle roster. Daily snapshots from October 9; earlier points are press reports of the same roster. A listing means a car is authorized to carry riders, not that it is.</figcaption></figure>
<h2>It arrives in batches</h2>
<p>Since May 28, Texas has required every company running driverless commercial vehicles to hold an authorization from its DMV, and the authorization lists each car. That makes Texas the one state where Cybercabs can be counted. The first seven appeared on August 31, then 45 the same day, three days before Tesla’s launch event in Austin. The count reached 126 by September 26, jumped by 150 on October 9 and by 81 on October 27, and has stood at 573 since October 28. Across the whole record it has doubled every 14 days; over the last 30, every 16. On October 27 Cybercabs passed Model Ys on the roster for the first time, and by November 1 Tesla listed 1,244 vehicles in Texas to Waymo’s 1,259.</p>
<p>A listing is not a ride. A crowd-sourced tracker had spotted 229 different Cybercabs on the road by late October. Tesla said on October 10 that more than 300 were serving Austin, and on its October 21 call that the Austin Cybercab fleet had grown tenfold in its first month.</p>
<h2>Tesla is building faster than it deploys</h2>
<figure><a href="https://06.ai/packs/cybercab-rollout#builds"><img src="https://06.ai/newsletter/cybercab-rollout/fig-02.png" alt="Tesla is building faster than it deploys" width="694" height="382"></a><figcaption>The highest Cybercab VIN serial seen on each date, against Cybercabs on the Texas roster and the number spotted on the road. Serials run in sequence but can be skipped or reserved, so they are a ceiling on cars built. The dashed extension assumes August and September’s pace held.</figcaption></figure>
<h2>The factory isn’t the constraint</h2>
<p>Tesla assigned VIN number 1 at the start of production on April 24. Serial 2,172 was photographed on August 9, and serial 3,358 joined the Texas roster on September 18, when the roster held fewer than 70 Cybercabs. If the serials are close to sequential, Tesla was building about 30 a day in August and September; at that pace it would have built about 4,700 by November 1, eight times what Texas lists. Tesla’s Lars Moravy said in October that the line, rated at more than 125,000 cars a year, doesn’t run every day because it is outpacing deployments. About 200 Cybercabs were parked at Tesla’s Houston robotaxi lot in early October, and residents near Miami’s airport had watched another lot fill for about two months. What limits the fleet is where Tesla may operate and how fast it opens cities.</p>
<h2>Extrapolating Texas gives nonsense either way</h2>
<figure><a href="https://06.ai/packs/cybercab-rollout#extrapolate"><img src="https://06.ai/newsletter/cybercab-rollout/fig-03.png" alt="Extrapolating Texas gives nonsense either way" width="694" height="490"></a><figcaption>Cybercabs on the Texas roster, two simple extensions of it from November 1, and our three scenarios for the whole US at each month-end. Log scale.</figcaption></figure>
<h2>Cities, not curves</h2>
<p>If the roster kept doubling every two weeks, Texas alone would hold about 12,000 Cybercabs at the end of 2026, 100,000 by mid-February and a million by the end of March 2027, eight times what the line is rated to build in a year. If it kept October’s pace of about 415 a month, the country would have about 6,400 at the end of 2027. Neither is how a robotaxi fleet grows. It grows when a city opens: the Texas roster gained more than 500 Cybercabs in the two months after Austin opened. So we model the fleet as a rollout of cities, each with a date it opens, a pace at which it fills and a number of cars it can use.</p>
<h2>Where the law already allows a car with no driver</h2>
<figure><a href="https://06.ai/packs/cybercab-rollout#states"><img src="https://06.ai/newsletter/cybercab-rollout/fig-04.png" alt="Where the law already allows a car with no driver" width="694" height="630"></a><figcaption>State rules for driverless commercial service, as surveyed in July 2026, with where Tesla and Waymo carry riders today. Hover or tap a state for details.</figcaption></figure>
<h2>Twenty-two states, most of the Sun Belt</h2>
<p>Twenty-one states allow driverless commercial service in statute, and Colorado allows it if the system can follow traffic laws, which is how Waymo runs in Denver. Texas requires the DMV authorization above; Florida requires no state permit; Arizona runs on self-certification, and Tesla already holds its statewide ride-hail permit. Nevada’s Transportation Authority approved Tesla for up to 5,000 cars in Las Vegas on August 20, with 120 days from the permit’s issue to start service. Georgia, Tennessee, North Carolina, Louisiana, Michigan, Oklahoma and Utah allow it too, and Pennsylvania does once PennDOT reviews a company’s self-certification. California allows it but requires DMV driverless and CPUC deployment permits; Tesla holds neither and runs its Bay Area service with safety drivers. New York and Washington, DC still require a human in the car.</p>
<p>The federal question is bigger. The Cybercab has no steering wheel or pedals, and Tesla certified it as meeting every federal safety standard instead of asking NHTSA for an exemption. NHTSA opened an audit of that certification hours after public rides began on September 4, and on September 10 ordered Tesla to justify it, starting with FMVSS 135’s rule that service brakes be applied by a foot control. Tesla asked for until November 20; NHTSA gave it until October 30. If NHTSA disagrees, the fallback is an exemption like the one Zoox received in July, which caps a pedal-less fleet at 2,500 cars a year. NHTSA proposed changes to FMVSS 135 in June, but a proposal doesn’t change the standard, and Congress has bills that would raise the exemption cap to 90,000 a year; none is law.</p>
<h2>Copying Waymo’s map</h2>
<figure><a href="https://06.ai/packs/cybercab-rollout#waymo-map"><img src="https://06.ai/newsletter/cybercab-rollout/fig-05.png" alt="Copying Waymo’s map" width="694" height="908"></a><figcaption>Metro areas where Waymo carries riders or says it is going next, plus Tesla’s named cities, grouped by what the state allows. The dot is when the base case opens Cybercab rides; the bar runs from the bull case’s date to the bear case’s.</figcaption></figure>
<h2>Most of Waymo’s map is open to Tesla</h2>
<p>Waymo carries riders in 15 US metro areas. Twelve are in states that already allow driverless service without permits Tesla lacks; the other three are in California. Tesla runs driverless rides in six of the twelve (Austin, Dallas, Houston, Miami, Orlando and Tampa) and has named three more (Las Vegas, Phoenix and San Antonio), plus Jacksonville, which Waymo doesn’t serve. That leaves Atlanta, Nashville and Denver. Waymo’s list of next cities adds five more in friendly states (Charlotte, Detroit, New Orleans, Philadelphia and Pittsburgh), Sacramento, and nine that need a state law or permit change first, led by New York, Chicago and Washington.</p>
<p>The base case copies the friendly part of that map. Cybercab rides open in Dallas in November, Houston in December, Miami and Las Vegas in January, the rest of Tesla’s named cities by May, then Atlanta, Nashville, Denver, Charlotte and New Orleans from June to October 2027. That puts about 59 million people in Cybercab metros by the end of 2027, against 2.6 million today. Detroit, Pittsburgh and Philadelphia wait for spring 2028. The bull case runs the same list faster, adds 20 smaller friendly-state metros from mid-2027 and California from October 2027. The bear case gets through Tesla’s named cities only by early 2028.</p>
<h2>The forecast, month by month</h2>
<figure><a href="https://06.ai/packs/cybercab-rollout#forecast"><img src="https://06.ai/newsletter/cybercab-rollout/fig-06.png" alt="The forecast, month by month" width="694" height="462"></a><figcaption>Cybercabs in robotaxi service in the US at each month-end: cars listed in a fleet, as Texas counts them. October is the Texas roster. Horizontal lines mark 10,000, 50,000 and 100,000. Shown: Linear.</figcaption></figure>
<figure><a href="https://06.ai/packs/cybercab-rollout#forecast"><img src="https://06.ai/newsletter/cybercab-rollout/fig-07.png" alt="Oct 2026 (actual) — Bear: 573 · Base: 573 · Bull: 573 · Weighted: 573 · Base metros: 1; Nov 2026 — Bear: 765 · Base: 1,100 · Bull: 1,820 · Weighted: 1,200 · Base metros: 2; Dec 2026 — Bear: 991 · Base: 2,010 · Bull: 4,450 · Weighted: 2,360 · Base metros: 3; Jan 2027 — Bear: 1,490 · Base: 3,390 · Bull: 8,480 · Weighted: 4,190 · Base metros: 5; Feb 2027 — Bear: 2,250 · Base: 5,160 · Bull: 13,900 · Weighted: 6,610 · Base metros: 5; Mar 2027 — Bear: 3,650 · Base: 7,330 · Bull: 20,600 · Weighted: 9,730 · Base metros: 7; Apr 2027 — Bear: 3,650 · Base: 10,200 · Bull: 29,000 · Weighted: 13,300 · Base metros: 9; May 2027 — Bear: 3,650 · Base: 13,500 · Bull: 38,600 · Weighted: 17,300 · Base metros: 10; Jun 2027 — Bear: 3,650 · Base: 16,800 · Bull: 49,000 · Weighted: 21,600 · Base metros: 11; Jul 2027 — Bear: 3,850 · Base: 19,800 · Bull: 63,000 · Weighted: 26,600 · Base metros: 12; Aug 2027 — Bear: 4,060 · Base: 22,200 · Bull: 78,000 · Weighted: 31,600 · Base metros: 13; Sep 2027 — Bear: 4,270 · Base: 23,900 · Bull: 92,300 · Weighted: 36,100 · Base metros: 14; Oct 2027 — Bear: 4,480 · Base: 25,900 · Bull: 106,000 · Weighted: 40,600 · Base metros: 15; Nov 2027 — Bear: 4,690 · Base: 27,800 · Bull: 120,100 · Weighted: 45,100 · Base metros: 15; Dec 2027 — Bear: 4,900 · Base: 29,400 · Bull: 134,100 · Weighted: 49,400 · Base metros: 15" width="694" height="675"></a></figure>
<p>Weighted: 25% bear, 50% base, 25% bull.</p>
<figure><a href="https://06.ai/packs/cybercab-rollout#forecast"><img src="https://06.ai/newsletter/cybercab-rollout/fig-08.png" alt="The forecast, month by month — Log scale" width="694" height="493"></a><figcaption><strong>Log scale.</strong></figcaption></figure>
<h2>Q4 2026 and 2027</h2>
<p><strong>The fourth quarter.</strong> Tesla said Cybercab rides start in Dallas in November and in “some other cities” before the year ends. The base case opens Dallas in November and Houston in December, where about 200 Cybercabs were already parked in October, and reaches about 2,000 on December 31, up from 573: about 700 a month, against about 415 in October. The bull case also opens Houston in November and Miami and Las Vegas in December, for about 4,500. The bear case adds only Dallas, for about 1,000.</p>
<p><strong>2027.</strong> In the base case each city adds Cybercabs at about Austin’s pace, 275 a month for a metro of Austin’s size (scaled by the square root of population), until it reaches 0.4 Cybercabs per 1,000 residents, a ceiling that rises to 0.5 by December 2027. The fleet passes Waymo’s whole current fleet of about 4,000 in February 2027 and 10,000 in April 2027, and ends 2027 at about 29,000 in 15 metro areas. On the October call Musk said robotaxis start to matter to Tesla’s results in the second quarter of 2027 and become a big deal by the third or fourth; Tesla’s Ashok Elluswamy said the limits would shift from technology to regulation around the first half.</p>
<p><strong>The range.</strong> The bull case has Cybercabs in 35 metros by December 2027 at up to 1.5 per 1,000 residents, and is held back by the factory in the second half, ending at about 134,000. The bear case assumes NHTSA rejects the self-certification: no new Cybercabs from April to June 2027 while Tesla files for an exemption, then 2,500 a year until the foot-control rule changes in mid-2028. It ends 2027 at about 4,900. Weighting the bear and bull cases at a quarter each and the base at half gives about 49,000, pulled up by the bull case.</p>
<h2>Texas falls to about a third</h2>
<figure><a href="https://06.ai/packs/cybercab-rollout#where"><img src="https://06.ai/newsletter/cybercab-rollout/fig-09.png" alt="Texas falls to about a third" width="694" height="422"></a><figcaption>Cybercabs in service by state group at each month-end. Shown: Base.</figcaption></figure>
<p>Texas holds every Cybercab today. In the base case it is still the largest group at the end of 2027, with about 37% of the fleet, ahead of the other friendly states (26%), Florida (25%) and Arizona and Nevada (13%). Its share falls because the second wave of cities is outside it. In the bull case California adds about 7,200 by December 2027, three months after its first city opens.</p>
<h2>When the fleet passes 10,000, 50,000, 100,000 and a million</h2>
<figure><a href="https://06.ai/packs/cybercab-rollout#milestones"><img src="https://06.ai/newsletter/cybercab-rollout/fig-10.png" alt="When the fleet passes 10,000, 50,000, 100,000 and a million" width="694" height="320"></a><figcaption>First month-end with at least this many Cybercabs in US robotaxi service.</figcaption></figure>
<ul><li><strong>10,000</strong> — Bear: Jul 2028 · Base: Apr 2027 · Bull: Feb 2027</li><li><strong>50,000</strong> — Bear: May 2029 · Base: Jul 2028 · Bull: Jul 2027</li><li><strong>100,000</strong> — Bear: Oct 2030 · Base: Jun 2029 · Bull: Oct 2027</li><li><strong>1,000,000</strong> — Bear: after 2034 · Base: Dec 2033 · Bull: Jun 2030</li></ul>
<h2>A million is a 2030s number</h2>
<p>The base case passes 10,000 in April 2027, between the bull case’s February 2027 and the bear case’s July 2028. 50,000 and 100,000 need cities beyond Waymo’s current map: the base case gets there in July 2028 and June 2029, the bull case in July 2027 and October 2027, when the factory is the limit. A million needs about five Cybercabs per 1,000 residents across most of the country’s big metros, ten times our 2027 density. The base case gets there in December 2033 and the bull case in June 2030; the bear case doesn’t by 2034. Beyond 2027 these are scenarios, not forecasts: they rest on how far prices fall and how many states change their rules.</p>
<h2>What a million would have to do</h2>
<figure><a href="https://06.ai/packs/cybercab-rollout#demand"><img src="https://06.ai/newsletter/cybercab-rollout/fig-11.png" alt="What a million would have to do" width="694" height="280"></a><figcaption>Rides a day at 15 per car (Waymo’s cars average about 21–28 in its busiest cities; Tesla’s run lower), against Waymo’s current pace and all US Uber and Lyft trips in 2025. Log scale.</figcaption></figure>
<h2>Demand is the ceiling</h2>
<p>At 15 rides a car a day, 29,000 Cybercabs would give about 440,000 rides a day, six times Waymo’s current pace. 100,000 would give 1.5 million, about a fifth of all US Uber and Lyft trips in 2025. A million would give 15 million, 2.2 times what Uber and Lyft carry in the whole country. Tesla can build that many cars. The open questions are whether rides get cheap enough to replace driving, city by city, and whether regulators let Tesla follow the riders.</p>
<h2>What to watch</h2>
<p><strong>NHTSA’s answer.</strong> Tesla’s sworn answers were due October 30. A closed audit removes most of the bear case; a Part 555 exemption petition in the Federal Register would confirm it.</p>
<p><strong>Dallas in November.</strong> The base case needs the Texas roster to add about 700 a month through December; October added about 415.</p>
<p><strong>Las Vegas by mid-to-late December,</strong> when the 120-day clock on Tesla’s Nevada permit runs out; it counts from the permit’s issue, after the August 20 approval. Tesla can start with Model Ys; Cybercabs there would show that a second state is open to them.</p>
<p><strong>Tesla’s fourth-quarter letter,</strong> due in late January, the first chance for an official fleet count, and whether Phoenix, San Antonio and Jacksonville move from “coming soon” to live.</p>
<h2>What this is anchored on</h2>
<p>Roster counts, permits and statements are as reported; build estimates, launch dates and fleet paths are ours.</p>
<ul><li><strong>Texas roster</strong> — Every vehicle listed under Tesla Robotaxi LLC’s and Waymo LLC’s automated-motor-vehicle authorizations · <a href="https://txmccs.txdmv.gov/truckstop/companies/81edcff1-8a6e-4ed0-be1f-60668515e223/automated-motor-vehicles">Texas DMV (TxDMV)</a></li><li><strong>Texas roster</strong> — Daily snapshots of the roster, October 9–November 1 (via fsddb.com) · <a href="https://github.com/matthewsvm/cybercab-dashboard">Texas Robotaxi Fleet Tracker (open source)</a></li><li><strong>Texas roster</strong> — November 1: 1,244 Tesla vehicles, including 40 entered as “Modey Y” with Model Y VINs · <a href="https://github.com/derekbeau1/tx-robotaxi-tracker/issues/6">tx-robotaxi-tracker, Nov 1</a></li><li><strong>Texas roster</strong> — First Cybercabs listed August 31: seven, then 45 the same day · <a href="https://driveteslacanada.ca/news/tesla-registers-first-cybercabs-austin-robotaxi-fleet/">Drive Tesla, Aug 31</a></li><li><strong>Texas roster</strong> — 45 Cybercabs in a Texas robotaxi fleet of about 420 when rides began · <a href="https://en.wikipedia.org/wiki/Tesla_Robotaxi">Wikipedia, Tesla Robotaxi</a></li><li><strong>Texas roster</strong> — 49 Cybercabs on September 8 · <a href="https://teslanorth.com/2026/09/09/tesla-cybercab-fleet-49-texas/">Tesla North, Sep 9</a></li><li><strong>Texas roster</strong> — 420 Model Ys and 126 Cybercabs on September 26 · <a href="https://www.teslaoracle.com/2026/09/26/tesla-robotaxi-and-cybercab-fleet-in-texas-surpasses-the-500-mark-420-126/">Tesla Oracle, Sep 26</a></li><li><strong>Texas roster</strong> — 158 Cybercabs on October 2 · <a href="https://www.basenor.com/blogs/news/tesla-cybercab-fleet-hits-158-in-texas-your-questions-answered">Basenor, Oct 2 update</a></li><li><strong>Texas roster</strong> — 69 Cybercabs on September 23; 169 before a record 150 were added on October 9 · <a href="https://evwire.com/p/tesla-cybercab-texas-registrations-record-150-almost-doubled-319">EVWire, Oct 9</a></li><li><strong>Texas roster</strong> — 544 Cybercabs and 507 Model Ys on October 27; Cybercabs pass Model Ys · <a href="https://driveteslacanada.ca/news/tesla-cybercabs-outnumber-model-y-texas-robotaxi-fleet/">Drive Tesla, Oct 27</a></li><li><strong>Texas law</strong> — TxDMV authorization required for commercial driverless operation from May 28, 2026 (SB 2807) · <a href="https://www.robotaxiatlas.com/rules/texas/">Robotaxi Atlas</a></li><li><strong>Use</strong> — 229 Cybercabs spotted on the road against 573 listed, late October · <a href="https://pioneerlands.org/cybercabtracker">Pioneerlands Cybercab Counter (RobotaxiTracker data)</a></li><li><strong>Builds</strong> — Start of production and VIN #1 on April 24; VIN #2172 photographed August 9 · <a href="https://www.basenor.com/blogs/news/cybercab-vin-2172-spotted-how-far-along-is-production">Basenor</a></li><li><strong>Builds</strong> — VIN #3358 added to the Texas roster, September 18 · <a href="https://x.com/tslaming/status/2101143120046096731">Ming (@tslaming) on X</a></li><li><strong>Builds</strong> — Line outpacing deployments, doesn’t run every day; 25 final-assembly stations; 10-second cycle target · <a href="https://teslanorth.com/2026/10/26/tesla-cybercab-line-outpacing-robotaxi-munro/">Tesla North, on Munro Live, Oct 26</a></li><li><strong>Builds</strong> — About 190–200 Cybercabs at Tesla’s Houston robotaxi lot; a Cybercab lot near Miami’s airport · <a href="https://www.notateslaapp.com/news/4776/tesla-may-be-getting-ready-to-launch-cybercabs-in-houston">Not a Tesla App, Oct 6</a></li><li><strong>Tesla</strong> — Austin Cybercab fleet up 10× since launch; “we do 6 cities now”; customer Cybercabs by end-2026; customer cars in the fleet around Q2 2027; limits turn “regulatory perhaps in Q1, Q2”; a “big deal by Q3 or Q4 of next year”; 125,000+ a year is the rate the line can run today · <a href="https://www.fool.com/earnings/call-transcripts/2026/10/22/tesla-tsla-q3-2026-earnings-call-transcript/">Tesla Q3 2026 earnings call, Oct 21 (transcript)</a></li><li><strong>Tesla</strong> — Call recap: Cybercab rides in Dallas next month; staging lots across the US · <a href="https://www.notateslaapp.com/news/4867/tesla-q3-2026-earnings-call-recap-everything-you-need-to-know">Not a Tesla App, Oct 22</a></li><li><strong>Tesla</strong> — Call recap: Cybercab in Dallas next month and in “some other cities” before the end of 2026 · <a href="https://teslanorth.com/2026/10/21/tesla-q3-2026-call-takeaways/">Tesla North, Oct 21</a></li><li><strong>Tesla</strong> — Las Vegas, Phoenix, San Antonio and Jacksonville listed as coming soon; driverless miles by city · <a href="https://teslanorth.com/2026/10/21/tesla-robotaxi-safety-report-four-new-cities/">Tesla North, Oct 21</a></li><li><strong>Tesla</strong> — Driverless service: Austin; Dallas and Houston from April 18; Miami from July 3; Orlando and Tampa from July 21 · <a href="https://en.wikipedia.org/wiki/Tesla_Robotaxi">Wikipedia, Tesla Robotaxi</a></li><li><strong>Nevada</strong> — Transportation Authority approves Tesla’s, Waymo’s and Aviari’s applications at its August 20 meeting · <a href="https://www.business.nv.gov/news-media/press-releases/2026/ntata/nevada-transportation-authority-approves-autonomous-vehicle-network-applications-for-tesla-waymo-and-aviari/">Nevada Dept. of Business &amp; Industry, Aug 21</a></li><li><strong>Nevada</strong> — Up to 5,000 cars for Tesla and 1,000 for Waymo in Clark County; service must start within 120 days · <a href="https://thenevadaglobe.com/702times/nevada-regulators-approve-commercial-robotaxi-permits-for-tesla-waymo-and-uber-authorizing-up-to-7000-driverless-vehicles-in-clark-county/">Nevada Globe</a></li><li><strong>Arizona</strong> — Tesla’s statewide ride-hail (TNC) permit, November 2025; self-certification regime · <a href="https://www.inkl.com/news/tesla-gets-approval-to-launch-robotaxi-service-in-arizona-without-safety-monitors">inkl (InsideEVs)</a></li><li><strong>States</strong> — State-by-state rules for driverless operation, July 2026 (built on NCSL’s database) · <a href="https://elevationprovinggrounds.com/article/where-are-autonomous-vehicles-legal/">Elevation Proving Grounds</a></li><li><strong>States</strong> — Tesla said 300+ Cybercabs were serving Austin (Oct 10); Tesla holds only a California testing permit with a driver; Arizona first-responder plan; Zoox exemption of up to 2,500 cars a year (July 30) · <a href="https://thechargeport.com/robotaxi-tracker">The Charge Port robotaxi tracker, Oct 25</a></li><li><strong>NHTSA</strong> — Audit query AQ26002 into the Cybercab’s self-certification, opened September 4 · <a href="https://www.karmactive.com/tesla-cybercab-nhtsa-audit-query-aq26002-september-2026/">Karmactive</a></li><li><strong>NHTSA</strong> — Special Order of September 10 citing FMVSS 135’s foot-control rule; a June 26 proposal to change FMVSS 135; Part 555 cap of 2,500 a year; a House committee vote to lift the cap to 90,000 · <a href="https://thefogsignal.substack.com/p/the-pedal-is-the-permit">The Fog Signal</a></li><li><strong>NHTSA</strong> — Tesla asked for until November 20; NHTSA set October 30 · <a href="https://finance.yahoo.com/markets/stocks/articles/tesla-stocks-move-lower-cybercab-190652852.html">Yahoo Finance, Oct 30</a></li><li><strong>Congress</strong> — SELF DRIVE Act draft would raise the exemption cap to 90,000 a year, then 100,000 · <a href="https://www.notateslaapp.com/news/3508/congress-ready-to-pass-federal-autonomous-vehicle-regulations">Not a Tesla App, Jan 15</a></li><li><strong>Waymo</strong> — About 4,000 cars (September); more than 500,000 paid rides a week; 15 US metro areas; its list of next cities · <a href="https://thechargeport.com/robotaxi-tracker">The Charge Port, Oct 25</a></li><li><strong>Waymo</strong> — City-by-city status and launch dates · <a href="https://www.driverlessrider.com/waymo-cities/">Driverless Rider, Oct 21</a></li><li><strong>Waymo</strong> — Fleet of 2,500 in November 2025 and 3,800 in May 2026 · <a href="https://www.thedriverlessdigest.com/p/waymo-stats-2025-funding-growth-coverage">The Driverless Digest</a></li><li><strong>Rides</strong> — About 2.50 billion US Uber and Lyft trips in 2025 (estimate); Waymo’s cars at about 21–28 rides a day · <a href="https://www.wolferesearch.com/wp-content/uploads/2026/03/Uber-Lyft-Deep-Dive-Updated-AV-Market-Model-and-Share-Shifts.pdf">Wolfe Research</a></li><li><strong>Population</strong> — Metro-area populations, July 2024 Census Bureau estimates · <a href="https://www.citypopulation.de/en/usa/metro">City Population</a></li></ul>
<h2>Method</h2>
<p><strong>Texas record.</strong> TxDMV’s roster lists every vehicle under each operator’s automated-motor-vehicle authorization. Daily snapshots of Tesla Robotaxi LLC and Waymo LLC from October 9 to November 1 come from fsddb.com’s copy of the roster, collected by the open-source Texas Robotaxi Fleet Tracker; earlier points are press reports of the same roster. On November 1 the roster also listed 40 vehicles entered as “Modey Y”, all with Model Y VINs; we count them as Model Ys. The doubling time is a least-squares fit of the log count on the day, across the daily record from August 31 to November 1 with gaps carried forward.</p>
<p><strong>Builds.</strong> VIN serials are the last digits of Cybercab VINs seen in photos or on the roster. Serials can be skipped or reserved, so they are a ceiling on cars built. The November figure extends the pace between August 9 and September 18, about 30 a day.</p>
<p><strong>Fleet.</strong> Each metro area opens on a scenario date. From then it adds R₀ × √(population ÷ Austin’s) Cybercabs a month (half that in its first month) until it reaches population × D, where D is Cybercabs per 1,000 residents. R₀ is 275 in the base case, about Austin’s pace in its first two months; 450 in the bull case and 250 in the bear. D starts at 0.4 (bull 0.8, bear 0.3), reaches 0.5 (1.5, 0.45) by December 2027 and rises in a straight line between year-ends after that, to 0.8, 1.25, 1.9, 2.8, 3.9 and 5.1 by 2033 in the base case. R₀ rises in proportion to D. At 15 rides a car a day, 0.5 per 1,000 residents is about 0.0075 rides per resident a day, a little over a third of what Americans take on Uber and Lyft per person today. National additions are capped by production: 10,400 a month, the line’s rating, through 2027 in the base and bear cases, then 20,000 in 2028, 35,000 in 2029, 50,000 in 2030 and 80,000 after; in the bull case 15,000 from mid-2027, 30,000 in 2028, 60,000 in 2029 and 100,000 after. The bear case adds a federal limit: no additions from April to June 2027, then 2,500 a year to June 2028. When a limit binds, every city’s additions are scaled down in proportion.</p>
<p><strong>Cities.</strong> Launch dates are ours. The base case follows Tesla’s guidance for Dallas and the order of its named cities, then Waymo’s markets and next cities in states that allow driverless service; 20 more friendly-state metros open one a month from January 2028, California’s five metros every other month from 2029 and states that must change their rules from 2030. In service means listed in a robotaxi fleet, as Texas counts; on any day fewer carry riders. Metro populations are Census Bureau estimates for July 2024, rounded. The weighted path puts 25% on the bear case, 50% on the base and 25% on the bull.</p>
<p><strong>Rides.</strong> 2025 US Uber and Lyft trips are Wolfe Research’s estimate, about 2.50 billion, or 6.9 million a day for a population of about 340 million. Waymo’s pace is its last reported figure, more than 500,000 paid rides a week.</p>
<hr>
<p>Every chart above links to its interactive version on 06.ai, with the data behind it. The daily trackers are at <a href="https://06.ai/trackers">06.ai/trackers</a>.</p>
<p><em>Not investment advice. Estimates are labeled as estimates.</em></p>]]></content:encoded>
</item>
<item>
  <title>Can Anthropic’s compute keep up with its demand?</title>
  <link>https://inthesingularity.ai/packs/anthropic-s1</link>
  <guid isPermaLink="true">https://inthesingularity.ai/packs/anthropic-s1</guid>
  <pubDate>Sat, 31 Oct 2026 12:17:00 GMT</pubDate>
  <description>Anthropic’s S-1 counts about 250 customers paying it $10M or more a year, out of roughly 15,000 companies big enough to spend that at 1% of revenue. We size the market at 1% and 5% of revenue, chart how much less compute each dollar of demand needs, and carry demand and compute forward by quarter to 2029.</description>
  <category>AI demand</category>
  <enclosure url="https://inthesingularity.ai/newsletter/anthropic-s1/cover.png" length="61843" type="image/png" />
  <media:content url="https://inthesingularity.ai/newsletter/anthropic-s1/cover.png" medium="image" type="image/png" />
  <content:encoded><![CDATA[<figure><a href="https://06.ai/packs/anthropic-s1"><img src="https://06.ai/newsletter/anthropic-s1/thumb.png" alt="Can Anthropic’s compute keep up with its demand?" width="421" height="392"></a><figcaption><a href="https://06.ai/packs/anthropic-s1">Interactive version and data on 06.ai →</a></figcaption></figure>
<p>Anthropic’s S-1 counts about <strong>250</strong> customers paying it $10M or more a year. At 1% of revenue, about <strong>15,000</strong> companies could; at 5%, about 72,000. Serving each dollar of that demand now costs Anthropic <strong>15¢</strong>, down from 88¢ a year earlier. So far it has kept the savings as margin, now 85%. With little margin left to gain, its compute has to keep pace with its revenue from here, unless it moves compute from research to customers. Carried forward by quarter, demand outruns the roughly 10 GW Anthropic has lined up for the end of 2027 on its reported trend and on our tracker; only its own plan fits. The question for the buildout isn’t whether the demand is there. It’s whether the compute arrives in time.</p>
<ul><li><strong>1.7%</strong> — $10M+ customers, of the companies that could be. 250 of about 15,000 companies with $1B+ of revenue, at 1% of revenue. At 5%, 250 of about 72,000 (0.35%)</li><li><strong>0.16%</strong> — AI model spend, share of big-company revenue. Anthropic and OpenAI run at about $153B a year, against $97T of revenue at $1B+ companies. 1% is $0.97T; 5% is $6.0T, about all IT spending today</li><li><strong>15¢</strong> — Anthropic’s cost to serve $1 of revenue cost of revenue per dollar, H1 2026 and Q3 2026. 88¢ in H1 2025 and $1.95 in 2024</li><li><strong>13–18 GW</strong> — Compute needed at the end of 2027 for the reported trend ($309B a year) or our tracker ($425B), if research keeps three-quarters of compute. About 10 GW is lined up</li></ul>
<h2>Revenue up 12× in a year, compute up 4.8×</h2>
<figure><a href="https://06.ai/packs/anthropic-s1#revenue"><img src="https://06.ai/newsletter/anthropic-s1/fig-01.png" alt="Revenue up 12× in a year, compute up 4.8×" width="694" height="332"></a><figcaption>Anthropic’s revenue and its compute and infrastructure expense by quarter, $B. Q1 2025 to Q2 2026 from the S-1’s quarterly results; Q3 2026 is the revenue Anthropic gave investors on October 14, with compute backed out of the $1.70 of revenue per $1 of compute it gave for the quarter. The figure above each pair is revenue per $1 of compute.</figcaption></figure>
<h2>What the S-1 says about demand</h2>
<p><strong>Customers are spending more, faster than anything we can compare it to.</strong> Net revenue retention is about 900%: customers billed in last year’s third quarter paid nine times as much this quarter. Customers paying $100K a year went from about 950 to more than 11,900 in a year, and doubled in the third quarter alone. Over half of them now use both the platform and one of Anthropic’s apps.</p>
<p><strong>Demand follows the models.</strong> Over 80% of third-quarter revenue came from models released in the six months before. Revenue growth slowed to 45% in the third quarter, from 144% in the second; part of that was Anthropic cutting off customers who broke its terms, mostly by distilling Claude, which had brought in 5–7% of revenue.</p>
<p><strong>And compute is the limit Anthropic names.</strong> The filing says demand for compute “has historically exceeded available capacity” across research, serving and its own use, which it runs as one pool; that it has asked its cloud providers for more compute than they were willing to provide; and that it has changed usage and rate limits. It doesn’t say how much customer demand went unserved.</p>
<h2>Demand in the S-1</h2>
<ul><li><strong>Customers paying $100K+ a year</strong> — Figure: More than 11,900 on September 30, 2026, from more than 950 a year earlier · Trailing twelve months</li><li><strong>Paying $1M+ / $10M+</strong> — Figure: More than 1,900 / about 250 · Same date and basis; no year-earlier counts given</li><li><strong>Net revenue retention</strong> — Figure: About 900% · This quarter’s revenue from last year’s customers against the same quarter a year ago; Pro and Max plans excluded</li><li><strong>How customers pay</strong> — Figure: 80% consumption (tokens), 20% subscriptions · First half of 2026; 83% and 17% in 2025</li><li><strong>Through AWS, Google Cloud, Azure</strong> — Figure: 48% of revenue · First half of 2026, from 11% in 2023</li><li><strong>Concentration</strong> — Figure: Top five customers about 25% of revenue; none above 10% · Nine months to September; 79% of Q3 revenue came from outside the top five, from 64% in Q1</li><li><strong>More than one product</strong> — Figure: Over half of $100K+ customers use the platform and an app · From about 30% at the end of 2025</li><li><strong>New models</strong> — Figure: Over 80% of Q3 revenue · From models released in the previous six months</li><li><strong>Revenue from rule-breakers</strong> — Figure: 5–7% of revenue · Nine months; customers later found to break its terms, mostly distillation, cut back materially in Q3</li><li><strong>Price per task</strong> — Figure: Opus 5.5 about 40% cheaper than Opus 5 · Fable 5.1 costs about 25% less than Fable 5 on typical workloads, up to 45% less on heavy coding and agent work</li><li><strong>Efficiency</strong> — Figure: Q2 revenue +1,368% on a year earlier, compute +300% · Revenue per employee $6.6M (trailing year), from $1.4M</li><li><strong>Capacity</strong> — Figure: “Demand for compute has historically exceeded available capacity” · Across research, serving and internal use, which it runs as one pool. It has asked clouds for more compute than they would provide, and has changed usage and rate limits</li></ul>
<h2>Customers paying $100K+ a year</h2>
<figure><a href="https://06.ai/packs/anthropic-s1#customers"><img src="https://06.ai/newsletter/anthropic-s1/fig-02.png" alt="Customers paying $100K+ a year" width="694" height="272"></a><figcaption>Trailing-twelve-month revenue of $100K or more, the S-1’s measure of enterprise adoption. The end-2025 count is Daloopa’s read of the filing; the June counts are DataLearner’s.</figcaption></figure>
<h2>250 of about 15,000</h2>
<p>How many companies could spend $10M a year on AI if they spent a given share of revenue on it, and what that spend adds up to. A company spending 1% reaches $10M at $1B of revenue; at 5%, at $200M. Company counts are a power law fitted to the 2026 Fortune Global 500 and McKinsey’s count of companies with $1B+ of revenue; the range is 10,000 to 20,000 such companies.</p>
<p>Revenue to spend $10M <strong>$1B</strong> · Companies that big <strong>15,000 </strong><em>10,000–20,000</em> · AI spend a year <strong>$0.97T </strong><em>$0.82–1.10T</em> · Anthropic’s 250, share <strong>1.7%</strong> · Anthropic’s run-rate, share <strong>10.0%</strong></p>
<figure><a href="https://06.ai/packs/anthropic-s1#tam"><img src="https://06.ai/newsletter/anthropic-s1/fig-03.png" alt="250 of about 15,000" width="694" height="402"></a><figcaption>Shown: Share of revenue spent on AI 1%.</figcaption></figure>
<figure><a href="https://06.ai/packs/anthropic-s1#tam"><img src="https://06.ai/newsletter/anthropic-s1/fig-04.png" alt="Revenue needed to spend $10M — At 1% of revenue: $1B · At 5% of revenue: $200M; Companies that big — At 1% of revenue: 15,000 10,000–20,000 · At 5% of revenue: 72,000 40,000–109,000; Their revenue — At 1% of revenue: $97T · At 5% of revenue: $120T; Spend on AI a year — At 1% of revenue: $0.97T $0.82T–$1.1T · At 5% of revenue: $6.0T $4.7T–$7.3T; Anthropic’s 250 $10M+ customers, share of the companies — At 1% of revenue: 1.7% · At 5% of revenue: 0.35%; Anthropic’s $96B run-rate, share of the spend — At 1% of revenue: 10% · At 5% of revenue: 1.6%" width="694" height="387"></a></figure>
<h2>How far along is that?</h2>
<p><strong>About 2% of the way, counted by customers.</strong> If every company with $1B of revenue put 1% of it into AI, about 15,000 would spend $10M a year. Anthropic has about 250 at that level, 1.7%. At 5% the field widens to about 72,000 companies with $200M of revenue or more, and Anthropic has 0.35% of them.</p>
<p><strong>The real share is probably lower.</strong> The S-1 doesn’t say who the 250 are. Some are likely AI-native companies whose products run on Claude, which aren’t among the big companies counted here. Counting only companies with $1B of revenue, the share paying Anthropic $10M a year is below 1.7%.</p>
<p><strong>Counted in dollars, further along, but not far.</strong> Anthropic’s run-rate of about $96B is 10% of the $0.97T a year that 1% of big-company revenue would be, and 1.6% of the $6.0T at 5%. All AI model spend, Anthropic’s and OpenAI’s together, comes to about 0.16% of the revenue of the world’s $1B+ companies today, and not all of it comes from them.</p>
<p><strong>Is 5% a stretch?</strong> Worldwide IT spending is about $6.3T this year, by Gartner’s count; 5% of the revenue of every company that could spend $10M on AI is about the same, so 5% would make AI about as big as all of IT. The S-1 puts the work of knowledge workers at more than $40T a year, and 1% of revenue is small next to a payroll. Nothing in the S-1 says where spending settles, but it isn’t near either number yet.</p>
<h2>Each dollar of demand takes less compute, and so far Anthropic has kept the savings</h2>
<figure><a href="https://06.ai/packs/anthropic-s1#efficiency"><img src="https://06.ai/newsletter/anthropic-s1/fig-05.png" alt="Each dollar of demand takes less compute, and so far Anthropic has kept the savings" width="694" height="368"></a><figcaption>Revenue per $1 of compute and infrastructure, research and training included, by quarter. From Q4 2026, our projection: either research and training keep about three-quarters of compute, as in Q3, or Anthropic moves compute to serving customers until serving takes 40%. The gross margin stays at 85% in both.</figcaption></figure>
<p>Cost of revenue per $1 of revenue, and the share of compute spending it takes. 2024 and H1 2025 are backed out of the margins in the S-1’s note on expenses by function; Q3 2026 is the 85% gross margin Anthropic gave investors.</p>
<figure><a href="https://06.ai/packs/anthropic-s1#efficiency"><img src="https://06.ai/newsletter/anthropic-s1/fig-06.png" alt="Each dollar of demand takes less compute, and so far Anthropic has kept the savings" width="694" height="308"></a></figure>
<p>Index, first half of 2025 = 1, log scale. The cost of serving a token is the average list price paid per Claude token on OpenRouter times one minus the gross margin: a rough proxy, which assumes Anthropic’s mix of models and discounts moves like OpenRouter’s. From Q4 2026, our projection: 2× a year, with 1.5× and 3× as the band. The last year ran at 4.7×.</p>
<figure><a href="https://06.ai/packs/anthropic-s1#efficiency"><img src="https://06.ai/newsletter/anthropic-s1/fig-07.png" alt="Each dollar of demand takes less compute, and so far Anthropic has kept the savings" width="694" height="322"></a></figure>
<h2>Where the efficiency comes from, and where it goes</h2>
<p><strong>Serving got 6× cheaper per dollar of revenue in a year.</strong> Cost of revenue was 88¢ per $1 of revenue in the first half of 2025, 51¢ in the second and 15¢ in the first half of 2026, an 85% gross margin Anthropic kept in the third quarter. On OpenRouter, the average list price paid per Claude token rose over the same period, from $3.03 to $3.87 per million as usage moved to Opus. Put the two together and the cost of serving a Claude token fell about 4.7× in a year. The S-1 credits its mix of chips (Amazon’s Trainium, Google’s TPUs, NVIDIA and AMD) and what it calls compute multipliers, improvements in training and serving that it says Claude now helps its researchers find faster.</p>
<p><strong>So far it has kept the savings.</strong> The gross margin went from 12% to 85% in a year while the price paid per token rose. That is how revenue grew 12× in a year while compute grew 4.8×. It can’t repeat: at 15¢ of cost per dollar of revenue, there is little margin left to gain.</p>
<p><strong>From here, the savings go to prices.</strong> The S-1 says efficiency lets Anthropic cut what customers pay “while preserving our margins,” and October shows it starting. Opus 5.5 does a task for about 40% less than Opus 5 did, and lists at $4 and $20 per million input and output tokens, against $15 and $75 for Opus 4.1 fourteen months ago; Haiku 5.5 lists at a tenth of Haiku 4.5’s price. On OpenRouter, the average price paid per Claude token fell 30% in October from the third quarter. Revenue per dollar of compute serving customers has held near $6.70 since the start of 2026, and we assume it stays there.</p>
<p><strong>Research takes most of the compute.</strong> Serving customers took 25.5% of compute and infrastructure in the third quarter, from 43.6% in the first half of 2025; the rest goes to research, training and Anthropic’s own use of Claude. Revenue per $1 of all compute rose from $0.40 in the first quarter of 2025 to $2.08 in the second quarter of 2026, then fell to $1.70 in the third, as compute grew 78% in a quarter, mostly for research.</p>
<p><strong>Our projection.</strong> We expect the tokens Anthropic serves per dollar to keep doubling each year, 45× the first half of 2025 by the end of 2029 (1.5× to 3× a year gives 17× to 175×), slower than the last year’s 4.7×. And with the margin near its ceiling, we expect Anthropic to pass nearly all of it to customers, as it began to in October, so revenue per gigawatt rises only as fast as what a gigawatt costs, which we take as about 5% a year. Revenue per $1 of compute then stays at $1.70 if research keeps its share, or rises to $2.67 if Anthropic shifts compute to serving until serving takes 40%. Moving compute from research is the biggest lever it has: each point of compute moved to serving adds about 4% to the revenue it can serve.</p>
<h2>Demand outruns the compute on every path but Anthropic’s own plan</h2>
<p>Q4 2027 average, at an annual rate: the compute can serve $213B, 24% short of the reported trend ($281B) and 44% short of our tracker ($379B). Q4 2029: $613B, against $841B and $1.42T.</p>
<figure><a href="https://06.ai/packs/anthropic-s1#demand"><img src="https://06.ai/newsletter/anthropic-s1/fig-08.png" alt="Demand outruns the compute on every path but Anthropic’s own plan" width="694" height="496"></a><figcaption>Annualized revenue by quarter (quarterly revenue × 4), $B, log scale. To Q3 2026 from the S-1 and investor day. Three demand paths from there: <strong>our tracker</strong> (06.ai’s revenue tracker as published, weighted to OpenRouter) and <strong>the reported trend</strong> (Q3’s 45% growth over Q2, from Q3’s revenue), both with growth halving every 18 months, and <strong>Anthropic’s plan</strong> ($110B a year at the end of 2026, about $195B of revenue in 2028, steady growth in between and after). Against them, the revenue the compute Anthropic has lined up can serve. The text quotes run-rates on December 31, as the tracker does; while revenue grows they sit above the fourth quarter’s average, by 16–20% at the end of 2026 and 4–12% at the end of 2027.</figcaption></figure>
<p>What each demand path needs to be served with today’s share of compute serving customers (or 40%, with the switch above), against what Anthropic has lined up: base, with the late and on-time cases as the band.</p>
<figure><a href="https://06.ai/packs/anthropic-s1#demand"><img src="https://06.ai/newsletter/anthropic-s1/fig-09.png" alt="Demand outruns the compute on every path but Anthropic’s own plan" width="694" height="390"></a></figure>
<figure><a href="https://06.ai/packs/anthropic-s1#demand"><img src="https://06.ai/newsletter/anthropic-s1/fig-10.png" alt="Demand outruns the compute on every path but Anthropic’s own plan — Shift to 40%" width="694" height="527"></a><figcaption><strong>Shift to 40%</strong>: Q4 2027 average, at an annual rate: the compute can serve $322B, 14% more than the reported trend ($281B) and 15% short of our tracker ($379B). Q4 2029: $961B, against $841B and $1.42T.</figcaption></figure>
<h2>Short by the end of 2027, on two of three paths</h2>
<p><strong>On the reported trend, Anthropic needs about 13 GW at the end of 2027</strong>, a third more than the roughly 10 GW it has lined up, and 34 GW at the end of 2029 against about 25. At today’s split between research and serving, its compute can serve about $231B a year of revenue at the end of 2027, 25% short of the trend’s $309B. Our tracker’s faster path, $425B, would need 18 GW.</p>
<p><strong>Both paths start above what Anthropic expects.</strong> The reported trend ends this year at $113B and our tracker at $126B. At its investor presentation the company expected about $110B (Reuters), close to the $110B we estimate its compute can serve on December 31; the FT reported annualized revenue was expected to stay below $100B, without giving a date. Slower demand, a cautious forecast or compute that is already binding would each explain the gap, and the filing doesn’t say which.</p>
<p><strong>Only Anthropic’s plan fits.</strong> $110B at the end of this year and about $195B of revenue in 2028 would leave its compute 44% spare at the end of 2027. On reported plans, Anthropic is lining up compute for more demand than the plan behind its IPO valuation shows.</p>
<p><strong>It can close much of the gap by moving compute from research to customers.</strong> Serving takes only about a quarter of its compute. At 40% the same gigawatts serve about $363B at the end of 2027, more than the reported trend needs, at the cost of slower model progress, which the S-1 itself lists as a risk. The other valve is price: if Anthropic held prices instead of passing the next round of efficiency on, each gigawatt would earn about twice as much a year later. Either way a shortage shows up as rate limits, rationing and prices that stop falling, and demand bends down toward the compute line.</p>
<p><strong>The answer leans on one estimate:</strong> how much compute Anthropic ran in the third quarter. We take about 3.1 GW, on the way from about 1.4 GW at the end of 2025 to 5 GW at the end of 2026, which makes a gigawatt worth about $22B of revenue a year. It is the softest number here: on that path, compute spending per gigawatt rose from $9.8B a year in the second quarter to $12.7B in the third, 29% in one quarter. At the second quarter’s cost, Q3 ran on about 4.0 GW, and the 10 GW lined up would serve about $179B at the end of 2027: 42% short of the reported trend, and only 12% spare on the plan. Starting from the 2 GW 22V Research estimated for early 2026 gives $203B. Our assumption that Q3’s serving compute was fully used cuts the other way: if it had room to spare, the same gigawatts serve more.</p>
<h2>At 1% of revenue the buildout is enough. At 5% it isn’t close</h2>
<figure><a href="https://06.ai/packs/anthropic-s1#market"><img src="https://06.ai/newsletter/anthropic-s1/fig-11.png" alt="At 1% of revenue the buildout is enough. At 5% it isn’t close" width="694" height="332"></a><figcaption>Gigawatts it would take to serve AI spend at 1%, 2% and 5% of the revenue of companies big enough to spend $10M, at Anthropic’s Q3 revenue per gigawatt, against the AI capacity installed worldwide at each year-end (the <a href="https://06.ai/packs/ai-compute-balance">compute-balance</a> pack’s base case). The gigawatts carry research and training in Anthropic’s proportions; serving customers alone would take about a quarter of them. Company revenue and revenue per gigawatt both grow about 5% a year, so the gigawatts needed stay flat while capacity grows.</figcaption></figure>
<h2>The question for the $1T+ raise is compute, not demand</h2>
<p><strong>The market’s worry runs the other way.</strong> The hyperscalers, neoclouds and labs need about $1.5T of outside money through 2030 (<a href="https://06.ai/packs/ai-funding-scale">the </a><a href="https://06.ai/packs/ai-funding-scale">funding-scale</a><a href="https://06.ai/packs/ai-funding-scale"> pack</a>), and the fear is that the demand won’t be there to pay it back. Sized against company revenue, demand at 1% would need about 44 GW, research and training included, a quarter of the 175 GW of AI capacity the world has by the end of 2029. At 2% it needs 97 GW, more than half, with consumer apps and advertising still to serve. At 5% it needs 272 GW, more than all of it.</p>
<p><strong>The S-1 points up the range, not down.</strong> Customers billed a year ago spend nine times as much; the number spending $100K a year grew 12.5×; and Anthropic, which knows its pipeline, is lining up compute, on reported plans, well beyond what its own IPO plan needs. <a href="https://06.ai/packs/ai-compute-balance">The </a><a href="https://06.ai/packs/ai-compute-balance">compute-balance</a><a href="https://06.ai/packs/ai-compute-balance"> pack</a> has the whole market about 19% short at the end of this year and about level at the end of 2027 before demand pulls ahead again, on demand that doesn’t assume any particular share of revenue.</p>
<p><strong>What would make us wrong.</strong> Growth slowed to 45% in the third quarter, and some of the cushion was distillation revenue Anthropic chose to give up. Five customers brought in about a quarter of revenue in the nine months to September, and about a fifth in the third quarter; the S-1 doesn’t name them. And demand is price-elastic: the S-1 credits lower prices per task for driving usage, so if a shortage keeps prices up, demand grows more slowly. A shortage would look like that: slower growth at high prices rather than a glut.</p>
<h2>What the S-1 changes for our end-2027 forecast</h2>
<p>Anthropic’s annualized revenue at the end of 2026 and 2027, $B. The tracker is unchanged; these are the reasons it differs from this pack and changes we’d consider.</p>
<figure><a href="https://06.ai/packs/anthropic-s1#tracker"><img src="https://06.ai/newsletter/anthropic-s1/fig-12.png" alt="Our tracker — End-2026: $126B · End-2027: $425B $243B–$790B · Basis: Reported figures and OpenRouter spend, growth 14.7% a month fading; Reported trend — End-2026: $113B · End-2027: $309B · Basis: Q3 revenue and its 45% growth over Q2, the same fade; Anthropic’s plan — End-2026: $110B · End-2027: $160B · Basis: $110B at end-2026 (Reuters), about $195B of revenue in 2028, no fade; What the compute can serve — End-2026: $110B · End-2027: $231B–$363B · Basis: 10 GW, research keeping its share or serving rising to 40%" width="694" height="343"></a></figure>
<p>Why the two models differ</p>
<ul><li><strong>Where they start.</strong> The tracker’s $96B for October 30 leans on OpenRouter’s October, the first full month of Opus 5.5, when spend on Claude there rose about 80% on September. It ends 2026 at $126B, above the $110B the company expected at its investor presentation (Reuters) and the FT’s “below $100 billion.”</li><li><strong>How fast they grow.</strong> The tracker grows 14.7% a month from its last three months; the last reported step, Q2 to Q3, is 13.3% a month.</li><li><strong>Supply.</strong> The tracker has no ceiling. Its end-2027 figure needs about 18 GW at today’s split, against about 10 GW lined up.</li><li><strong>Basis.</strong> 48% of Anthropic’s revenue comes through the clouds’ marketplaces. Whether it books that revenue gross, with the clouds’ fees as a cost, is unconfirmed; if it does, the tracker’s head-to-head with OpenAI isn’t like-for-like.</li></ul>
<p>Changes we’d consider, not made</p>
<ul><li>Load the S-1’s quarterly revenue, replacing three estimates (Q1 2025 was $0.376B, not the $0.33B the tracker inferred).</li><li>Log the $110B year-end expectation Reuters reported next to the FT’s under-$100B, already logged as a target, and use them as a soft ceiling for December.</li><li>Add a compute-capped line: run-rate no higher than the gigawatts lined up times revenue per gigawatt.</li><li>Weight OpenRouter less in the month after a major launch: launches drive revenue, but OpenRouter’s users switch faster than contracts do.</li><li>Adjust for the holidays: customer spend fell about 20% in the two weeks to January 2.</li><li>If marketplace revenue turns out to be booked gross, publish a net-basis Anthropic series alongside, for the comparison with OpenAI.</li></ul>
<h2>What this is anchored on</h2>
<p>Where the S-1 leaves a figure blank or we couldn’t read a section directly, the source is a named reading of the filing.</p>
<ul><li><strong>Anthropic S-1</strong> — Revenue and compute and infrastructure expense by year and half-year; customers by spend tier; net revenue retention; revenue per employee; revenue vs compute growth; price per task; compute risk factors; knowledge work worth over $40T · <a href="https://www.sec.gov/Archives/edgar/data/0002130357/000162828026067168/anthropicpbc-sx1.htm">SEC, filed October 26, 2026</a></li><li><strong>S-1, quarterly</strong> — Revenue and compute and infrastructure by quarter, Q1 2025–Q2 2026 (they sum to the audited totals); revenue mix; purchase commitments by counterparty · <a href="https://www.allocations.com/insights/anthropic-quarterly-revenue-what-the-s-1-shows">Allocations, reading the S-1</a></li><li><strong>S-1, by function</strong> — Cost of revenue $2,773M in 2025 and $2,380M in H1 2026; gross margin about 40% and 85% · <a href="https://tomtunguz.com/anthropics-s1-grew-eleven-fold/">Tomasz Tunguz</a></li><li><strong>S-1, details</strong> — Margins by period (2024 −94.8%, H1 2025 12.4%); marketplace share of revenue by year; $265.6B of fixed or minimum compute commitments at June 30 · <a href="https://www.alpha-sense.com/api/html-embed/file-4c88b5685d7120f5cdcdc99545773cca679d25fd-html/">AlphaSense primer</a></li><li><strong>S-1, details</strong> — $100K+ customers about 1,500 at end-2025; multi-product customers; concentration by quarter; rule-breaking revenue 5–7%; over 80% of Q3 revenue from new models; nearly $590B of commitments · <a href="https://daloopa.com/?p=17863">Daloopa</a></li><li><strong>Investor day, Q3</strong> — $16.8B of Q3 revenue; $1.70 of revenue per $1 of compute; adjusted operating margin 11%; distillation 5–7% of revenue · <a href="https://simonwillison.net/2026/Oct/15/anthropic-tells-investors/">FT, via Simon Willison, Oct 15</a></li><li><strong>Investor day, margin</strong> — 85% gross margin (GAAP, per Bloomberg’s Ed Ludlow); briefing on October 14 · <a href="https://fourweekmba.com/ai-anthropic-tells-investors-q3-revenue-rose-45-to-16-8b/">FourWeekMBA</a></li><li><strong>Year-end expectation</strong> — Q3 revenue up 45% to $16.8B; the company “was expecting an annualised revenue run rate of $110 billion by the end of the year,” per a person at the investor presentation · <a href="https://www.business-standard.com/world-news/anthropic-s-revenue-rises-45-to-16-8-billion-in-q3-ahead-of-ipo-126101500124_1.html">Reuters, via Business Standard, Oct 15</a></li><li><strong>Year-end expectation</strong> — Investors expect $110B of annualized revenue by the end of the year · <a href="https://www.pymnts.com/news/investment-tracker/ipo/2026/anthropic-quarterly-revenue-jumps-45-before-planned-ipo/">Wall Street Journal, via PYMNTS</a></li><li><strong>Year-end expectation</strong> — Annualized revenue “expected to remain below the closely watched $100 billion threshold” (no date given) · <a href="https://finance.yahoo.com/technology/ai/articles/anthropic-reports-45-revenue-growth-102235753.html">Financial Times, via Yahoo Finance, Oct 15</a></li><li><strong>IPO plan</strong> — 2028 revenue of roughly $190–200B underpinning the IPO valuation · <a href="https://finance.yahoo.com/technology/ai/articles/exclusive-anthropic-ipo-valuation-hinges-001158885.html">Reuters, via Yahoo Finance, Aug 14</a></li><li><strong>Compute plan</strong> — About 5 GW available by the end of 2026, roughly double by the end of 2027 · <a href="https://www.newsquawk.com/headlines/anthropic-reportedly-plans-to-have-5gw-worth-of-compute-by-the-end-of-2026-reports-nyt">New York Times, via Newsquawk, Sep 18</a></li><li><strong>Compute plan</strong> — 14.8 GW locked in since October 2025, worth up to $517B; AWS Trainium 5 GW, Google/Broadcom TPU 5 GW · <a href="https://aiweekly.co/alerts/the-information-anthropic-has-locked-in-148-gw-of-compute-worth-up-to-517b">The Information, via AI Weekly, Sep 6</a></li><li><strong>Compute plan</strong> — Up to 5 GW with Amazon, nearly 1 GW of it by the end of 2026 · <a href="https://www.anthropic.com/news/anthropic-amazon-compute">Anthropic</a></li><li><strong>Compute plan</strong> — About 2 GW early in 2026; about 5 GW placed with Broadcom for 2027 and 10 GW for 2028; over 30 GW by 2030 · <a href="https://06.ai/packs/ai-compute-balance">22V Research, Sep 4 (via </a><a href="https://06.ai/packs/ai-compute-balance">the </a><a href="https://06.ai/packs/ai-compute-balance">compute-balance</a><a href="https://06.ai/packs/ai-compute-balance"> pack</a><a href="https://06.ai/packs/ai-compute-balance">)</a></li><li><strong>Compute today</strong> — About 1.4 GW at the end of 2025 (press reports of investor discussions) · <a href="https://fourweekmba.com/ai-anthropic-gigawatt-threshold-compute-unit-of-account/">FourWeekMBA</a></li><li><strong>Large companies</strong> — About 8,000 companies with $1B+ of revenue, and 7,000 more by 2025 · <a href="https://www.carriermanagement.com/news/2013/10/02/113901.htm">McKinsey Global Institute, via Carrier Management</a></li><li><strong>Large companies</strong> — More than 28,000 companies with $200M+ of revenue in its sample (17,000 listed, 11,400 private) · <a href="https://www.mckinsey.com/capabilities/strategy-and-corporate-finance/our-insights/the-new-global-competition-for-corporate-profits">McKinsey Global Institute</a></li><li><strong>Large companies</strong> — Fortune Global 500 2026: $43.1T of revenue, $33.2B to make the list · <a href="https://www.energytrend.com/print/51848">EnergyTrend</a></li><li><strong>Large companies</strong> — Forbes Global 2000 2026: $56T of sales · <a href="https://logos-pres.md/en/news/companies-set-new-records-on-the-forbes-global-2000/">Forbes, via Logos Press</a></li><li><strong>IT spending</strong> — Worldwide IT spending of $6.31T in 2026 · <a href="https://www.businesswire.com/news/home/20260422301495/en/">Gartner, Apr 22</a></li><li><strong>Run-rates now</strong> — Anthropic $96B and OpenAI $57B on 2026-10-30 · <a href="https://06.ai/trackers/openai-anthropic-revenue">06.ai revenue tracker</a></li><li><strong>Claude token prices</strong> — Average list price paid per million Claude tokens, by period; list prices by model (Opus 4.1 $15/$75, Opus 5.5 $4/$20, Haiku 4.5 $1/$5, Haiku 5.5 $0.10/$0.50) · <a href="https://openrouter.ai/rankings">OpenRouter rankings and catalog (CC BY 4.0), priced as in </a><a href="https://openrouter.ai/rankings">the </a><a href="https://openrouter.ai/rankings">AI-demand</a><a href="https://openrouter.ai/rankings"> pack</a></li><li><strong>S-1, details</strong> — Customer spend fell about 20% in the two weeks to January 2, 2026; Fable 5.1’s cost on typical workloads · <a href="https://www.buildmvpfast.com/blog/anthropic-s1-public-filing-claude-builders-2026">BuildMVPFast, reading the S-1</a></li><li><strong>S-1, details</strong> — $100K+ customers about 600 at June 30, 2025 and about 6,000 at June 30, 2026 · <a href="https://www.datalearner.com/ai-organizations/anthropic/s-1">DataLearner, reading the S-1</a></li><li><strong>World AI capacity</strong> — 40 GW operating in September 2026; switched on each year to 2030, base case · <a href="https://06.ai/packs/ai-compute-balance">06.ai </a><a href="https://06.ai/packs/ai-compute-balance">compute-balance</a><a href="https://06.ai/packs/ai-compute-balance"> pack</a></li></ul>
<h2>Method</h2>
<p><strong>Revenue and compute</strong> are the S-1’s, by quarter to June 2026. Compute is its “compute and infrastructure” expense, which covers training and research as well as serving customers. Cost of revenue, the compute that serves customers plus a little else, comes from the S-1’s note on expenses by function. Q3 2026 is blank in the filing: revenue is the $16.8B Anthropic gave investors, compute is revenue divided by the $1.70 per $1 it gave, and the gross margin is the 85% it gave. We take the $1.70, which the FT reported per dollar of “computing power,” to be measured against the S-1’s compute and infrastructure line; if it is narrower, Q3’s compute was larger, serving’s share of it smaller and a gigawatt’s cost higher.</p>
<p><strong>Company counts</strong> are a power law through two points: 500 companies with $33.2B of revenue or more (the 2026 Fortune Global 500, $43.1T between them) and about 15,000 with $1B or more (McKinsey’s projection for 2025; 10,000 and 20,000 for the range). That makes about 72,000 companies with $200M or more, against the 28,000 McKinsey found in its own sample of listed and private companies a decade ago, and gives $97T of revenue for companies with $1B or more, more than the Global 2000’s $56T because it includes private and smaller companies. Spend at a share of revenue counts every company big enough to reach $10M at that share.</p>
<p><strong>Demand paths</strong> are run-rates by day. Our tracker’s path is its published forecast ($425B at the end of 2027), carried forward with its own formula: growth starts at 14.7% a month and the monthly rate halves every 18 months. The reported trend starts from Q3’s revenue at mid-quarter, $67.2B a year, and grows 13.3% a month, Q3’s growth over Q2, halving the same way. That puts it at $81B on September 30, where the tracker’s weekly estimates also were ($81B). Q3 is blank in the S-1, so the trend’s last step comes from investor day. Anthropic’s plan rises to $110B on December 31, the run-rate the company expected at its investor presentation (Reuters; the WSJ reported investors expect the same, while the FT reported revenue was expected to stay below $100B, with no date). After that it grows a steady 3.2% a month with no fade, the rate that gives $195B of revenue in 2028. Quarterly revenue is the run-rate summed over each quarter.</p>
<p><strong>Compute lined up</strong> is all-in gigawatts at each year-end: 1.4 at the end of 2025; 5, 10, 18 and 25 at the ends of 2026–29 in the base case; 4.5, 8, 13 and 18 if deliveries run late; 5, 12, 22 and 32 if the 14.8 GW The Information counts, AMD’s 2 GW and more TPUs arrive on time. The late and on-time cases split from the base after September 30, 2026. These are reported plans and our extensions, not figures in the S-1, which gives commitments in dollars and leaves the latest total blank. Readings of the filing put fixed or minimum compute commitments at $265.6B on June 30 (AlphaSense) and all commitments at nearly $590B (Daloopa), depending on how minimums, ceilings and leases are counted; the S-1 expects the part after 2030 to be incurred between 2031 and 2045. Between year-ends we interpolate geometrically.</p>
<p><strong>Revenue the compute can serve</strong> is Q3 2026’s revenue per gigawatt, about $22B a year on about 3.1 GW, times the gigawatts lined up, scaled by the share of compute serving customers: 25.5% in Q3, held, or rising to 40% by the end of 2027. Revenue per gigawatt rises 5% a year, our assumption for how much more a gigawatt of newer hardware costs; at a steady 85% margin, revenue per gigawatt moves with that cost, and the extra work each gigawatt does goes to customers as lower prices. We assume Q3’s serving compute was fully used, so Q3’s revenue is what it could serve; the S-1 says demand for compute has exceeded capacity across research, serving and its own use, but not that customer demand went unmet in Q3. Gigawatts needed are the same calculation run backward from each demand path. Compute spend is $12.7B per gigawatt a year in Q3, rising 5% a year with revenue per gigawatt: the base case spends about $96B on compute in 2027 and $312B in 2029, about $599B over the three years, about as much as the nearly $590B of commitments tallied so far, which stretch to 2045. Anthropic says it plans to secure more.</p>
<p><strong>The whole market</strong> is AI spend at each share of revenue, served at Anthropic’s Q3 revenue per gigawatt, research included, against the <a href="https://06.ai/packs/ai-compute-balance">compute-balance</a> pack’s base case: 40 GW operating in September 2026 plus what is switched on each year. The gigawatts carry research and training in Anthropic’s Q3 proportions; serving alone is about a quarter of them, 11, 25 and 69 GW at 1%, 2% and 5%.</p>
<hr>
<p>Every chart above links to its interactive version on 06.ai, with the data behind it. The daily trackers are at <a href="https://06.ai/trackers">06.ai/trackers</a>.</p>
<p><em>Not investment advice. Estimates are labeled as estimates.</em></p>]]></content:encoded>
</item>
<item>
  <title>Who owns the most Blackwell per dollar?</title>
  <link>https://inthesingularity.ai/packs/blackwell-per-dollar</link>
  <guid isPermaLink="true">https://inthesingularity.ai/packs/blackwell-per-dollar</guid>
  <pubDate>Wed, 28 Oct 2026 12:16:00 GMT</pubDate>
  <description>Nvidia’s Blackwell GPUs owned by each listed company per $1B of market cap, and who is lined up for the most of what comes next. The neoclouds lead by one to two orders of magnitude, mostly because they buy their GPUs with debt.</description>
  <category>AI infrastructure</category>
  <enclosure url="https://inthesingularity.ai/newsletter/blackwell-per-dollar/cover.png" length="61401" type="image/png" />
  <media:content url="https://inthesingularity.ai/newsletter/blackwell-per-dollar/cover.png" medium="image" type="image/png" />
  <content:encoded><![CDATA[<figure><a href="https://06.ai/packs/blackwell-per-dollar"><img src="https://06.ai/newsletter/blackwell-per-dollar/thumb.png" alt="Who owns the most Blackwell per dollar?" width="421" height="391"></a><figcaption><a href="https://06.ai/packs/blackwell-per-dollar">Interactive version and data on 06.ai →</a></figcaption></figure>
<p>CoreWeave owns about <strong>16,525</strong> Nvidia Blackwell GPUs per $1B of market cap, <strong>77×</strong> Microsoft’s 214. Most of that gap is debt: CoreWeave’s enterprise value is 2.4× its market cap, and per $1B of enterprise value IREN leads at 11,275. Amazon has the largest disclosed claim on what comes next, at least <strong>3M</strong> Nvidia GPUs through 2028; per dollar, the neoclouds’ unbuilt contracted power is bigger still, if they can finance it.</p>
<ul><li><strong>16,525</strong> — Blackwell GPUs per $1B of market cap, CoreWeave. Microsoft 214, Alphabet 86. Nvidia GPUs only, estimates</li><li><strong>11,275</strong> — Per $1B of enterprise value, IREN. CoreWeave falls to 6,882: its enterprise value is 2.4× its market cap</li><li><strong>846K</strong> — Largest Blackwell owner, Microsoft. B200 and B300 shipped through Q1 2026, per Epoch AI. Meta 519K, Amazon 401K</li><li><strong>3M+</strong> — Largest disclosed next allocation, AWS. Nvidia GPUs: more than 1M from 2026, then 2M more through 2027–28</li></ul>
<h2>Blackwell GPUs per $1B of market cap</h2>
<figure><a href="https://06.ai/packs/blackwell-per-dollar#per-dollar"><img src="https://06.ai/newsletter/blackwell-per-dollar/fig-01.png" alt="Blackwell GPUs per $1B of market cap" width="694" height="376"></a><figcaption>Base estimate, with the range from low to high. Log scale. Market caps at the October 26–27, 2026 closes. Nvidia GPUs only: Alphabet’s TPUs and Amazon’s Trainium aren’t counted.</figcaption></figure>
<h2>Why the neoclouds top the list</h2>
<p><strong>It’s a leverage screen.</strong> GPUs per dollar of market cap rise as the equity shrinks against the fleet. CoreWeave’s roughly 543K Blackwell GPUs sit under $78.9B of enterprise value but $32.9B of market cap; the rest is debt and leases, $31.4B of it recourse debt at June 30. IREN and Nebius carry less (1.16× and 1.05× their market caps), so per dollar of enterprise value IREN moves to first and CoreWeave’s figure falls by more than half, to 6,882.</p>
<p><strong>And the market has marked the equity down.</strong> CoreWeave closed at $59.57 on October 27, against a 52-week high of $141.79; its credit default swaps peaked at 855bp in the run-up to its second-quarter results. IREN closed at $27.44, against $69.58 the night it announced its Nvidia partnership in May. Part of the per-dollar lead is the market discounting the debt behind the fleets.</p>
<p><strong>Owning the GPU isn’t owning the upside.</strong> Most neocloud capacity is contracted out at fixed prices: Microsoft at IREN and Nebius, Meta at Nebius and at CoreWeave, where it signed $21B through 2032 on top of $14B. Shareholders earn what’s left after interest and depreciation, and the customer keeps the margin on what runs on the chips. CoreWeave’s adjusted operating margin was 5% in the second quarter.</p>
<h2>Blackwell GPUs owned</h2>
<figure><a href="https://06.ai/packs/blackwell-per-dollar#owned"><img src="https://06.ai/newsletter/blackwell-per-dollar/fig-02.png" alt="Blackwell GPUs owned" width="694" height="346"></a><figcaption>The same estimates in units. Hyperscaler counts run through Q1 2026; the high end adds two more quarters at Q4 2025’s pace.</figcaption></figure>
<h2>Blackwell shipped by quarter</h2>
<figure><a href="https://06.ai/packs/blackwell-per-dollar#quarterly"><img src="https://06.ai/newsletter/blackwell-per-dollar/fig-03.png" alt="Blackwell shipped by quarter" width="694" height="380"></a><figcaption>Epoch AI’s estimate, B200 and B300 by owner, on one scale. Q1 2026 (paler) is incomplete in the source. Other is everyone Epoch doesn’t name in this series: neoclouds, sovereigns and enterprises. Oracle is counted from its Abilene site instead.</figcaption></figure>
<h2>Who owns the most</h2>
<p><strong>Microsoft, by Epoch AI’s estimate:</strong> about 846K B200 and B300 through the first quarter, ahead of Meta (519K), Amazon (401K) and Alphabet (332K). Epoch’s estimate for Microsoft’s Fairwater Atlanta site alone is about 304K.</p>
<p><strong>The biggest single site is xAI’s.</strong> Epoch puts Colossus 2 at 1.112 million H100-equivalents, about 440K Blackwell GPUs, owned by SpaceX since it bought xAI in February. Abilene, the site Oracle runs for OpenAI, is about 201K.</p>
<p><strong>The counts lag the shipments.</strong> Epoch’s owner series ends in the first quarter and Nvidia has shipped two more quarters since, so every hyperscaler owns more than its base figure today. Two more quarters at the fourth quarter’s pace lift the hyperscalers’ ratios by about half, which doesn’t change the order: the lowest neocloud is still 12× the highest hyperscaler.</p>
<h2>Who is lined up for the next allocation, per $1B of market cap</h2>
<figure><a href="https://06.ai/packs/blackwell-per-dollar#next"><img src="https://06.ai/newsletter/blackwell-per-dollar/fig-04.png" alt="Who is lined up for the next allocation, per $1B of market cap" width="694" height="376"></a><figcaption>Commitments and unbuilt contracted power for deliveries from late 2026 on: mostly Vera Rubin, some Blackwell Ultra, in proportions nobody discloses. Power converted at 3 kW per GPU (2.6–3.4 for the range). Commitments given only as “millions” or “multi-gigawatt” are read at their floor, marked ≥. Log scale.</figcaption></figure>
<h2>What each forward figure rests on</h2>
<ul><li><strong>Nebius NBIS</strong> — GPUs: 1.33M · Per $1B: 31,997 · Announced: Aug 2026 · Target of 5 GW contracted by year-end, less the 0.8–1 GW it expects connected by then. Low: the 3 GW it targeted in February, less 1 GW · In Nvidia’s first Rubin cloud cohort; deployments from late 2026 or early 2027</li><li><strong>CoreWeave CRWV</strong> — GPUs: 900K · Per $1B: 27,397 · Announced: Aug 2026 · 2.7 GW contracted but not yet active: 4.2 GW contracted as of August 11 against 1.5 GW active · First production Vera Rubin NVL72 customer, Cognition, on September 30</li><li><strong>IREN IREN</strong> — GPUs: 167K · Per $1B: 15,015 · Announced: May &amp; Aug 2026 · The 500 MW of IT load it plans to add in 2027. Its Nvidia partnership covers up to 5 GW, a framework rather than orders · No Rubin orders disclosed</li><li><strong>Oracle ORCL</strong> — GPUs: ≥667K · Per $1B: ≥1,745 · Announced: Oct 27, 2026 · “Multi-gigawatt” Vera Rubin expansion on OCI, read at its floor of 2 GW · Vera Rubin NVL72 generally available on OCI now</li><li><strong>Meta META</strong> — GPUs: ≥2M · Per $1B: ≥1,242 · Announced: Feb 18, 2026 · Multiyear Nvidia deal for “millions” of Blackwell and Rubin GPUs, read at its floor of 2 million. No unit count or value disclosed · Vera CPUs from 2027; part of the capacity comes through CoreWeave and Crusoe</li><li><strong>Amazon (AWS) AMZN</strong> — GPUs: ≥3M · Per $1B: ≥1,115 · Announced: Mar 16 &amp; Aug 26, 2026 · More than 1 million Nvidia GPUs from 2026, then 2 million more through 2027–28, across Blackwell Ultra, Rubin and Rubin Ultra. The largest disclosed · In Nvidia’s first Rubin cloud cohort</li><li><strong>SpaceX (xAI) SPCX</strong> — GPUs: 530K · Per $1B: 265 · Announced: 2026 · Colossus build-out toward 1 million GPUs across its sites, less the roughly 470,000 Blackwell counted above · No Rubin volumes disclosed</li><li><strong>Microsoft MSFT</strong> — GPUs: ≥400K · Per $1B: ≥101 · Announced: Jun 2026 · Fairwater sites scaling to “hundreds of thousands” of Vera Rubin superchips, two GPUs each, read at the floor. Rubin it rents from neoclouds, such as 30,000 from Nscale, belongs to them and isn’t counted · In Nvidia’s first Rubin cloud cohort</li><li><strong>Alphabet GOOGL</strong> · No committed Nvidia volume disclosed. Google Cloud’s network design links up to 960,000 Rubin GPUs, a capability rather than an order · In Nvidia’s first Rubin cloud cohort</li></ul>
<h2>Who gets Rubin first</h2>
<p>Nvidia named eight clouds first in line for Vera Rubin: AWS, Google Cloud, Microsoft Azure, Oracle Cloud, CoreWeave, Lambda, Nebius and Nscale. Production shipments began this fall. CoreWeave has had the first customer in production, Cognition, since September 30, and Oracle made Vera Rubin generally available on its cloud on October 27, with a multi-gigawatt expansion to come. Nvidia’s 2026 Rubin output is estimated at only 200K–300K GPUs, and it has historically sent 60%–70% of a new generation’s first year to the hyperscalers, so most of these allocations land in 2027 and 2028.</p>
<p><strong>In GPUs, Amazon and Meta.</strong> AWS committed to more than 1 million Nvidia GPUs from 2026 and 2 million more through 2027–28, the largest disclosed. Meta’s February deal covers “millions” of Blackwell and Rubin GPUs, with no number given. Microsoft is building Fairwater for hundreds of thousands of Rubin superchips and also rents Rubin from neoclouds, 30,000 of them from Nscale in Norway.</p>
<p><strong>Per dollar, the neoclouds, on paper.</strong> CoreWeave has 2.7 GW contracted that isn’t built, Nebius is aiming for 5 GW contracted by year-end, and IREN plans 500 MW more in 2027. That is about 31,997 next-generation GPUs per $1B of Nebius’s market cap and 27,397 per $1B of CoreWeave’s. But capex guided for this year is 1.13× CoreWeave’s market cap and 0.54× Nebius’s, and IREN’s for its fiscal 2027 is 2.48× its own. The next allocation will be paid for by lenders and by customers’ prepayments, not by today’s shareholders.</p>
<h2>The ranking, with leverage</h2>
<p>Sorted by GPUs per $1B of market cap. Enterprise value and guided capex for the four names where debt changes the picture.</p>
<ul><li><strong>CoreWeave CRWV · Neocloud</strong> — Blackwell GPUs: 543K 397K–600K · As of: Jun 2026 · Market cap: $32.9B · Per $1B mkt cap: 16,525 · EV: $78.9B · Per $1B EV: 6,882 · EV ÷ mkt cap: 2.40× · Capex ÷ mkt cap: 1.13×</li><li><strong>IREN IREN · Neocloud</strong> — Blackwell GPUs: 145K 36K–150K · As of: Year-end 2026 (secured) · Market cap: $11.1B · Per $1B mkt cap: 13,063 · EV: $12.9B · Per $1B EV: 11,275 · EV ÷ mkt cap: 1.16× · Capex ÷ mkt cap: 2.48×</li><li><strong>Nebius NBIS · Neocloud</strong> — Blackwell GPUs: 161K 100K–200K · As of: Jun 2026 · Market cap: $41.7B · Per $1B mkt cap: 3,863 · EV: $43.7B · Per $1B EV: 3,684 · EV ÷ mkt cap: 1.05× · Capex ÷ mkt cap: 0.54×</li><li><strong>Oracle ORCL · Leveraged builder</strong> — Blackwell GPUs: 201K 100K–450K · As of: Oct 2026 · Market cap: $382B · Per $1B mkt cap: 527 · EV: $470B · Per $1B EV: 428 · EV ÷ mkt cap: 1.23× · Capex ÷ mkt cap: 0.24×</li><li><strong>Meta META · Hyperscaler</strong> — Blackwell GPUs: 519K 477K–818K · As of: Q1 2026 · Market cap: $1.61T · Per $1B mkt cap: 322</li><li><strong>SpaceX (xAI) SPCX · Leveraged builder</strong> — Blackwell GPUs: 470K 140K–585K · As of: Oct 2026 · Market cap: $2T · Per $1B mkt cap: 235</li><li><strong>Microsoft MSFT · Hyperscaler</strong> — Blackwell GPUs: 846K 780K–1.32M · As of: Q1 2026 · Market cap: $3.95T · Per $1B mkt cap: 214</li><li><strong>Amazon AMZN · Hyperscaler</strong> — Blackwell GPUs: 401K 373K–607K · As of: Q1 2026 · Market cap: $2.69T · Per $1B mkt cap: 149</li><li><strong>Alphabet GOOGL · Hyperscaler</strong> — Blackwell GPUs: 332K 308K–509K · As of: Q1 2026 · Market cap: $3.87T · Per $1B mkt cap: 86</li></ul>
<h2>What a Blackwell allocation is worth</h2>
<p>Rubin is a big step: CoreWeave measured about 10× the tokens per megawatt of a GB200 NVL72 on DeepSeek R1 at matched interactivity. A Blackwell fleet contracted at today’s prices earns them until its contracts end. What it re-contracts for once Rubin is in volume is the main unknown in what it’s worth, and it matters most to owners who borrowed to buy it. <a href="https://06.ai/packs/inference-margins">The </a><a href="https://06.ai/packs/inference-margins">inference-margin</a><a href="https://06.ai/packs/inference-margins"> pack</a> has what a Blackwell GPU earns serving open models now, and <a href="https://06.ai/packs/neocloud-capex-funding">the </a><a href="https://06.ai/packs/neocloud-capex-funding">neocloud</a><a href="https://06.ai/packs/neocloud-capex-funding"> pack</a> what the neoclouds still have to raise.</p>
<p>So owning the allocation through neocloud equity is a levered position on fleet value and refinancing more than on AI margins. Per dollar of enterprise value, with the contracts and their counterparties in view, is the fairer comparison.</p>
<h2>How each count was built</h2>
<p>IREN’s count is the company’s own target. The rest are Epoch AI’s model, Epoch’s site estimates or our conversion of disclosed power, so treat them as estimates.</p>
<ul><li><strong>CoreWeave CRWV</strong> — Implied from power: 1.5 GW active on June 30, less the 360 MW its ~250,000 mostly-Hopper GPUs used at its IPO filing, at about 2.1 kW per Blackwell GPU. Range: 1.9–2.3 kW, and 20% of new power not Blackwell at the low end</li><li><strong>IREN IREN</strong> — Secured fleet: its 150,000-GPU target for year-end 2026 (over 50,000 B300 ordered in March plus Microsoft’s GB300), less a few thousand Hoppers. About a quarter is live: $1B of run-rate revenue against $4B contracted. The low end counts only that quarter</li><li><strong>Nebius NBIS</strong> — Implied from power, low confidence: 170 MW active at the end of 2025, scaled by run-rate revenue ($1.25B to $3B) to about 408 MW in June, less about 70 MW of Hopper, at about 2.1 kW per GPU</li><li><strong>Oracle ORCL</strong> — Abilene alone: Epoch AI’s 509,000 H100-equivalents for the site, about 201,000 Blackwell. Oracle says the Abilene supercluster has more than 100,000; Larry Ellison’s target for the site is more than 450,000. Other OCI regions aren’t counted</li><li><strong>Meta META</strong> — Epoch AI’s chip-owners model: B200 and B300 shipped through Q1 2026 (incomplete). Low: through Q4 2025. High: two more quarters at Q4 2025’s pace</li><li><strong>SpaceX (xAI) SPCX</strong> — Colossus 2: Epoch AI’s 1.112M H100-equivalents, about 440,000 Blackwell, plus 30,000 GB200 in Colossus 1. Low: Epoch’s owner data (140,000 added in Q3 2025). High: Colossus 2 at the reported 555,000</li><li><strong>Microsoft MSFT</strong> — Epoch AI’s chip-owners model, as for Meta. Its site estimate for Fairwater Atlanta alone is about 304,000</li><li><strong>Amazon AMZN</strong> — Epoch AI’s chip-owners model, as for Meta. Nvidia only: Trainium isn’t counted</li><li><strong>Alphabet GOOGL</strong> — Epoch AI’s chip-owners model, as for Meta. Nvidia only: Google’s TPUs, most of its compute, aren’t counted</li></ul>
<h2>What this is anchored on</h2>
<p>Market caps at the October 26–27, 2026 closes.</p>
<ul><li><strong>Hyperscaler Blackwell counts</strong> — B200 and B300 units shipped by owner and quarter, Q4 2024 to Q1 2026 (Q1 2026 incomplete) · <a href="https://epoch.ai/data/ai-chip-owners">Epoch AI, AI Chip Owners (updated May 12, 2026)</a></li><li><strong>Site estimates</strong> — Colossus 2 (SpaceX) 1.112M, Fairwater Atlanta (Microsoft) 769,000, Abilene (Oracle) 509,000 H100-equivalents · <a href="https://epoch.ai/data/ai-data-centers?view=table">Epoch AI, AI data centers (updated Oct 26, 2026)</a></li><li><strong>CoreWeave</strong> — 1.5 GW active and 3.7 GW contracted on June 30, 4.2 GW by August 11; backlog $104B; Meta $21B through 2032 on top of $14B; recourse debt $31.4B; CDS peaked at 855bp · <a href="https://www.techtimes.com/articles/324054/20260812/coreweave-q2-2026-129b-backlog-grew-296b-six-weeks-despite-50-default-odds.htm">TechTimes, Aug 12, 2026</a></li><li><strong>CoreWeave</strong> — Adjusted operating margin 5% in Q2; FY2026 capex guide $35–39B · <a href="https://futurumgroup.com/insights/coreweave-q2-fy-2026-ai-demand-drives-pricing-and-capacity-growth/">Futurum, Aug 2026</a></li><li><strong>CoreWeave baseline</strong> — About 250,000 GPUs, mostly Hopper, on 360 MW at its IPO filing · <a href="https://www.nextplatform.com/2025/03/05/coreweaves-250000-strong-gpu-fleet-undercuts-the-big-clouds/">The Next Platform, Mar 5, 2025</a></li><li><strong>CoreWeave, Rubin</strong> — First production Vera Rubin NVL72 customer (Cognition); about 10× the tokens per megawatt of GB200 NVL72 on DeepSeek R1 · <a href="https://www.coreweave.com/news/coreweave-delivers-nvidia-vera-rubin-nvl72-performance-at-production-scale-starting-with-cognition">CoreWeave, Sep 30, 2026</a></li><li><strong>Nebius</strong> — About 170 MW active at the end of 2025; run-rate revenue $1.25B; Microsoft and Meta capacity delivered · <a href="https://www.sec.gov/Archives/edgar/data/1513845/000110465926013946/tm266173d1_ex99-2.htm">Nebius Q4 2025 shareholder letter</a></li><li><strong>Nebius</strong> — Run-rate revenue $3B at June 30; 5 GW contracted-power target for year-end; FY2026 capex $20–25B; Rubin from late 2026 or early 2027 · <a href="https://www.fool.com/earnings/call-transcripts/2026/08/19/nebius-nbis-q2-2026-earnings-call-transcript/">Nebius Q2 2026 call</a></li><li><strong>IREN</strong> — 150,000-GPU target for year-end 2026; over 50,000 B300 ordered; Nvidia partnership up to 5 GW; $69.58 after hours on May 7 · <a href="https://finance.yahoo.com/sectors/technology/articles/iren-partners-nvidia-5-gw-205851949.html">Blockspace via Yahoo Finance, May 7, 2026</a></li><li><strong>IREN</strong> — $4B contracted run-rate revenue, $1B live after Microsoft accepted Horizon 1; 500 MW more IT load in 2027; FY2027 capex $25–30B · <a href="https://finance.yahoo.com/technology/ai/articles/iren-q4-earnings-call-highlights-230429246.html">Yahoo Finance, Aug 27, 2026</a></li><li><strong>Oracle</strong> — Vera Rubin NVL72 generally available on OCI; multi-gigawatt Vera Rubin expansion planned; more than 100,000 Blackwell in the Abilene supercluster · <a href="https://www.oracle.com/news/announcement/ai-world-nvidia-and-oracle-expand-nvidia-vera-rubin-to-multi-gigawatt-scale-2026-10-27/">Oracle, Oct 27, 2026</a></li><li><strong>Oracle</strong> — Abilene to house more than 450,000 GB200 eventually · <a href="https://www.datacenterdynamics.com/en/news/openai-and-oracle-to-deploy-450000-gb200-gpus-at-stargate-abilene-data-center/">DCD</a></li><li><strong>Oracle</strong> — Net debt $88.3B at August 31, 2026; FY2027 capex guide $90–95B · <a href="https://www.cestriancapitalresearch.com/oracle-corporation-nyse-orcl-q1-fy2027-earnings-record-cds-spreads-88b-net-debt-and-the-openai-question/">Cestrian Capital Research</a></li><li><strong>AWS</strong> — More than 1 million Nvidia GPUs over the next 12 months · <a href="https://www.datacenterdynamics.com/en/news/aws-to-deploy-more-than-one-million-nvidia-gpus-over-next-12-months/">DCD, Mar 16, 2026</a></li><li><strong>AWS</strong> — 2 million additional Nvidia GPUs through 2027–28: Blackwell Ultra, Rubin, Rubin Ultra · <a href="https://theenergymag.com/news/market-news/aws-and-nvidia-to-deliver-2-million-additional-gpus-and-next-generation-infrastructure-for-agentic-and-physical-ai">The Energy Mag, Aug 26, 2026</a></li><li><strong>Meta</strong> — Multiyear deal for millions of Blackwell and Rubin GPUs · <a href="https://pureai.com/Articles/2026/02/18/NVIDIA-expands-multiyear-AI-infrastructure-deal-with-Meta.aspx">PureAI, Feb 18, 2026</a></li><li><strong>Rubin</strong> — First cohort: AWS, Google Cloud, Azure, OCI, CoreWeave, Lambda, Nebius, Nscale; 2026 output estimated at 200,000–300,000, 60–70% to hyperscalers; Fairwater to scale to hundreds of thousands of Vera Rubin superchips; Google network up to 960,000 Rubin GPUs · <a href="https://www.techtimes.com/articles/319203/20260627/nvidia-vera-rubin-ships-this-fall-8-cloud-partners-10x-lower-token-cost-hbm4-triples-bandwidth.htm">TechTimes, Jun 27, 2026</a></li><li><strong>Microsoft</strong> — Rents 30,000 Vera Rubin from Nscale in Norway · <a href="https://letsdatascience.com/news/microsoft-rents-30000-nvidia-vera-rubin-chips-7119dfbe">Let’s Data Science, Apr 14, 2026</a></li><li><strong>SpaceX</strong> — Listed June 12, 2026; owns xAI since February; Colossus across sites toward 1 million GPUs · <a href="https://en.wikipedia.org/wiki/Initial_public_offering_of_SpaceX">Wikipedia</a></li><li><strong>SpaceX</strong> — Colossus 2 reported at 555,000 GPUs · <a href="https://introl.com/blog/xai-colossus-2-gigawatt-expansion-555k-gpus-january-2026">Introl, Jan 2026</a></li><li><strong>Market caps</strong> — Closes on October 26: Microsoft, Alphabet, Amazon, Meta, Oracle, SpaceX · <a href="https://www.stocktitan.net/rankings/companies-market-cap">StockTitan</a></li><li><strong>CoreWeave market</strong> — $59.57 close Oct 27 (52-week range $51.24–$141.79); EV · <a href="https://www.fool.com/coverage/stock-market-today/2026/10/27/stock-market-today-oct-27-coreweave-surges-on-denmark-data-center-launch/">Motley Fool; StockAnalysis</a></li><li><strong>CoreWeave market</strong> — Market cap and enterprise value, Oct 23 (re-marked to Oct 27’s close) · <a href="https://stockanalysis.com/stocks/crwv/market-cap/">StockAnalysis</a></li><li><strong>Nebius market</strong> — $41.66B market cap at the Oct 26 close; debt $10.06B, cash $8.04B · <a href="https://companiesmarketcap.com/nebius-group/marketcap/">CompaniesMarketCap (Nasdaq); TipRanks</a></li><li><strong>IREN market</strong> — $27.44 close Oct 27, $11.10B market cap; debt $7.84B, cash $6.08B · <a href="https://companiesmarketcap.com/iris-energy/marketcap/">CompaniesMarketCap; TipRanks</a></li></ul>
<h2>Method</h2>
<p><strong>Blackwell GPUs</strong> are counted one per B200, B300, GB200 or GB300 GPU package, not per die. For Microsoft, Meta, Amazon and Alphabet they are Epoch AI’s estimates of B200 and B300 shipped through Q1 2026, which Epoch flags as incomplete. Site estimates in H100-equivalents are converted at Epoch’s own ratio, 2.527 per Blackwell GPU.</p>
<p><strong>Power to GPUs.</strong> CoreWeave and Nebius don’t publish GPU counts, so we convert active power at about 2.1 kW of IT load per Blackwell GPU (1.9–2.3): a GB200 or GB300 NVL72 rack draws about 120–140 kW for 72 GPUs, plus networking and storage. CoreWeave’s base removes the 360 MW its mostly-Hopper fleet used at its IPO filing. Nebius’s removes about 70 MW of Hopper from the roughly 408 MW we infer for June from its run-rate revenue. Forward power converts at 3 kW per Rubin-generation GPU (2.6–3.4).</p>
<p><strong>Market cap and enterprise value</strong> use closing prices on October 26 (October 27 for CoreWeave and IREN) and all share classes. Enterprise value adds net debt from the latest balance sheet: CoreWeave’s from StockAnalysis, re-marked to the October 27 price; Oracle’s $88.3B at August 31; Nebius’s and IREN’s from TipRanks. Capex is the latest full-year guide at its midpoint (fiscal 2027 for Oracle and IREN).</p>
<p><strong>Not counted:</strong> GPUs a company rents rather than owns (they count for the owner, so Microsoft’s capacity at Nebius, IREN and Nscale counts for them), Hopper and older GPUs, and other chips: TPUs, Trainium and AMD. Private owners, among them OpenAI, Anthropic, Lambda, Nscale and Crusoe, have no market cap to divide by.</p>
<hr>
<p>Every chart above links to its interactive version on 06.ai, with the data behind it. The daily trackers are at <a href="https://06.ai/trackers">06.ai/trackers</a>.</p>
<p><em>Not investment advice. Estimates are labeled as estimates.</em></p>]]></content:encoded>
</item>
<item>
  <title>Who makes money serving open models?</title>
  <link>https://inthesingularity.ai/packs/inference-margins</link>
  <guid isPermaLink="true">https://inthesingularity.ai/packs/inference-margins</guid>
  <pubDate>Tue, 27 Oct 2026 12:15:00 GMT</pubDate>
  <description>Open-weight models now carry most of OpenRouter’s US traffic, served by dozens of hosts competing on price. What a GPU earns serving them, chip by chip and model by model, who keeps the margin, and who wins if GPU rents invert.</description>
  <category>AI infrastructure</category>
  <enclosure url="https://inthesingularity.ai/newsletter/inference-margins/cover.png" length="60621" type="image/png" />
  <media:content url="https://inthesingularity.ai/newsletter/inference-margins/cover.png" medium="image" type="image/png" />
  <content:encoded><![CDATA[<figure><a href="https://06.ai/packs/inference-margins"><img src="https://06.ai/newsletter/inference-margins/thumb.png" alt="Who makes money serving open models?" width="421" height="391"></a><figcaption><a href="https://06.ai/packs/inference-margins">Interactive version and data on 06.ai →</a></figcaption></figure>
<p>Serving DeepSeek’s V4.1-Flash on a Blackwell rack earns about <strong>$39</strong> an hour per GPU at DeepSeek’s off-peak prices, so a host renting the rack covers its rent at <strong>37%</strong> utilization. On an 8-GPU B200 box running NVIDIA’s standard recipe it takes 60%, and an H200 can’t cover its rent at those prices at all. The margin sits with rack-scale serving, frontier-size models and owned GPUs; on the most popular model, 33 hosts compete on price.</p>
<ul><li><strong>$39</strong> — Revenue per GPU-hour, Blackwell rack. V4.1-Flash at DeepSeek’s off-peak prices on a GB300 NVL72, per vLLM. The rack rents for $14.59 a GPU-hour</li><li><strong>37%</strong> — Utilization a renter needs to break even. Same rack. 60% on an 8-GPU B200 box with NVIDIA’s recipe; an H200 would need 120%, so never</li><li><strong>33</strong> — Hosts selling one model for V4.1-Flash on OpenRouter. 15 list exactly DeepSeek’s peak rate; 8 discount by 30–84%</li><li><strong>24.6%</strong> — Zhipu’s gross margin on API sales. H1 2026, from −0.4% a year earlier, after raising API prices about 101% this year</li></ul>
<h2>What a GPU earns, and the utilization a renter needs</h2>
<figure><a href="https://06.ai/packs/inference-margins#earn"><img src="https://06.ai/newsletter/inference-margins/fig-01.png" alt="What a GPU earns, and the utilization a renter needs" width="694" height="504"></a><figcaption>Revenue per GPU-hour at full load. The rack is vLLM’s own figure on SemiAnalysis AgentX traffic. The rest are NVIDIA’s Dynamo recipes on its agent trace (64K-token prompts, 400-token answers, 90% of the prompt cached), priced at each model’s going rate: DeepSeek’s off-peak rate for V4.1-Flash, with the paler bar at the peak rate most OpenRouter hosts post; Z.ai’s list for GLM-5.3-Flash; Moonshot’s $3/$15 for Kimi K3. Ticks are typical on-demand rent on October 20. The right-hand column is the share of hours a renter must sell to cover it.</figcaption></figure>
<p>An owner needs far less. At SemiAnalysis’s owner costs ($1.95 an hour for a B200, $2.65 for a GPU in a GB300 rack) the B200 box covers its costs at 15% utilization and the GB300 rack at 7%. Kimi K3 on H200s doesn’t cover even the cost of owning them.</p>
<h2>How the money comes in</h2>
<p><strong>Routing is price-first.</strong> OpenRouter’s default spreads requests across hosts that have been stable, weighted by the inverse square of price, so a host at half the price draws four times the traffic. It adds no markup to host prices and earns a 5.5% fee when buyers load credits, so a host keeps the price it posts.</p>
<p><strong>The traffic is agents, and mostly input.</strong> In the week to October 14, V4.1-Flash processed 36.6 trillion prompt tokens and 807 billion completion tokens, about 45 to 1, and its biggest senders are coding and agent harnesses. In the week to October 26 it carried 52.8T tokens, 28% of everything in OpenRouter’s daily top 50. Chinese models carry 60% of OpenRouter’s tokens over the last four weeks but 26% of the dollars (<a href="https://06.ai/trackers/openrouter">tracker</a>).</p>
<p><strong>So it’s a cache business priced per input token.</strong> On NVIDIA’s trace, 70% of a V4.1-Flash request’s revenue comes from the 10% of prompt tokens that miss the cache; output is 17%. At OpenRouter’s 45-to-1 mix, uncached input is still about 48%. The cache-read price, which hosts set separately, decides what an agent pays. <a href="https://06.ai/packs/deepinfer">The </a><a href="https://06.ai/packs/deepinfer">DeepInfer</a><a href="https://06.ai/packs/deepinfer"> pack</a> shows the same split from DeepSeek’s side.</p>
<p><strong>Posted isn’t paid.</strong> 15 of the 33 hosts list exactly DeepSeek’s peak rate, $0.30 per million input tokens and $1.20 per million output; 8 run discounts of 30% to 84%. OpenRouter now shows next to the posted prices what buyers actually pay on average, which caching and discounts pull well below the list.</p>
<h2>What one agent request costs at each host</h2>
<figure><a href="https://06.ai/packs/inference-margins#hosts"><img src="https://06.ai/newsletter/inference-margins/fig-02.png" alt="What one agent request costs at each host" width="694" height="650"></a><figcaption>Dollars per 1,000 requests on NVIDIA’s agent trace (64K-token prompt, 90% of it cached, 400-token answer) at each host’s cheapest standard endpoint for V4.1-Flash on October 27, discounts included. Listed companies in pink. Dashed lines: DeepSeek’s own off-peak and peak rates. Two hosts are left out: one lists no cache price and one listing couldn’t be read.</figcaption></figure>
<p>Headline prices mislead. Decart has the 2nd-lowest input-plus-output price but ranks 11th on an agent’s bill, because its cache reads cost $0.018 per million, three times DeepSeek’s peak rate. Alibaba Cloud drops from 10th to 19th the same way. Modal and Google Vertex charge $0.03 for cache reads and cost 3.0× DeepSeek’s off-peak rate per request.</p>
<h2>Which models pay best</h2>
<p><strong>Frontier-size open models, on the newest racks.</strong> Kimi K3 has 2.8 trillion parameters, 104 billion of them active, and nine of the ten platforms that served it in its first weeks charged Moonshot’s $3 input, $15 output. That price holds because the economics only work at rack scale: on NVIDIA’s recipes a GB300 GPU earns $34.91 an hour on K3, a GB200 $12.58 and an H200 $1.35.</p>
<p><strong>New models, for a few days.</strong> StepFun’s Step 5 Preview reached second place on OpenRouter in its first week, and K3 was live on nearly every major platform within 48 hours of its weights. The window before thirty hosts pile in is short.</p>
<p><strong>The Flash tier is where the volume is.</strong> V4.1-Flash, GLM-5.3-Flash and MiniMax’s M3.1-Flash carry the most tokens, and their prices follow cost down: over about 90 days the cheapest input price fell 77% for V4.1-Flash and 90% for GLM-5.3-Flash. GLM-5.3-Flash at Z.ai’s list earns $15.05 an hour on a B200, slightly more than V4.1-Flash at DeepSeek’s off-peak rate. Margins in this tier go to whoever runs the rack-scale stack.</p>
<p><strong>Hopper loses on every model here.</strong> An H200 can’t cover its rent on V4.1-Flash at DeepSeek’s off-peak prices, and NVIDIA doesn’t publish an H100 recipe for the model at all. Free endpoints, MiniMax’s M3.1-Flash and NVIDIA’s Nemotron 3 Ultra among them, carried 12.6T and 5.5T tokens in the week to October 26, free to users, so someone else pays.</p>
<h2>Who keeps the margin</h2>
<p>Reported figures. The measures differ, so compare rows with care.</p>
<ul><li><strong>DeepSeek</strong> — Layer: Lab serving its own open weights · Measure: Theoretical cost-profit ratio, V3 and R1 on H800s priced at $2 an hour · Feb 2025 · Value: 545% · At R1 list prices; DeepSeek said actual revenue was far lower</li><li><strong>Nebius NBIS</strong> — Layer: GPU cloud with a token API · Measure: Adjusted EBITDA margin, AI business · Q2 2026 · Value: 50% · Mostly GPU rental; Token Factory is its fastest-growing line</li><li><strong>Fireworks</strong> — Layer: Inference platform, private · Measure: Gross margin · 2026 · Value: ~50% · Targeting 60%</li><li><strong>Cerebras CBRS</strong> — Layer: Inference cloud on its own chips · Measure: Core gross margin · Q2 2026 · Value: 41% · Cloud services 53% (core) in Q1</li><li><strong>Zhipu 2513.HK</strong> — Layer: Lab selling its open weights by API · Measure: Gross margin on cloud API · H1 2026 · Value: 24.6% · −0.4% a year earlier, after raising API prices about 101% this year</li><li><strong>MiniMax 0100.HK</strong> — Layer: Lab selling its open weights by API · Measure: Gross margin, whole company · H1 2026 · Value: 17.9% · Open-platform revenue up 703%</li></ul>
<h2>How the margin moved</h2>
<p><strong>2023 to 2025: compute got cheap.</strong> H100 rents fell from peaks near $8 an hour to $2.00 on Silicon Data’s index by December 2025. In February 2025 DeepSeek said its V3 and R1 fleet cost $87,072 a day at $2 an H800-hour against $562,027 of theoretical revenue, about $12.90 per GPU-hour, while warning that actual revenue was far lower.</p>
<p><strong>2026: compute got scarce again.</strong> H100 and H200 contract prices rose about 40% between October 2025 and March 2026 as memory costs passed through. Silicon Data’s H100 index reached $2.88 on October 24, and Ornn’s settled at $3.17 on October 26, up 24.8% in a month.</p>
<p><strong>Token prices split.</strong> The Flash tier fell on architecture: V4.1-Flash activates 8B parameters to read and 16B to write. The frontier tier went up: K3 launched at $3/$15, Zhipu raised API prices about 101% this year, and DeepSeek warned of a significant price increase. Revenue per GPU-hour on Blackwell racks is about three times DeepSeek’s 2025 figure even after its price cuts (<a href="https://06.ai/packs/deepinfer">the </a><a href="https://06.ai/packs/deepinfer">DeepInfer</a><a href="https://06.ai/packs/deepinfer"> pack</a>), so hosts running the rack-scale stack kept ahead of rising rents and hosts on smaller setups fell behind.</p>
<p>The clearest public read is Zhipu: its API sales went from a −0.4% gross margin to 24.6% in a year, with prices about doubled. Selling tokens is still a thin business for anyone without the cheapest capacity. Owning the rack, or the model, is where the margin is.</p>
<h2>If the GPU curve inverts</h2>
<p>GPU rents aren’t inverted now. Silicon Data says its modeled forward curve sloped down in 2025 and moved to contango this year. In mid-2026 its B200 curve was only about 8% lower at 36 months ($5.62 to $5.17), and H100 was flat past 15 months. CME has distributed that curve to subscribers since October 15, and the H100 and B200 futures it planned to list on October 5 were reported as still pending regulatory review. Our own index (<a href="https://06.ai/trackers/gpu-compute-price-index">tracker</a>) started this week and doesn’t yet have enough term quotes to read a curve.</p>
<p>A steep inversion, a high near month and a low back end, would mean scarce capacity now and a supply wave or fast obsolescence later. A GPU-hour can’t be stored, so, as with power or freight, nothing ties the two ends together. Who gains, and who loses:</p>
<h2>Winners and losers in a steeply inverted curve</h2>
<p>Our assessment, roughly in order of how much each gains.</p>
<ul><li>Wins: <strong>Owners of installed GPUs that aren’t under contract</strong> — They sell at the expensive near end and stay off the cheap back end. Their cash margin is the spot rent less power.</li><li>Wins: <strong>Hosts that own their GPUs</strong> — Token prices settle where hosts renting at prompt rates break even. Owners keep the difference, and the inversion widens it.</li><li>Wins: <strong>Buyers who can wait</strong> — Training runs, RL and batch jobs buy the cheap back end. A long position in a deferred future gains as it rolls up toward spot, if spot holds.</li><li>Wins: <strong>Holders of leases fixed near the 2025 low</strong> — They are long compute below market and can undercut renters or sublet where contracts allow.</li><li>Wins: <strong>Capacity that can switch on now</strong> — Energized sites and the bottleneck parts, such as HBM. An inverted curve pays for speed.</li><li>Wins: <strong>Brokers and market makers</strong> — Steep curves and wide gaps between chips, regions and indices bring hedging demand and spreads to capture.</li><li>Loses: <strong>New builds that need contracts to borrow</strong> — Lower contracted rates mean less debt per GPU, which slows new supply and pulls the curve back.</li><li>Loses: <strong>Hosts renting at the prompt price</strong> — They pay the scarce price for GPUs and sell tokens at prices set by owners.</li><li>Loses: <strong>Long-dated bets on resale value</strong> — An inverted curve is the market pricing faster depreciation than six-year accounting lives.</li></ul>
<h2>How new hosts get in</h2>
<p>Most of the 33 hosts on V4.1-Flash fit four groups: inference-engine and model companies (Modular, Relace, Morph, Decart, Reka, Cohere), GPU clouds selling tokens as well as GPU-hours (CoreWeave, Nebius, DigitalOcean, Crusoe, GMI, Prime Intellect), Chinese clouds (Baidu, Alibaba, SiliconFlow), and privacy-focused hosts (Phala, Venice).</p>
<p><strong>Uptime first.</strong> Over the day to October 27, OpenRouter served V4.1-Flash 99.91% of the time; without its failover between hosts it would have been 87.81%. Hosts that stay up get the traffic. <strong>Then price:</strong> with inverse-square routing, a launch discount buys share fast. <strong>Pass the quality checks:</strong> OpenRouter publishes each host’s benchmark scores and has a routing mode for tool-calling accuracy, so heavy quantization shows. <strong>Be early, or fast:</strong> day-zero support catches a model before prices converge, and a speed tier sells at a premium (Baseten Fast posts $0.60/$2.40).</p>
<p>The arithmetic sets the floor. A host renting 8-GPU B200 boxes and pricing at DeepSeek’s off-peak rate needs 60% utilization to cover its rent. The hosts that last own their GPUs, fill idle ones, run the rack-scale stack, or have someone else paying.</p>
<h2>Listed ways in</h2>
<p>No listed company only rents GPUs and resells open-model tokens. The listed exposure is GPU clouds with token APIs on their own GPUs, Cerebras on its own chips, and the Hong Kong-listed labs selling their own open weights.</p>
<ul><li><strong>Nebius NBIS</strong> — What it sells: Token Factory: per-token API on 60+ open models, on its own GPUs. Hosts V4.1-Flash on OpenRouter at $0.30/$1.20. · Latest figures: Q2 revenue $582M; AI business 50% adjusted EBITDA margin; bought Eigen AI for $643M</li><li><strong>CoreWeave CRWV</strong> — What it sells: Serverless Inference: per-token API on an open-model catalog, on its own GPUs. Hosts V4.1-Flash on OpenRouter at $0.15/$0.65. · Latest figures: Managed-inference booked ARR went from $1M to over $100M within months of launch</li><li><strong>Cerebras CBRS</strong> — What it sells: Fast inference cloud on its own wafer-scale chips. · Latest figures: IPO in May raised $6.4B; Q2 core cloud revenue $127.7M, up 287%; core gross margin 41%</li><li><strong>DigitalOcean DOCN</strong> — What it sells: Hosts V4.1-Flash ($0.225/$0.90) and Kimi K3 on OpenRouter.</li><li><strong>Zhipu 2513.HK</strong> — What it sells: Sells its own open-weight GLM models by API, including through OpenRouter. · Latest figures: API revenue RMB 825M in H1, up 2,736%, 86.5% of revenue; API gross margin 24.6%</li><li><strong>MiniMax 0100.HK</strong> — What it sells: Sells its own M-series models by API, including through OpenRouter. · Latest figures: Open-platform revenue up 703% in H1; gross margin 17.9%</li><li><strong>Alphabet, Alibaba, Baidu GOOGL · BABA · BIDU</strong> — What it sells: Host V4.1-Flash on OpenRouter: Vertex at the posted rate, Alibaba Cloud at 50% off, Baidu Qianfan at 84% off.</li></ul>
<p>The pure-play hosts are private: Fireworks (over $1B annualized revenue, valued at $17.5B in July), Together ($8.3B), Baseten ($13B) and DeepInfra. Lambda, a GPU cloud, is raising up to $4B before a possible IPO at a $14.5B pre-money valuation.</p>
<h2>What this is anchored on</h2>
<p>Host prices as posted on October 27; rents as of October 20 (the <a href="https://06.ai/packs/deepinfer">DeepInfer</a> pack) and the dates shown.</p>
<ul><li><strong>Rack-scale yield</strong> — $38.96 per GPU-hour at DeepSeek's off-peak prices on one GB300 NVL72 rack, SemiAnalysis AgentX traffic; 954K tokens/s per GPU counting cached input · <a href="https://vllm.ai/blog/2026-10-23-deepseek-v41-flash-pd-wide-ep">vLLM blog, Oct 23</a></li><li><strong>Recipe yields</strong> — Output tokens/s per GPU on the agentic trace: V4.1-Flash 1,155 (GB200), 1,088 (B200), 359 (H200), disaggregated · <a href="https://docs.nvidia.com/dynamo/recipes/deepseek-v4-1-flash">NVIDIA Dynamo, DeepSeek-V4.1-Flash recipe</a></li><li><strong>Recipe yields</strong> — Kimi K3: 91.3 (GB300), 32.9 (GB200), 3.54 (H200) output tokens/s per GPU; trace: 64K input, 400 output, 90% cache hits · <a href="https://docs.nvidia.com/dynamo/recipes/kimi-k3">NVIDIA Dynamo, Kimi-K3 recipe</a></li><li><strong>Recipe yields</strong> — GLM-5.3-Flash: 579 output tokens/s per GPU on four B200s · <a href="https://docs.nvidia.com/dynamo/recipes/glm-5-3-flash">NVIDIA Dynamo, GLM-5.3-Flash recipe</a></li><li><strong>Hosts</strong> — 33 hosts for V4.1-Flash with input, output and cache-read prices; pooled availability 99.91% against 87.81% without routing; top apps are agent harnesses · <a href="https://openrouter.ai/deepseek/deepseek-v4.1-flash-20260910">OpenRouter, DeepSeek V4.1-Flash</a></li><li><strong>Routing</strong> — Default routing weights stable hosts by the inverse square of price · <a href="https://openrouter.ai/docs/provider-routing">OpenRouter docs, provider routing</a></li><li><strong>Fees</strong> — No markup on host prices; 5.5% fee on credit purchases · <a href="https://openrouter.ai/docs/faq">OpenRouter FAQ</a></li><li><strong>Traffic mix</strong> — V4.1-Flash, week to Oct 14: 36.62T prompt and 807B completion tokens; Step 5 Preview reached #2 in its first week · <a href="https://metaprompt.ai/posts/most-used-models-openrouter">Metaprompt, from OpenRouter's weekly table</a></li><li><strong>Traffic mix</strong> — Open-weight models carry about 60% of tokens from US-originating requests on OpenRouter, mostly Chinese models · <a href="https://thenewstack.io/openrouter-us-region-routing/">The New Stack</a></li><li><strong>Prices</strong> — V4.1-Flash off-peak $0.15 input, $0.003 cached, $0.60 output; peak hours double · <a href="https://apidog.com/blog/deepseek-v4-1-flash-pricing/">Apidog</a></li><li><strong>Prices</strong> — Kimi K3: 2.8T parameters, 104B active; providers' prices · <a href="https://deepinfra.com/blog/kimi-k3-pricing-providers-cost">DeepInfra</a></li><li><strong>Prices</strong> — Nine of ten K3 platforms at $3/$15 with $0.30 cached; most live within 48 hours of the weights · <a href="https://dev.to/sophiaashi/every-kimi-k3-api-provider-compared-pricing-speed-and-the-one-50-discount-53ge">DEV Community</a></li><li><strong>Prices</strong> — GLM-5.3-Flash $0.15 input, $0.03 cached, $0.50 output · <a href="https://www.orcarouter.ai/blog/minimax-m3-1-flash-vs-glm-5-3-flash">OrcaRouter</a></li><li><strong>Prices</strong> — DeepSeek warned of a significant price increase after launching V4 at very low rates · <a href="https://www.vktr.com/ai-platforms/deepseeks-v41-flash-beats-its-own-flagship-model/">VKTR</a></li><li><strong>Price falls</strong> — V4.1-Flash cheapest input price down 76.7% over about 90 days, $0.30 to $0.07 · <a href="https://pricepertoken.com/pricing-page/model/deepseek-deepseek-v4.1-flash">PricePerToken</a></li><li><strong>Price falls</strong> — GLM-5.3-Flash cheapest input price down 90% over about 90 days · <a href="https://pricepertoken.com/pricing-page/model/z-ai-glm-5.3-flash">PricePerToken</a></li><li><strong>Rents</strong> — H100 rents have come down from historical peaks near $8 an hour · <a href="https://www.thundercompute.com/blog/ai-gpu-rental-market-trends">Thunder Compute</a></li><li><strong>Rents</strong> — H100 $3.17 an hour on Oct 26, up 24.8% in a month · <a href="https://data.ornn.com/">Ornn Compute Price Index</a></li><li><strong>Rents</strong> — Silicon Data H100 index $2.00 (Dec 9, 2025) to $2.88 (Oct 24, 2026); its forward curve moved from sloping down to contango · <a href="https://fourweekmba.com/ai-nvidia-gpu-rents-are-now-priced-to-rise-for-36-months/">FourWeekMBA</a></li><li><strong>Rents</strong> — Mid-2026 forward curve: B200 about 8% lower at 36 months; H100 flat past 15 months · <a href="https://www.silicondata.com/blog/gpu-futures">Silicon Data</a></li><li><strong>Rents</strong> — H100 and H200 contract prices up about 40% from Oct 2025 to Mar 2026 as HBM costs passed through · <a href="https://intuitionlabs.ai/articles/data-center-gpu-pricing-2026">IntuitionLabs</a></li><li><strong>Futures</strong> — CME distributes Silicon Data's forward curves from Oct 15; H100 and B200 futures planned for Oct 5, pending regulatory review · <a href="https://cryptobriefing.com/cme-group-gpu-rental-forward-curve-pricing/">Crypto Briefing; Spheron</a></li><li><strong>Owner cost</strong> — Owner cost per GPU-hour: B200 $1.95, GB200 $2.21, GB300 $2.65 · <a href="https://inferencex.semianalysis.com/compare-per-dollar/deepseek-r1-b200-vs-gb300">SemiAnalysis via InferenceX</a></li><li><strong>Margins</strong> — DeepSeek's Feb 2025 disclosure: $87,072 a day of GPU cost against $562,027 of theoretical revenue · <a href="https://www.scmp.com/tech/big-tech/article/3300734/deepseek-shows-power-v3-r1-models-theoretical-545-profit-margin">South China Morning Post</a></li><li><strong>Margins</strong> — Zhipu cloud API gross margin −0.4% to 24.6%; API prices up about 101% this year · <a href="https://eu.36kr.com/en/p/3964873497320966">36Kr</a></li><li><strong>Margins</strong> — Zhipu API revenue and share; Zhipu and MiniMax gross margins, H1 2026 · <a href="https://eu.36kr.com/en/p/3964573800586498">36Kr</a></li><li><strong>Margins</strong> — MiniMax open-platform revenue up 703.1% · <a href="https://www1.hkexnews.hk/listedco/listconews/sehk/2026/0826/2026082600680.pdf">MiniMax interim results, HKEX</a></li><li><strong>Margins</strong> — Fireworks gross margin about 50%, targeting 60% · <a href="https://sacra.com/c/fireworks-ai/">Sacra</a></li><li><strong>Companies</strong> — Cerebras Q2: core cloud revenue $127.7M, core gross margin 41%, $6.4B IPO · <a href="https://www.sec.gov/Archives/edgar/data/0002021728/000162828026056186/cbrsannouncesfinancialresu.htm">Cerebras 8-K</a></li><li><strong>Companies</strong> — Nebius Q2 revenue $582M; AI business 50% adjusted EBITDA margin · <a href="https://www.cnhinews.com/news/article_1f4b4f84-df6d-59ed-9487-e8775c51781b.html">Zacks via CNHI</a></li><li><strong>Companies</strong> — CoreWeave managed-inference booked ARR from $1M to over $100M · <a href="https://futurumgroup.com/?p=94191">Futurum</a></li><li><strong>Companies</strong> — Fireworks, Together and Baseten valuations and revenue · <a href="https://www.vccafe.com/the-rise-of-the-inference-economy/">VC Cafe</a></li><li><strong>Companies</strong> — Lambda raising up to $4B at $14.5B pre-money · <a href="https://letsdatascience.com/news/lambda-pursues-4-billion-pre-ipo-funding-round-909f4022">Let's Data Science</a></li></ul>
<h2>Method</h2>
<p><strong>Revenue per GPU-hour.</strong> For NVIDIA’s recipes: output tokens per second per GPU ÷ 400 output tokens per request × dollars per request × 3,600. Dollars per request: 64,000 prompt tokens, 10% at the uncached input price and 90% at the cache-read price, plus 400 output tokens. Full load means the recipe’s measured concurrency around the clock. NVIDIA’s output figures include reasoning tokens; its GLM and Kimi runs use a 15% subset of the trace and a simulated speculative-decoding acceptance rate.</p>
<p><strong>The rack</strong> is vLLM’s figure on one GB300 NVL72 at DeepSeek’s off-peak prices, on SemiAnalysis’s AgentX traffic, a different trace from NVIDIA’s. The paler bars double the V4.1-Flash figures, since DeepSeek’s peak rates double every price.</p>
<p><strong>Break-even utilization</strong> is rent ÷ revenue at full load. Rents are typical on-demand prices on October 20, as in the <a href="https://06.ai/packs/deepinfer">DeepInfer</a> pack; committed rents are lower, so these overstate what a long-term renter needs. Owner costs are SemiAnalysis’s total cost of ownership from its 2025 survey.</p>
<p><strong>Not counted:</strong> CPUs, networking, cache storage, staff, a GPU cloud’s own margin on top of the GPU, and capacity sitting idle between peaks. NVIDIA’s trace is more input-heavy than OpenRouter’s 45-to-1 mix.</p>
<p><strong>Hosts</strong> are each host’s cheapest standard-routing endpoint on OpenRouter’s V4.1-Flash page on October 27, at discounted prices where a discount runs; list prices are kept in the data. Token counts for the week to October 26 are from our <a href="https://06.ai/trackers/openrouter">OpenRouter tracker</a>’s daily top 50.</p>
<hr>
<p>Every chart above links to its interactive version on 06.ai, with the data behind it. The daily trackers are at <a href="https://06.ai/trackers">06.ai/trackers</a>.</p>
<p><em>Not investment advice. Estimates are labeled as estimates.</em></p>]]></content:encoded>
</item>
<item>
  <title>How many tokens can a GPU make?</title>
  <link>https://inthesingularity.ai/packs/deepinfer</link>
  <guid isPermaLink="true">https://inthesingularity.ai/packs/deepinfer</guid>
  <pubDate>Tue, 27 Oct 2026 12:14:00 GMT</pubDate>
  <description>DeepSeek open-sourced the engine behind its own API. Its benchmarks show what a Blackwell GPU now yields, what that is worth, and which chips get the gain. vLLM matched it three days later.</description>
  <category>AI infrastructure</category>
  <enclosure url="https://inthesingularity.ai/newsletter/deepinfer/cover.png" length="51751" type="image/png" />
  <media:content url="https://inthesingularity.ai/newsletter/deepinfer/cover.png" medium="image" type="image/png" />
  <content:encoded><![CDATA[<figure><a href="https://06.ai/packs/deepinfer"><img src="https://06.ai/newsletter/deepinfer/thumb.png" alt="How many tokens can a GPU make?" width="421" height="391"></a><figcaption><a href="https://06.ai/packs/deepinfer">Interactive version and data on 06.ai →</a></figcaption></figure>
<p>DeepSeek open-sourced DeepInfer, the engine its own API runs on. Its benchmarks show a Blackwell GPU producing <strong>26,300</strong> output tokens a second at about 103 tokens a second per user, and DeepSeek says it earns <strong>$38</strong> an hour per GPU at its off-peak prices, about three times its 2025 figure. Only Blackwell data-center GPUs and Huawei’s Ascend 950 can run it. Three days later vLLM published results within 6% of it, so the yield is now available to anyone with a Blackwell rack.</p>
<ul><li><strong>6.3×</strong> — Output per decode GPU vs 2025. 11,573 tokens/s at DeepSeek’s production point, against 1,850 for V3 on H800s, at 11.5× the speed per user</li><li><strong>$38</strong> — Revenue per GPU-hour, off-peak. DeepSeek’s figure; vLLM gets $38.96 on agent traffic. $12.91 for V3 in 2025; a GB200 rents for $10.58</li><li><strong>155T</strong> — Tokens a day through one cluster. 94% of them cached input, on our reconstruction; about 2,700 GPUs</li><li><strong>≤6%</strong> — vLLM’s gap to DeepInfer on decode. Oct 23, on GB200 at batch 8 to 256; 954K tokens/s per GPU on agent traffic on a GB300 NVL72</li></ul>
<h2>Output per GPU falls as each user gets faster</h2>
<figure><a href="https://06.ai/packs/deepinfer#decode"><img src="https://06.ai/newsletter/deepinfer/fig-01.png" alt="Output per GPU falls as each user gets faster" width="694" height="442"></a><figcaption>DeepInfer decode on Blackwell at a 131,072-token context, as DeepSeek benchmarked it: output tokens per second per GPU against tokens per second per user, one dot per batch size. EP64 spreads the model’s experts over 64 GPUs in one NVLink domain, which takes a 72-GPU rack; EP8 over one 8-GPU box. Benchmark figures, before DeepSeek’s 80% haircut for live traffic.</figcaption></figure>
<h2>What DeepSeek released</h2>
<p>On October 20 DeepSeek put DeepInfer, the engine behind its own API, on GitHub under the Apache 2.0 license. It serves one model, V4.1-Flash, and runs only on NVIDIA’s Blackwell data-center GPUs (B200, B300 and their 72-GPU racks) and Huawei’s Ascend 950. Hopper GPUs such as the H100 and H200 fail its startup check, and AMD isn’t a build target. It is the offline core: the cache store, request routing and prefill/decode split a live service needs are left out, and the smallest efficient deployment is 64 prefill plus 64 decode GPUs. DeepSeek says the open build is within 5% of its internal one and that live traffic reaches about 80% of these benchmarks.</p>
<h2>Does it change GPU usage?</h2>
<p>Not for DeepSeek, which has run this stack since V4.1-Flash launched in September. For everyone else it set the bar, and the bar was cleared fast. On the dashboard we built on release day, DeepInfer’s curves implied about 2.0× the tokens per GPU vLLM had published on agent traffic a few days earlier (830K a second). On October 23 vLLM, the most widely used open-source engine, published decode results within 6% of DeepInfer’s across its batch sizes, using DeepSeek’s kernels, and 954K tokens a second per GPU on the same agent workload on a GB300 NVL72 rack, with the online serving DeepInfer lacks. That is 15% more than its October 17 figure, with a P90 speed of 196 tokens a second per user against a 100 target then. Anyone hosting V4.1-Flash on a Blackwell rack can now get close to DeepSeek’s own yield with an open engine, which cuts the GPUs a host needs for the same traffic.</p>
<h2>Does it change token yields?</h2>
<p>Against 2025, by a lot. At its production point DeepSeek gets 6.3× the output per decode GPU of its V3 fleet on H800s, at 11.5× the speed per user, and $38 of revenue per GPU-hour against $12.91, even after cutting prices by about three-quarters. Most of that arrived with the model: V4.1-Flash stores 890 bytes of cache per token, against 48,068 for V3.2. The engine is how DeepSeek cashes it in. But the yield depends on the speed sold: 26,300 tokens a second per GPU at 103 per user, falling to 26,300 at 103.</p>
<h2>Do GPU types change differently?</h2>
<p>Yes, and topology matters more than the chip. Decode wants 64 to 128 GPUs in one NVLink domain, which only GB200 and GB300 NVL72 racks offer. An 8-GPU B200 or B300 box keeps up at relaxed speeds and tops out at 239 tokens a second per user. Hopper and AMD gain nothing directly. Ascend 950 is a first-class target, which matters more for China’s chip mix than for anyone’s throughput.</p>
<h2>Cost per million output tokens at market rental</h2>
<figure><a href="https://06.ai/packs/deepinfer#cost"><img src="https://06.ai/newsletter/deepinfer/fig-02.png" alt="Cost per million output tokens at market rental" width="694" height="392"></a><figcaption>Typical rental per GPU-hour on October 20 ÷ output per GPU-hour, after the 80% live-traffic factor. GB200 NVL72 on the EP64 curve at $10.58 an hour; HGX B200 on the EP8 curve at $8.02, scaled to its 90% peak rate. The dashed line is DeepSeek’s off-peak output price. GPU time only: prefill, cache storage and the rest of the stack are extra.</figcaption></figure>
<p>The 8-GPU box is the cheaper way to buy output up to about 130 tokens a second per user; past that the rack wins, by a widening margin. The box costs more per output token than DeepSeek charges off-peak beyond about 225; the rack holds out to about 310. At equal speed the rack makes 1.1× the box’s output per GPU at 105 tokens a second per user and 6.4× at 235.</p>
<h2>What DeepInfer changes, chip by chip</h2>
<p>Per-GPU specs from vendor sheets and coverage, HBM as vendors list it. Scale-up domain is the most GPUs one interconnect domain holds. Rental is the typical on-demand price per GPU-hour on October 20. “What changes” is our assessment.</p>
<ul><li><strong>GB200 NVL72 NVIDIA</strong> — DeepInfer: Runs · FP4 PFLOPS: 10 · HBM GB: 186 · Scale-up: 72 · Rent / hr: $10.58 · Full effect. The setup DeepSeek benchmarks: EP64 decode inside one 72-GPU NVLink domain, Mega MoE kernels, native FP4 weights.</li><li><strong>GB300 NVL72 NVIDIA</strong> — DeepInfer: Runs · FP4 PFLOPS: 15 · HBM GB: 279 · Scale-up: 72 · Rent / hr: $14.59 · Full effect expected; FlashMLA and DeepGEMM build for SM 10.3. Extra HBM matters less now that KV no longer binds; extra FP4 compute helps prefill. Rents for about 38% more than GB200 per GPU-hour.</li><li><strong>HGX B200 NVIDIA</strong> — DeepInfer: Runs · FP4 PFLOPS: 9 · HBM GB: 180 · Scale-up: 8 · Rent / hr: $8.02 · Partial. Runs DeepInfer, but the NVLink island caps decode at EP8: close to parity at relaxed speeds, falls behind fast. DeepSeek uses 8-GPU boxes for prefill.</li><li><strong>HGX B300 NVIDIA</strong> — DeepInfer: Runs · FP4 PFLOPS: 13.5 · HBM GB: 270 · Scale-up: 8 · Rent / hr: $9.99 · Partial, as B200. More HBM and FP4 compute don't lift the EP8 ceiling on fast decode; well suited to prefill.</li><li><strong>H200 NVIDIA</strong> — DeepInfer: No · HBM GB: 141 · Scale-up: 8 · Rent / hr: $5.33 · No direct effect. The engine's startup check rejects anything but SM 10.x and Hopper has no FP4 tensor cores. Gains arrive only as vLLM and SGLang port the ideas.</li><li><strong>H100 NVIDIA</strong> — DeepInfer: No · HBM GB: 80 · Scale-up: 8 · Rent / hr: $4.17 · No direct effect, as H200, with less HBM.</li><li><strong>H800 NVIDIA</strong> — DeepInfer: No · HBM GB: 80 · Scale-up: 8 · No direct effect. The chip DeepSeek served V3 and R1 on in 2025.</li><li><strong>H20 NVIDIA</strong> — DeepInfer: No · HBM GB: 96 · Scale-up: 8 · No direct effect; also the weakest compute in this table.</li><li><strong>MI355X AMD</strong> — DeepInfer: No · FP4 PFLOPS: 10.07 · HBM GB: 288 · Scale-up: 8 · Rent / hr: $5.60 · No direct effect. FP4-capable, but DeepInfer's kernels are CUDA or Ascend only; runs V4.1-Flash through vLLM, SGLang or ATOM.</li><li><strong>MI300X AMD</strong> — DeepInfer: No · HBM GB: 192 · Scale-up: 8 · Rent / hr: $3.65 · No direct effect; no FP4 matrix units, so FP4 experts are dequantized in software.</li><li><strong>Ascend 950 Huawei</strong> — DeepInfer: Runs · FP4 PFLOPS: 2 · HBM GB: 144 · Scale-up: 8,192 · Strategic. First-class target (DeepGEMM-Ascend, DeepEP-Ascend, TileLang). DeepSeek and Huawei built a 128-chip Ascend 950 supernode that fits the decode design; Huawei's SuperPoD scales to 8,192 chips. Peak FP4 per chip is a fifth of GB200's.</li></ul>
<h2>Most tokens are cache reads; most revenue isn’t</h2>
<figure><a href="https://06.ai/packs/deepinfer#recon"><img src="https://06.ai/newsletter/deepinfer/fig-03.png" alt="Most tokens are cache reads; most revenue isn’t" width="694" height="178"></a><figcaption>How $38 per GPU-hour adds up at V4.1-Flash off-peak prices: one prefill GPU per decode GPU, as DeepSeek runs them, both at the production point and the 80% factor. We solve for the share of input that hits the cache so the pair averages DeepSeek’s figure: 95%, close to the 96% that agent benchmarks measure.</figcaption></figure>
<p>At those rates about 2,700 GPUs would carry the 155T tokens a day DeepSeek reports and bill about $0.9B a year off-peak, or $1.1B counting peak hours at double. Reported annualized revenue for all of DeepSeek was about $1.0B at the start of October, so the figures hang together. vLLM’s independent $38.96 on a GB300 NVL72 points the same way.</p>
<h2>Memory no longer limits the batch</h2>
<figure><a href="https://06.ai/packs/deepinfer#kv"><img src="https://06.ai/newsletter/deepinfer/fig-04.png" alt="Memory no longer limits the batch" width="694" height="202"></a><figcaption>How many 131,072-token sessions fit in 100 GB of HBM, by generation, log scale. At DeepInfer’s largest decode batch, 256 such sessions take about 30 GB per GPU, so bandwidth and communication now set decode speed, not memory.</figcaption></figure>
<h2>Then and now</h2>
<p>DeepSeek’s February 2025 fleet averages for V3 and R1 on H800s against V4.1-Flash on DeepInfer at its production point, after the 80% factor. Different chips and price decks; V3 revenue is DeepSeek’s theoretical figure at R1 list prices.</p>
<figure><a href="https://06.ai/packs/deepinfer#gens"><img src="https://06.ai/newsletter/deepinfer/fig-05.png" alt="Output tokens/s per decode GPU — V3 / R1 on H800 · Feb 2025: 1,850 · V4.1-Flash on DeepInfer · Oct 2026: 11,573 · Change: 6.3×; Per-user output speed, tokens/s — V3 / R1 on H800 · Feb 2025: 21 · V4.1-Flash on DeepInfer · Oct 2026: 241 · Change: 11.5×; GPU-hours per billion output tokens — V3 / R1 on H800 · Feb 2025: 150 · V4.1-Flash on DeepInfer · Oct 2026: 24 · Change: −84%; Revenue per GPU-hour — V3 / R1 on H800 · Feb 2025: $12.91 · V4.1-Flash on DeepInfer · Oct 2026: $38.00 · Change: 2.9×; Output, per 1M tokens — V3 / R1 on H800 · Feb 2025: $2.19 · V4.1-Flash on DeepInfer · Oct 2026: $0.60 · Change: −73%; Uncached input, per 1M tokens — V3 / R1 on H800 · Feb 2025: $0.55 · V4.1-Flash on DeepInfer · Oct 2026: $0.15 · Change: −73%; Cached input, per 1M tokens — V3 / R1 on H800 · Feb 2025: $0.140 · V4.1-Flash on DeepInfer · Oct 2026: $0.003 · Change: −98%" width="694" height="457"></a></figure>
<h2>Where the gain comes from</h2>
<p><strong>The model.</strong> V4.1-Flash keeps 890 bytes of cache per token, 3.9× less than V4-Flash, and activates 8B parameters when reading and 16B when writing. Its launch prices show where costs fell: cached input 57% cheaper than V4-Flash, uncached input 32%, output only 9%.</p>
<p><strong>The engine.</strong> Wide expert parallelism is most of it. Bounded replay rebuilds only the last 140 tokens of sliding-window state, which DeepSeek says nearly halves prefill work. Confidence-scheduled drafting verifies fewer guessed tokens per step and adds 5% throughput at the production batch, 11% at the largest.</p>
<p><strong>The hardware.</strong> The model’s expert weights are in FP4, which Blackwell, AMD’s MI355X and Ascend 950 multiply natively and Hopper does not. Decode wants a big NVLink domain, which is why the rack beats the box.</p>
<h2>What it means</h2>
<p><strong>For GPU economics.</strong> A GB200 rents for about $10.58 an hour and, served this way, earns $38 at DeepSeek’s off-peak prices. Serving open models on Blackwell racks is profitable at market GPU prices, so long as the traffic is cache-heavy agent work, and vLLM now makes that available to every host. Third-party hosts charge about DeepSeek’s peak rate, so there is room for their prices to fall toward its off-peak ones (<a href="https://06.ai/packs/token-prices">the </a><a href="https://06.ai/packs/token-prices">token-price</a><a href="https://06.ai/packs/token-prices"> pack</a>).</p>
<p><strong>For older chips.</strong> Hopper rents for less but gets none of this directly, and it can’t do the FP4 math. The yield gap between a Blackwell rack and an H100 widens with each release like this one, which should weigh on Hopper rental prices over time.</p>
<p><strong>For China.</strong> DeepSeek built Ascend 950 support in from the start and is reported to be planning a 160,000-chip Huawei data center. DeepSeek hasn’t published any Ascend throughput, which is the number to watch.</p>
<h2>What this is anchored on</h2>
<p>Decode and prefill points are read off DeepSeek’s benchmark charts; rental prices and DeepInfer figures as of October 20, vLLM’s as of October 23.</p>
<ul><li><strong>DeepInfer</strong> — Engine, supported hardware, decode and prefill benchmark charts, $38 per GPU-hour, 155T tokens a day · <a href="https://github.com/deepseek-ai/DeepInfer">DeepInfer repository</a></li><li><strong>DeepInfer</strong> — Production concurrency, draft verification table, balance factor, KV storage, minimum deployment · <a href="https://github.com/deepseek-ai/DeepInfer/blob/main/doc/inference_deployment_qa.md">DeepInfer deployment Q&amp;A</a></li><li><strong>V4.1-Flash</strong> — Parameters, 890 bytes of KV cache per token, KV series by generation · <a href="https://arxiv.org/abs/2609.19969">DeepSeek-V4.1-Flash technical report</a></li><li><strong>V4.1-Flash</strong> — KV-cache series by generation · <a href="https://www.contextstudios.ai/blog/deepseek-v4-1-flash-890-bytes-kv-cache-per-token">Context Studios</a></li><li><strong>Prices</strong> — V4-Flash and V4.1-Flash off-peak and peak rates; peak hours a week · <a href="https://apidog.com/blog/deepseek-v4-1-flash-pricing/">Apidog</a></li><li><strong>vLLM</strong> — Within 6% of DeepInfer on decode; 954K tokens/s per GPU on AgentX, GB300 NVL72; $38.96 per GPU-hour (Oct 23) · <a href="https://vllm.ai/blog/2026-10-23-deepseek-v41-flash-pd-wide-ep">vLLM blog</a></li><li><strong>vLLM</strong> — 830K tokens/s per GPU on AgentX at 100 tok/s per user (Oct 17) · <a href="https://aicoder.com/news/news-20261017-vllm-openrouter-deepseek-v41-flash-1017-day0-serving">AICoder</a></li><li><strong>vLLM</strong> — 5× high-throughput gain on V4.1-Flash since day 0 · <a href="https://vllm.ai/blog/2026-10-07-deepseek-v41-flash">vLLM blog</a></li><li><strong>AgentX</strong> — Median input and output tokens, cache-hit rate · <a href="https://www.matthewswong.com/en/blog/deepseek-v4-1-flash-serving-vllm-vs-sglang/">AgentX workload</a></li><li><strong>AgentPerf</strong> — Tokens/s per MW by system on V4.1-Flash · <a href="https://artificialanalysis.ai/articles/vera-rubin-nvl72-agentperf">Artificial Analysis</a></li><li><strong>AgentPerf</strong> — MI355X, B200 and B300 results · <a href="https://cryptobriefing.com/amd-mi355x-aa-agentperf-deepseek-results/">CryptoBriefing</a></li><li><strong>Rental</strong> — Typical on-demand price per GPU-hour, Oct 20 · <a href="https://savrn.com/ai-index/pricing/gpus">Savrn AI index</a></li><li><strong>Ascend 950</strong> — Specs · <a href="https://www.techradar.com/pro/huawei-ascend-950-vs-nvidia-h200-vs-amd-mi300-instinct-how-do-they-compare">TechRadar</a></li><li><strong>Ascend</strong> — DeepSeek’s Ascend tooling; planned Huawei data center; $1B run-rate · <a href="https://thenextweb.com/news/deepseek-huawei-ascend-tilelang-open-source-cuda">The Next Web</a></li><li><strong>DeepInfer</strong> — Release coverage · <a href="https://www.marktechpost.com/2026/10/20/deepseek-deepinfer-open-source-inference-framework/">MarkTechPost</a></li><li><strong>DeepInfer</strong> — Release coverage · <a href="https://thenextweb.com/news/deepseek-deepinfer-open-source-inference-engine">The Next Web</a></li><li><strong>V3 / R1, 2025</strong> — DeepSeek’s Feb 2025 inference-system figures on H800s · <a href="https://www.analyticsvidhya.com/blog/2025/03/deepseek-inference-system-overview/">Analytics Vidhya</a></li></ul>
<h2>Method</h2>
<p><strong>Reading the charts.</strong> DeepSeek’s decode chart defines output per GPU as per-user speed × requests per GPU, and our readings meet that within 0.5%. Decode output assumes 75% acceptance on 3 draft tokens, as DeepSeek’s chart does; only the step rate is measured. The EP64 curve ran inside one NVLink domain, so on a GB200 or GB300 NVL72 rack; DeepSeek doesn’t say which.</p>
<p><strong>The production point</strong> is DeepSeek’s draft-verification table at 60 sequences per GPU, which it says is where production runs. “After the live-traffic factor” multiplies by 80%, DeepSeek’s ratio of live traffic to its perfectly balanced benchmarks.</p>
<p><strong>Costs</strong> use typical on-demand rental per GPU-hour, not DeepSeek’s own costs, which it doesn’t disclose. HGX B200 output is the EP8 curve scaled by DeepInfer’s own peak-rate table (2,250 against 2,500 TFLOPS).</p>
<p><strong>The revenue check</strong> assumes every GPU runs at the production point around the clock, with prefill at 32,768-token steps. The comparison with vLLM’s earlier result priced one agent request (131,072-token context, 4,096 uncached, 444 output tokens) in DeepInfer GPU-seconds; it set an offline benchmark against a live replay, so it was indicative, and vLLM’s like-for-like decode test on October 23 supersedes it.</p>
<p><strong>What DeepSeek doesn’t disclose:</strong> which Blackwell part ran the benchmarks, whether the cluster behind the daily token figure is NVIDIA, Ascend or both, fleet size, the real cache-hit share, and any Ascend 950 throughput.</p>
<hr>
<p>Every chart above links to its interactive version on 06.ai, with the data behind it. The daily trackers are at <a href="https://06.ai/trackers">06.ai/trackers</a>.</p>
<p><em>Not investment advice. Estimates are labeled as estimates.</em></p>]]></content:encoded>
</item>
<item>
  <title>Can the cloud pay for the AI capex?</title>
  <link>https://inthesingularity.ai/packs/cloud-capex-payback</link>
  <guid isPermaLink="true">https://inthesingularity.ai/packs/cloud-capex-payback</guid>
  <pubDate>Mon, 26 Oct 2026 12:13:00 GMT</pubDate>
  <description>Azure is growing 43% and AWS 37%, the fastest in years. How much of that is AI, what Microsoft’s and Amazon’s AI capex earns from the cloud alone with their OpenAI and Anthropic stakes at zero, and what happens if the labs can’t pay.</description>
  <category>AI returns</category>
  <enclosure url="https://inthesingularity.ai/newsletter/cloud-capex-payback/cover.png" length="60407" type="image/png" />
  <media:content url="https://inthesingularity.ai/newsletter/cloud-capex-payback/cover.png" medium="image" type="image/png" />
  <content:encoded><![CDATA[<figure><a href="https://06.ai/packs/cloud-capex-payback"><img src="https://06.ai/newsletter/cloud-capex-payback/thumb.png" alt="Can the cloud pay for the AI capex?" width="421" height="409"></a><figcaption><a href="https://06.ai/packs/cloud-capex-payback">Interactive version and data on 06.ai →</a></figcaption></figure>
<p>Before AI, both clouds were slowing: Azure from 50% growth in 2021 to 26% in mid-2023, AWS from 40% to 12%. Now Azure grows <strong>43%</strong> and AWS <strong>37%</strong>. We estimate AI accounts for about <strong>26 of Azure’s 43 points</strong> of growth this year and <strong>17 of AWS’s 33</strong>. Renting out compute takes a lot of capital, and so far each dollar has earned more at Amazon than at Microsoft: in 2026, about 40 cents a year of AI revenue for each dollar of AI capex already in service at Amazon, and 26 cents at Microsoft. Count only cloud revenue, charge all the AI capex, value the stakes in OpenAI and Anthropic at zero, and the spending through 2027 earns about <strong>12.8% a year at Amazon</strong> and <strong>3.0% at Microsoft</strong>. Both numbers turn on how quickly capex starts earning, and on two labs that don’t yet make money paying their bills. These are the middle of our ranges.</p>
<ul><li><strong>26 of 43 pts</strong> — AI’s share of Azure’s 2026 growth about $39B of revenue, 32% of Azure, against a no-AI trend of 17% a year</li><li><strong>17 of 33 pts</strong> — AI’s share of AWS’s 2026 growth about $31B, 18% of AWS, against 14.5%; in line with Amazon’s own AI run-rates</li><li><strong>12.8%</strong> — Amazon’s return on AI capex through 2027 a year from AWS alone, stakes at zero. 8.6% charging the $68B it put into OpenAI and Anthropic; 7.2% if the labs pay half their bills from 2028</li><li><strong>3.0%</strong> — Microsoft’s return on AI capex through 2027 from Azure alone; 6.0% counting the Copilots, 2.4% charging the money put into the labs, −6.0% if the labs pay half from 2028</li></ul>
<h2>Both clouds bottomed in 2023; AI turned them back up</h2>
<figure><a href="https://06.ai/packs/cloud-capex-payback#growth"><img src="https://06.ai/newsletter/cloud-capex-payback/fig-01.png" alt="Both clouds bottomed in 2023; AI turned them back up" width="694" height="352"></a><figcaption>Revenue growth against the same quarter a year earlier. Azure is Microsoft’s “Azure and other cloud services” as reported at the time (fiscal 2024 as restated in August 2024); its fiscal year ends in June. On the narrower definition Microsoft adopted in September, the June 2026 quarter was 42%. Open dots are estimates: Azure 44% in both quarters (Microsoft guides 44–45% in constant currency on the new definition for the September quarter), AWS 32.5%, the middle of analysts’ 30–35% for the third quarter.</figcaption></figure>
<h2>What changed</h2>
<p><strong>Microsoft.</strong> Azure’s growth fell from 50% in the quarter to September 2021 to 26% in the quarter to June 2023, as customers cut their cloud bills. It has climbed since, to 43% in the quarter to June 2026 (42% on the definition Microsoft now reports), and Microsoft guides 44–45% in constant currency for the quarter to September. Azure passed $100B of revenue in the year to June. Microsoft says demand still exceeds the capacity it can bring online and expects to roughly double that capacity in two years.</p>
<p><strong>Amazon.</strong> AWS slowed from 40% at the end of 2021 to 12% in mid-2023. It grew about 17% in the first half of 2025, then 20%, 24%, 28% and, in the second quarter of 2026, 37%, its fastest in 18 quarters. Amazon says it won’t have enough capacity for all of 2026’s demand and expects the same in 2027. The whole market is moving: Synergy Research puts cloud infrastructure spending up 43% in the second quarter, with generative-AI services up 165%.</p>
<h2>Microsoft said how much of Azure’s growth was AI, until AI was 16 points of it</h2>
<figure><a href="https://06.ai/packs/cloud-capex-payback#azpts"><img src="https://06.ai/newsletter/cloud-capex-payback/fig-02.png" alt="Microsoft said how much of Azure’s growth was AI, until AI was 16 points of it" width="694" height="352"></a><figcaption>Azure revenue growth against a year earlier, split into AI services and the rest. Through the quarter to March 2025 the split is Microsoft’s own; after that it is our estimate, at a 17% trend for the rest (lighter). Growth is as Microsoft first reported it, so the split adds up; the quarters to September and December 2026 are estimates.</figcaption></figure>
<p>Microsoft’s points show AI rising from about 1 point of Azure’s growth in mid-2023 to 16 points by March 2025, when AI services were about 19% of Azure. The rest of Azure slowed as it did so, from 24 points of growth in the quarter to March 2024 to 17 a year later. Microsoft has since given only its AI business’s run-rate, which “surpassed an annual revenue run rate of $37 billion” in April 2026, Copilots included. At a 17% trend for the rest, AI adds about 20 points of Azure’s growth through 2025 and 26 by mid-2026, when it is about 34% of Azure. By the same chain, AI was already about 5% of Azure in 2022. Earlier capex paid for that, so for the returns we leave it, and its growth, out.</p>
<h2>By our estimate AWS will spend about as much on capex this year as it takes in</h2>
<figure><a href="https://06.ai/packs/cloud-capex-payback#capex"><img src="https://06.ai/newsletter/cloud-capex-payback/fig-03.png" alt="By our estimate AWS will spend about as much on capex this year as it takes in" width="694" height="332"></a><figcaption>Capex as a share of revenue. Microsoft: capex including finance leases, calendar years; 2026 is its April guide of about $190B, before it moved more data-center leases off capex (about $175B on the new basis). Amazon: cash capex plus property bought under finance leases; 2026 is its guide of about $220B plus about $4B of finance leases. AWS: its share of Amazon’s property additions applied to that capex; 2026 at our 75%. The ticks mark Microsoft’s and AWS’s 2019–23 intensity, the baseline we treat as non-AI (in the model it applies to no-AI revenue).</figcaption></figure>
<h2>The backlogs say the demand is there, and that much of it is two labs</h2>
<figure><a href="https://06.ai/packs/cloud-capex-payback#backlog"><img src="https://06.ai/newsletter/cloud-capex-payback/fig-04.png" alt="The backlogs say the demand is there, and that much of it is two labs" width="694" height="332"></a><figcaption>Contracted revenue not yet recognized, end of quarter, $B. Microsoft: commercial remaining performance obligation, which also covers Microsoft 365 and other commercial contracts; at the end of 2025, 45% of it was OpenAI’s. AWS: Amazon’s performance obligations, primarily AWS; by our arithmetic, OpenAI’s $138B and Anthropic’s more than $100B of commitments make up about half.</figcaption></figure>
<h2>Is growth accelerating?</h2>
<p><strong>Yes, at both, and capacity is the limit.</strong> Azure grew 39–40% for four quarters, then 43%, and Microsoft guides 44–45%. AWS has accelerated for five straight quarters. Microsoft’s commercial backlog reached $678B in June, up 84%; excluding OpenAI it grew 25%, and all of the quarter’s increase came from customers other than the frontier labs. AWS’s went from $195B to $496B in a year, with an average remaining life of 6.4 years, after OpenAI added $100B of commitments and Anthropic more than $100B.</p>
<h2>By our estimate AI adds about $39B to Azure this year and $31B to AWS</h2>
<p>Azure, 2026: $122B of revenue, $39B of it (32%) from AI at a 17% no-AI trend. AI adds about 26 of this year’s 43 points of growth.</p>
<figure><a href="https://06.ai/packs/cloud-capex-payback#uplift"><img src="https://06.ai/newsletter/cloud-capex-payback/fig-05.png" alt="By our estimate AI adds about $39B to Azure this year and $31B to AWS" width="694" height="362"></a><figcaption>Revenue against a no-AI trend. Azure’s follows Microsoft’s own split of AI and the rest in 2023 and 2024, with the AI Azure already had in 2022 kept on the no-AI side; AWS’s starts from 2023. The gap is above-trend revenue; we credit AI with 90% of it from 2025 (at AWS, from 2024). 2026 and 2027 are estimates (dashed): 2027 is what the capacity bought through 2027 can rent for, as in our base case. Choose the business and the no-AI trend.</figcaption></figure>
<h3>By business</h3>
<figure><a href="https://06.ai/packs/cloud-capex-payback#uplift"><img src="https://06.ai/newsletter/cloud-capex-payback/fig-06.png" alt="By our estimate AI adds about $39B to Azure this year and $31B to AWS — AWS" width="694" height="393"></a><figcaption><strong>AWS</strong>, 2026: $171B of revenue, $31B of it (18%) from AI at a 14.5% no-AI trend. AI adds about 17 of this year’s 33 points of growth.</figcaption></figure>
<h2>How we split AI from trend</h2>
<p>Both trends are checked against what the companies themselves disclose. <strong>Azure, 17% a year from 2025:</strong> by Microsoft’s own split, Azure outside AI grew about 24% in 2024 and was down to 19% by early 2025. At 17%, Azure’s AI services ran at about $30B a year in early 2026; add the Copilots and that fits Microsoft’s AI run-rate of more than $37B in April. <strong>AWS, 14.5%:</strong> it grew 13% in 2023 and 19% in 2024, when AI was a small business. At 14.5%, AWS’s AI revenue ran at about $20B a year in the first quarter of 2026 and $32B in the second; Amazon says it was “over $15 billion in Q1 2026” and above $25B in the second quarter. Much above 16%, our method falls short of Amazon’s own figures, so the simulation drops draws that do. The rest of each gap we leave to things that aren’t AI.</p>
<p>At these settings AI adds 19 points to Azure’s growth in 2025 and 26 in 2026, 32% of this year’s revenue, and 5 and 17 points to AWS’s, 18%. AI’s points are the growth in AI’s revenue over last year’s revenue, so they and the rest add up to the year’s growth. Microsoft’s Copilots aren’t counted: Microsoft 365’s commercial cloud revenue has grown 14–16% a year in constant currency since mid-2024, no faster than before Copilot.</p>
<h2>From the cloud alone, Amazon’s AI capex clears 10%; Microsoft’s falls well short</h2>
<p>At a 60% cash margin and the base no-AI trend (Azure 17%, AWS 14.5%): Amazon 13.6% through 2026 and 12.8% with 2027, 8.1% all-in; Microsoft 4.1%, 3.0% and 0.7%.</p>
<figure><a href="https://06.ai/packs/cloud-capex-payback#irr"><img src="https://06.ai/newsletter/cloud-capex-payback/fig-07.png" alt="From the cloud alone, Amazon’s AI capex clears 10%; Microsoft’s falls well short" width="694" height="392"></a><figcaption>Internal rate of return on AI capex with the stakes in OpenAI and Anthropic, and each company’s own models and assistants, at zero. Every dollar of capex above the pre-buildout intensity is charged. 2027 capex is our estimate: $225B at Microsoft on its old lease definition, $290B at Amazon with 75% to AWS. All-in adds the cash put into the labs (Amazon $68B, Microsoft about $14B), half of Microsoft’s neocloud contracts (the half we assume serves its own teams, about $38B over 2026–33) and running costs for the share of the fleet that doesn’t rent to customers. The line is a 10% hurdle.</figcaption></figure>
<h2>How the return is counted</h2>
<p><strong>Returns.</strong> Cloud revenue needs capacity, so after 2026 we tie AI revenue to it. We assume capex starts earning two quarters after it is spent: servers earn within weeks of landing, data centers a year or more after work starts. On that basis each dollar of incremental capex brought in about 40 cents a year of AI revenue at Amazon in 2026 and 26 cents at Microsoft. Microsoft’s is lower, probably because more of its capex has gone into things Azure’s revenue doesn’t show: its own models and Copilots, capacity for OpenAI’s research, and data-center sites ahead of use. At Microsoft we also leave out the 10% of Azure’s AI revenue we estimate runs on capacity it rents from neoclouds (it paid CoreWeave about $3.4B in 2025), and that capacity’s cost. Each dollar of capacity keeps renting at that rate, less 2.5% a year as compute gets cheaper, at a 60% cash margin before depreciation. That is a little above AWS’s 52% (operating income plus depreciation) in 2025; AWS’s operating margin was 39% in the second quarter of 2026, up from 33% a year earlier.</p>
<p><strong>Costs.</strong> All capex above a pre-buildout intensity: 14% of Microsoft’s no-AI revenue (its 2019–23 average, finance leases included) and 31% of AWS’s, about $449B at Microsoft and $367B at Amazon over 2023–27. Microsoft’s 2026 capex is its April guide, $190B, before it moved more data-center leases off capex. Servers, 60% of capex (Microsoft has said half to two-thirds), are replaced every 6 years at Microsoft and 5 at Amazon, at 75% of their cost, for the 75% of the fleet that rents to customers. Servers get bonus depreciation for tax (100% since 2025), data centers go over 20 years. Nothing is counted after 2040, though the data centers will still stand.</p>
<p><strong>Results.</strong> Through 2026: 13.6% at Amazon, 4.1% at Microsoft. Adding 2027’s build: 12.8% and 3.0%, paid back in 2033 and 2037. The cloud’s return barely depends on whether 2027 is spent, since new capacity earns what the old does as long as it rents at today’s rates; if 2027’s rents for 20% less, the returns fall to 10.4% and 1.3%. The biggest unknown is how long capex takes to start earning, because that sets how much each dollar in service earns today. If capex earns from one quarter after it is spent, more of it is already in service, each dollar earns less, and the returns fall to 6.2% at Amazon and −1.3% at Microsoft; from three quarters, they rise to 20.3% and 7.5%. Every ten points of margin move the returns four to five points. On the same question, Meta’s ads return about 18% and Alphabet’s ads and Cloud 13% (<a href="https://06.ai/packs/ai-capex-payback">the </a><a href="https://06.ai/packs/ai-capex-payback">ad-payback</a><a href="https://06.ai/packs/ai-capex-payback"> pack</a>).</p>
<h2>Payback comes in 2033 at Amazon and 2037 at Microsoft</h2>
<figure><a href="https://06.ai/packs/cloud-capex-payback#jcurve"><img src="https://06.ai/newsletter/cloud-capex-payback/fig-08.png" alt="Payback comes in 2033 at Amazon and 2037 at Microsoft" width="694" height="362"></a><figcaption>Cumulative after-tax cash flow of each AI program, $B: AI’s share of cloud cash profit, less incremental capex, server replacements and tax; capex through 2027. Dashed: all-in, with the cash put into the labs, running costs for the fleet that doesn’t rent out and, at Microsoft, half its neocloud contracts. The dots mark payback.</figcaption></figure>
<h2>If the labs can’t pay, Microsoft’s return goes negative and Amazon’s drops below 10%</h2>
<figure><a href="https://06.ai/packs/cloud-capex-payback#labs"><img src="https://06.ai/newsletter/cloud-capex-payback/fig-09.png" alt="If the labs can’t pay, Microsoft’s return goes negative and Amazon’s drops below 10%" width="694" height="342"></a><figcaption>Internal rate of return on AI capex through 2027 if, from 2028, OpenAI and Anthropic pay only part of their bills, and the capacity they stop paying for is re-let at half the price. We put the labs at 80% of Azure’s AI revenue and 50% of AWS’s. Dashed: all-in. The chart stops at −20%; the labels give the values below it.</figcaption></figure>
<h2>Two customers, and two stakes</h2>
<p><strong>The labs are much of the AI revenue.</strong> Microsoft says nearly 90% of Microsoft Cloud’s $214B of revenue in fiscal 2026 came from customers other than frontier-model companies. That leaves the labs about $22–26B, against our estimate of about $29B for all of Azure’s AI revenue that year, so we put them at 80% of it; OpenAI was 45% of Microsoft’s commercial backlog at the end of 2025. At Amazon, OpenAI’s and Anthropic’s commitments are about half the backlog, and we use 50%. The labs pay with money they raise: OpenAI’s own plan has it burning about $278B of cash over 2026–30 (<a href="https://06.ai/packs/ai-labs-funding">the </a><a href="https://06.ai/packs/ai-labs-funding">lab-funding</a><a href="https://06.ai/packs/ai-labs-funding"> pack</a>), though both labs’ revenue is compounding fast (<a href="https://06.ai/packs/ai-lab-revenue">the </a><a href="https://06.ai/packs/ai-lab-revenue">lab-revenue</a><a href="https://06.ai/packs/ai-lab-revenue"> pack</a>). If from 2028 they paid only half their bills, and the capacity they left were re-let at half price, Microsoft’s return through 2027 would fall to −6.0% and Amazon’s to 7.2%.</p>
<p><strong>The stakes are where the payoff might be.</strong> This pack counts them at zero. Amazon carried its Anthropic notes and preferred stock at about $190B at the end of June, against $18B invested, and has booked about $80B of pre-tax gains on them since early 2025; it has also put $50B into OpenAI, which committed another $100B to AWS. Microsoft’s roughly 27% of OpenAI was valued at about $135B in October 2025; OpenAI has since raised money at $852B. Charge the cash put into the labs and value the stakes at nothing, and the returns through 2027 are 8.6% at Amazon and 2.4% at Microsoft; all-in, with running costs and Microsoft’s neocloud contracts too, 8.1% and 0.7%. Count the stakes at their marks and the AI build has already paid for a good part of itself, on paper.</p>
<h2>Amazon’s build clears 10% in about two-thirds of our draws; Microsoft’s in fewer than one in ten</h2>
<figure><a href="https://06.ai/packs/cloud-capex-payback#odds"><img src="https://06.ai/newsletter/cloud-capex-payback/fig-10.png" alt="Amazon’s build clears 10% in about two-thirds of our draws; Microsoft’s in fewer than one in ten" width="694" height="344"></a><figcaption>Internal rate of return across 8,000 draws, each assumption uniform across its range, centered on the base case: the no-AI trend (Azure 14–20%, AWS 13–16%), AI’s share of the gap (80–100%), the lag from capex to revenue (one to three quarters), growth in the second half of 2026 (Azure 42–46%, AWS 30–35%), the share of Azure’s AI revenue on rented capacity (5–15%), the cash margin (50–70%), the fade in revenue per unit of capacity (0–5% a year), the share of the fleet that rents to customers (60–90%), the capex baseline (Microsoft 12–16%, AWS 28–34%), 2027 capex ($195–255B Microsoft, $250–330B Amazon), AWS’s share of Amazon’s capex (65–85%), server life (Microsoft 5–7 years, Amazon 4–6), replacement cost (60–90%), all-in running costs (2–4%) and, in the labs rows, the share of the labs’ bills paid from 2028 (50–100%) and their share of AI revenue (Microsoft 65–95%, Amazon 35–65%). AWS draws that contradict Amazon’s own AI run-rates are dropped (3% of them). Fixed: tax rates, the horizon, 2026 capex, the server share of capex (60%), what new capacity rents for (today’s rate) and the re-let price (half). Bars run from the 10th to the 90th percentile; the dot is the median.</figcaption></figure>
<h2>When does the cash come back?</h2>
<p>AI capex through 2027 clears a 10% return in 64% of our draws for Amazon and 8% for Microsoft, and clears zero in 97% and 72%. All-in, it clears 10% in 36% and 3%. The medians, 12.5% and 3.0%, sit within a point of the base case. Andy Jassy’s case is that the returns come later: “The FCF and ROIC for these investments are cumulatively quite attractive a couple years after being in service.” Amazon’s free cash flow was −$7.6B over the twelve months to June. Microsoft’s was $19.6B in the June quarter, down 23% on a year earlier, and it guides capex up again in fiscal 2027.</p>
<h2>What to watch</h2>
<p><strong>Microsoft, October 28.</strong> Azure against the 44–45% guide, now on the narrower definition; capex against “over $50 billion” for the quarter and any figure for fiscal 2027; whether the backlog keeps growing outside OpenAI.</p>
<p><strong>Amazon, October 29.</strong> AWS against analysts’ 30–35%; capex against $220B for the year and a first word on 2027; the backlog after the OpenAI and Anthropic expansions; the AI run-rate, last above $25B.</p>
<p><strong>The labs’ funding.</strong> OpenAI’s next round and Anthropic’s IPO decide whether the bills behind half of AWS’s backlog, and most of Azure’s AI revenue, get paid. In our base case a lab dollar is worth the same as any other; the labs chart shows what happens when it isn’t.</p>
<h2>What this is anchored on</h2>
<p>No-AI trends, AI’s share of above-trend revenue, the lag, the second half of 2026, 2027 capex, margins, replacement, rented capacity, the labs’ share of AI revenue, timing of the all-in costs and the returns are our estimates, built from the sources below.</p>
<ul><li><strong>Microsoft FY26 Q4</strong> — Revenue $90.0B, +18%, and $331.8B for the fiscal year; Azure and other cloud services +43%; Azure revenue passed $100B for the first time; free cash flow $19.6B, down 23% · <a href="https://www.sec.gov/Archives/edgar/data/0000789019/000119312526323632/msft-ex99_1.htm">Microsoft FY26 Q4 results (8-K), Jul 29</a></li><li><strong>Microsoft FY26 Q4 call</strong> — Capex $41B, about two-thirds short-lived assets (GPUs and CPUs), including $5.6B of finance leases; data-center useful life extended from 15 to 25 years, so more future data-center leases will be operating leases, outside capex; calendar 2026 capex about $175B on that basis, plans otherwise unchanged; fiscal Q1 2027: Azure about 45% in constant currency, capex over $50B; commercial RPO $678B, +84%, +25% excluding OpenAI, all of the quarter’s increase from customers outside frontier-model companies; Microsoft Cloud gross margin 65%; over 30M paid Microsoft 365 Copilot seats; demand exceeds capacity; capacity to roughly double in two years · <a href="https://www.fool.com/earnings/call-transcripts/2026/08/07/microsoft-msft-q4-2026-earnings-call-transcript/">Microsoft FY26 Q4 call, via The Motley Fool</a></li><li><strong>Microsoft FY26 Q4 call</strong> — Nearly 90% of fiscal 2026 Microsoft Cloud revenue came from customers outside frontier-model companies; a $3.2B gain from the Anthropic investment · <a href="https://www.marketbeat.com/instant-alerts/microsoft-q4-earnings-call-highlights-2026-07-29/">MarketBeat, Jul 29</a></li><li><strong>Microsoft, Sep 2026</strong> — Azure revenue in dollars on a narrower definition: fiscal 2025 $72.6B (quarters $16.0B, $17.3B, $18.5B, $20.7B), fiscal 2026 $101.9B ($22.4B, $24.1B, $26.0B, $29.4B); June 2026 quarter +42% on this basis; fiscal Q1 2027 guide 44–45% in constant currency · <a href="https://www.sec.gov/Archives/edgar/data/0000789019/000119312526380280/d291965dex991.htm">Microsoft FY27 segments and metrics (8-K), Sep 2026</a></li><li><strong>Microsoft FY26 Q3</strong> — AI business “surpassed an annual revenue run rate of $37 billion”, up 123% · <a href="https://www.sec.gov/Archives/edgar/data/0000789019/000119312526191457/msft-ex99_1.htm">Microsoft FY26 Q3 results (8-K), Apr 29</a></li><li><strong>Microsoft FY26 Q3 call</strong> — Capex $31.9B; calendar 2026 capex about $190B, including about $25B from higher component prices · <a href="https://sherwood.news/tech/microsoft-is-planning-to-spend-190-billion-on-capex-this-year/">Sherwood News, Apr 30</a></li><li><strong>Microsoft FY26 Q2 call</strong> — “Approximately 45% of our commercial RPO balance is from OpenAI,” of $625B · <a href="https://www.roic.ai/quote/MSFT:AV/transcripts/2026-year/2-quarter">Microsoft FY26 Q2 call, via roic.ai</a></li><li><strong>Microsoft AI points</strong> — Azure growth included 13 points from AI services, which grew 157% (December 2024 quarter) · <a href="https://www.marketbeat.com/earnings/reports/2025-1-29-microsoft-co-stock">Microsoft FY25 Q2 call, via MarketBeat</a></li><li><strong>Microsoft AI points</strong> — 16 points from AI services in the March 2025 quarter · <a href="https://www.marketbeat.com/earnings/reports/2025-4-30-microsoft-co-stock">Microsoft FY25 Q3 call, via MarketBeat</a></li><li><strong>Microsoft AI points</strong> — Roughly 12 points from AI services in the September 2024 quarter, “similar to last quarter” · <a href="https://www.nasdaq.com/articles/microsoft-msft-q1-2025-earnings-call-transcript">Microsoft FY25 Q1 call, via Nasdaq</a></li><li><strong>Microsoft AI points</strong> — 8 points in the June 2024 quarter; 7 and 6 in the March 2024 and December 2023 quarters; roughly 3 in September 2023 and 1 in June 2023 · <a href="https://www.microsoft.com/en-us/investor/events/FY-2024/earnings-fy-2024-q4">Microsoft earnings calls, Jul 2023–Jul 2024</a></li><li><strong>Microsoft FY24 recast</strong> — Azure growth for fiscal 2024 restated to 31%, 33%, 35% and 34% by quarter · <a href="https://www.sec.gov/Archives/edgar/data/789019/000119312524204403/d846847dex991.htm">Microsoft metric recast (8-K), Aug 2024</a></li><li><strong>Microsoft capex</strong> — Capex including finance leases by quarter, $6.6B (September 2022) to $41B (June 2026); finance leases $11.1B, $6.7B, $4.7B and $5.6B in the quarters of fiscal 2026, mostly for large data-center sites · <a href="https://www.fool.com/earnings/call-transcripts/2025/10/29/microsoft-msft-q1-2026-earnings-call-transcript">Microsoft earnings calls, fiscal 2023–26</a></li><li><strong>Microsoft baseline</strong> — Cash capex from $13.9B in fiscal 2019 to $28.1B in fiscal 2023, about 12% of revenue; in fiscal 2023, $31.9B including finance leases. Our 14% baseline adds finance leases · <a href="https://www.microsoft.com/en-us/investor/events/FY-2023/earnings-FY-2023-Q4">Microsoft results and calls, fiscal 2019–23</a></li><li><strong>Microsoft leases</strong> — Data-center leases signed but not yet commenced of $329.1B at June 30, 2026, starting from fiscal 2027 through fiscal 2033 · <a href="https://btw.media/en/microsoft-329-1bn-uncommenced-datacentre-leases">btw.media, on the fiscal 2026 10-K</a></li><li><strong>Microsoft–OpenAI</strong> — Microsoft holds about 27% of OpenAI Group PBC, valued at about $135B; OpenAI contracted an incremental $250B of Azure services; Microsoft loses its right of first refusal on compute · <a href="https://blogs.microsoft.com/blog/2025/10/28/the-next-chapter-of-the-microsoft-openai-partnership/">Microsoft, Oct 28, 2025</a></li><li><strong>Microsoft–OpenAI</strong> — Microsoft stops paying OpenAI a revenue share; OpenAI’s payments to Microsoft continue through 2030 under a cap; Microsoft’s license becomes non-exclusive · <a href="https://openai.com/index/next-phase-of-microsoft-partnership">OpenAI, Apr 27</a></li><li><strong>Microsoft–OpenAI</strong> — $13B of funding commitments to OpenAI, $11.8B funded at Mar 31, 2026 · <a href="https://www.sec.gov/Archives/edgar/data/0000789019/000119312526191507/msft-20260331.htm">Microsoft 10-Q, quarter to Mar 31</a></li><li><strong>Microsoft–Anthropic</strong> — Anthropic committed to buy $30B of Azure compute; Microsoft to invest up to $5B in Anthropic · <a href="https://blogs.microsoft.com/blog/2025/11/18/microsoft-nvidia-and-anthropic-announce-strategic-partnerships/">Microsoft, Nov 18, 2025</a></li><li><strong>Microsoft neoclouds</strong> — Dedicated GPU capacity for Microsoft worth about $17.4B through 2031, up to about $19.4B with additional services · <a href="https://www.sec.gov/Archives/edgar/data/1513845/000110465925088312/tm2525580d1_6k.htm">Nebius (6-K), Sep 2025</a></li><li><strong>Microsoft neoclouds</strong> — A five-year, $9.7B AI cloud contract with Microsoft · <a href="https://www.sec.gov/Archives/edgar/data/1878848/000114036125040072/ef20058139_ex99-1.htm">IREN (8-K), Nov 2025</a></li><li><strong>Microsoft neoclouds</strong> — Neocloud contracts above $60B: Nscale about $23B, Nebius up to $19.4B, IREN $9.7B, Lambda and CoreWeave; mostly five-year terms · <a href="https://introl.com/blog/microsoft-60-billion-neocloud-spending-2025">Introl, citing Bloomberg, Nov 2025</a></li><li><strong>Microsoft neoclouds</strong> — The Nebius capacity is for Microsoft’s internal teams building models and a consumer assistant · <a href="https://finance.yahoo.com/news/microsoft-inks-33-billion-deals-191655037.html">Tom’s Hardware, citing Bloomberg, Oct 2025</a></li><li><strong>CoreWeave</strong> — About 67% of 2025 revenue of $5.1B came from Microsoft · <a href="https://www.sec.gov/Archives/edgar/data/1769628/000176962826000104/crwv-20251231.htm">CoreWeave 10-K, 2025</a></li><li><strong>Amazon Q2 2026</strong> — AWS revenue $42.2B, +37% (36.7%), operating income $16.6B; “Exceeded a $25 billion annual revenue run rate for AWS’s AI business”, and the same for Amazon’s chips business; Q3 guidance: revenue $197–202B, operating income $22.5–26.5B; free cash flow −$7.6B over twelve months · <a href="https://www.sec.gov/Archives/edgar/data/0001018724/000101872426000024/amzn-20260630xex991.htm">Amazon Q2 2026 results (8-K), Jul 30</a></li><li><strong>Amazon Q2 2026</strong> — Performance obligations, primarily AWS, about $496B with a 6.4-year average remaining life; OpenAI expanded its $38B commitment by $100B over 8 years, Anthropic its own by more than $100B over 10 years; $50B invested in OpenAI ($21.3B of it after June 30); $8B of Anthropic notes plus $10B of preferred, carried at about $190B · <a href="https://www.sec.gov/Archives/edgar/data/0001018724/000101872426000026/amzn-20260630.htm">Amazon 10-Q, quarter to Jun 30</a></li><li><strong>Amazon Q2 2026 call</strong> — Cash capex $53.1B in the quarter; about $220B for 2026, up from about $200B on the higher cost of memory; “we will still not have enough capacity to meet all the demand we have in 2026”, and likely the same in 2027 · <a href="https://www.fool.com/earnings/call-transcripts/2026/08/07/amazon-amzn-q2-2026-earnings-call-transcript/">Amazon Q2 2026 call, via The Motley Fool</a></li><li><strong>Amazon letter</strong> — “AWS’s AI revenue run rate is over $15 billion in Q1 2026”; “The FCF and ROIC for these investments are cumulatively quite attractive a couple years after being in service” · <a href="https://www.aboutamazon.com/md/news/company-news/amazon-ceo-andy-jassy-2025-letter-to-shareholders">Andy Jassy, 2025 letter to shareholders, Apr 9</a></li><li><strong>Amazon Q1 2026</strong> — AWS revenue $37.6B, +28%; chips business above a $20B run-rate; backlog about $364B · <a href="https://www.sec.gov/Archives/edgar/data/0001018724/000101872426000012/amzn-20260331xex991.htm">Amazon Q1 2026 results (8-K), Apr 29</a></li><li><strong>Amazon 2025</strong> — AWS revenue $128.7B, operating income $45.6B; 2025 cash capex $128.3B; about $200B of capex planned for 2026 · <a href="https://ir.aboutamazon.com/news-release/news-release-details/2026/Amazon-com-Announces-Fourth-Quarter-Results/">Amazon Q4 2025 results, Feb 5</a></li><li><strong>Amazon segments</strong> — AWS net additions to property and equipment $24.8B, $53.3B and $96.5B in 2023–25; AWS depreciation and amortization $21.5B in 2025 · <a href="https://www.sec.gov/Archives/edgar/data/1018724/000101872426000004/amzn-20251231.htm">Amazon 10-K, 2025 (segment note)</a></li><li><strong>Amazon segments</strong> — AWS net additions to property and equipment $13.1B, $16.5B, $22.0B and $27.8B in 2019–22 · <a href="https://www.sec.gov/Archives/edgar/data/1018724/000101872423000004/R30.htm">Amazon 10-Ks, 2020 and 2022 (segment notes)</a></li><li><strong>Amazon servers</strong> — Useful life of a subset of servers and networking equipment cut from six to five years from January 2025 · <a href="https://www.sec.gov/Archives/edgar/data/1018724/000101872425000004/Financial_Report.xlsx">Amazon 10-K, 2024</a></li><li><strong>Amazon–OpenAI</strong> — $15B invested, $35B more on conditions; the $38B AWS agreement expanded by $100B over 8 years, about 2 GW of Trainium · <a href="https://openai.com/index/amazon-partnership/">OpenAI, Feb 27</a></li><li><strong>Amazon–Anthropic</strong> — Anthropic commits more than $100B over ten years to AWS, up to 5 GW; Amazon invests $5B, up to $20B more · <a href="https://www.anthropic.com/news/anthropic-amazon-compute">Anthropic, Apr 20</a></li><li><strong>AWS Q3</strong> — Analysts expect AWS growth of 30% to 35% in the third quarter; Amazon reports October 29 · <a href="https://www.benzinga.com/trading-ideas/movers/26/10/62522059/amazon-earnings-are-a-week-away-heres-what-to-expect">Benzinga, Oct 22</a></li><li><strong>Microsoft FY27 Q1</strong> — Results after the close on October 28 · <a href="https://finviz.com/news/400058/microsoft-announces-quarterly-earnings-release-date">Microsoft, via Finviz</a></li><li><strong>Cloud market</strong> — Q2 2026 cloud infrastructure spending $143.4B, +43%, the fastest in eight years; GenAI-specific services +165% · <a href="https://www.srgresearch.com/articles/q2-cloud-market-passes-143-billion-highest-growth-rate-in-eight-years">Synergy Research, Jul 30</a></li><li><strong>OpenAI</strong> — Plan of July 2026: negative free cash flow of about $278B over 2026–30 · <a href="https://www.thestar.com.my/tech/tech-news/2026/09/19/openai-forecasts-cash-burn-near-280-billion-by-2030-ft-reports">FT, via Reuters, Sep 18</a></li><li><strong>OpenAI</strong> — Round closed March 31 at $122B and an $852B valuation · <a href="https://techcrunch.com/2026/03/31/openai-not-yet-public-raises-3b-from-retail-investors-in-monster-122b-fund-raise/">TechCrunch, Mar 31</a></li><li><strong>Anthropic</strong> — At least $518B of compute commitments over about a decade, including Amazon $110B and Microsoft $31.4B · <a href="https://www.zawya.com/en/business/anthropics-518bln-ai-buildout-hinges-largely-on-deals-that-cannot-be-canceled-filing-shows-1392642">Reuters, on the draft prospectus, via Zawya</a></li><li><strong>History</strong> — Microsoft quarterly revenue, Azure growth, capex and RPO, fiscal 2022–26; AWS quarterly revenue 2019–2026 · <a href="https://www.microsoft.com/en-us/investor">Microsoft and Amazon earnings releases and filings</a></li></ul>
<h2>Method</h2>
<p><strong>Clock.</strong> Reported through June 2026: Microsoft’s fiscal 2026 and Amazon’s second quarter, both mapped to calendar years. Azure’s revenue is Microsoft’s new dollar series for fiscal 2025–26; earlier quarters are backed out from the growth it reported. The second half of 2026: Azure up 44% and AWS up 32.5%. Capex 2026: Microsoft $190B (its guide before the lease change), Amazon $220B plus finance leases with 75% to AWS; 2027 our estimates.</p>
<p><strong>AI’s share.</strong> Azure: AI services as Microsoft split them through March 2025, less the AI Azure already had in 2022 grown with the rest of Azure; after that, 90% of revenue above a 17% trend. AWS: 90% of revenue above 2023’s revenue grown at 14.5% a year, consistent with Amazon’s AI run-rates. Trends hold through 2027 and fade to 6% by 2035.</p>
<p><strong>Returns.</strong> Cash flows from 2022 to 2040, no terminal value. In: AI revenue at a 60% cash margin. Through 2026 it is the estimate above (at Microsoft, less the 10% we put on rented capacity); after that, each dollar of incremental capex earns what 2026’s AI revenue implies per dollar of capex then in service (26 cents at Microsoft, 40 cents at Amazon), from two quarters after it is spent, fading 2.5% a year. Out: capex above 14% of Microsoft’s no-AI revenue and 31% of AWS’s; servers, 60% of capex, replaced every 6 or 5 years at 75% of their cost for the 75% of the fleet that rents to customers; cash tax at 19% and 17%, with servers getting first-year bonus depreciation (80% in 2023, 60% in 2024, 100% since; the rest over five years) and data centers over 20 years, losses offsetting other income. All-in adds the cash put into the labs (Amazon: $8B of Anthropic notes over 2023–25, $10B of preferred and $50B into OpenAI in 2026; Microsoft: about $9B of its 2023 OpenAI commitment over 2023–25, our timing, and up to $5B into Anthropic), half of Microsoft’s neocloud contracts (about $75B over 2026–33, our timing), and running costs of 3% a year of the cost of the fleet that doesn’t rent to customers, while its servers last. The labs cases cut the labs’ payments from 2028 and re-let what they stop paying for at half price. The base case is the middle of every range in the simulation.</p>
<p><strong>What’s left out.</strong> The stakes in OpenAI and Anthropic, and OpenAI’s revenue share to Microsoft. Microsoft’s Copilots (6.0% if counted at $7–10B a year), Amazon’s use of AI in retail and ads, and every company model or assistant. Operating leases Microsoft signs from fiscal 2027 sit outside its reported capex; our 2026–27 figures keep the old definition. Monte Carlo: 8,000 draws per case with a fixed seed. Model and data: scripts/packs/cloud_payback.py in the site’s repository.</p>
<hr>
<p>Every chart above links to its interactive version on 06.ai, with the data behind it. The daily trackers are at <a href="https://06.ai/trackers">06.ai/trackers</a>.</p>
<p><em>Not investment advice. Estimates are labeled as estimates.</em></p>]]></content:encoded>
</item>
<item>
  <title>How much memory can each AI chip get?</title>
  <link>https://inthesingularity.ai/packs/ai-memory-balance</link>
  <guid isPermaLink="true">https://inthesingularity.ai/packs/ai-memory-balance</guid>
  <pubDate>Sun, 25 Oct 2026 12:12:00 GMT</pubDate>
  <description>AI accelerator shipments nearly double by 2030, and HBM supply can’t keep up with the memory per chip on today’s roadmaps. We balance HBM supply and demand to 2030, size the memory per accelerator the market can afford, and weigh KV cache, Marvell’s optical pooling, China and Terafab.</description>
  <category>AI memory</category>
  <enclosure url="https://inthesingularity.ai/newsletter/ai-memory-balance/cover.png" length="67517" type="image/png" />
  <media:content url="https://inthesingularity.ai/newsletter/ai-memory-balance/cover.png" medium="image" type="image/png" />
  <content:encoded><![CDATA[<figure><a href="https://06.ai/packs/ai-memory-balance"><img src="https://06.ai/newsletter/ai-memory-balance/thumb.png" alt="How much memory can each AI chip get?" width="421" height="432"></a><figcaption><a href="https://06.ai/packs/ai-memory-balance">Interactive version and data on 06.ai →</a></figcaption></figure>
<p>AI accelerator shipments roughly double from about 14M packages in 2026 to 26.5M in 2030, and HBM can’t grow fast enough to give each one the memory on today’s roadmaps. At announced specs, HBM supply falls <strong>18% short</strong> of demand in 2027 and 29–38% short in 2028–30. To balance, the average accelerator can carry about <strong>270 GB</strong> of HBM in 2027 and <strong>430 GB</strong> in 2030, growth of 17% a year against 28% on the roadmaps. The trims now being reported, from Rubin Ultra’s cut to 8-Hi stacks on next year’s ASICs, do roughly that; price does the rest.</p>
<ul><li><strong>~270 GB</strong> — HBM per accelerator, 2027 the fleet average supply can support. Rubin and B300 carry 288 GB</li><li><strong>~430 GB</strong> — HBM per accelerator, 2030. +17% a year from 2027, against +28% on announced roadmaps (~690 GB)</li><li><strong>29% short</strong> — At announced specs, 2028. HBM supply against demand: 18% short in 2027, 38% by 2030. At the trimmed specs, within ±3% every year</li><li><strong>−22% to +52%</strong> — Biggest swing, 2028 unit volumes, from the vendors’ deal books to a power-limited build, at trimmed specs</li></ul>
<h2>At announced specs, HBM supply falls 18–38% short from 2027; the reported trims close the gap</h2>
<p>Base units, base supply: at announced specs supply is 18% short in 2027 and 38% short in 2030; at trimmed specs 2% short in 2027 and 2% short in 2030.</p>
<figure><a href="https://06.ai/packs/ai-memory-balance#balance"><img src="https://06.ai/newsletter/ai-memory-balance/fig-01.png" alt="At announced specs, HBM supply falls 18–38% short from 2027; the reported trims close the gap" width="694" height="382"></a><figcaption>HBM made by SK hynix, Samsung and Micron each year against what accelerators need, in exabytes (EB). Demand counts HBM bought about four months before the accelerator ships, plus about 6% for stacks lost in assembly. Labels: supply ÷ demand − 1 at announced specs. Choose the unit and supply cases.</figcaption></figure>
<h2>What the balance says</h2>
<p><strong>How we count.</strong> Supply is HBM wafer starts times the good terabytes each wafer yields. Demand is units times HBM per package, bought about four months before the accelerator ships, plus about 6% for stacks lost in 2.5D assembly and physical capacity above the usable figure. The lead matters when growth is steep: purchases run 37% above the bits inside 2026’s shipments, because 2027’s step-up is bought early, and 13–17% above in 2027–30.</p>
<p><strong>2026 is roughly balanced (−2%).</strong> That fits the price signal: HBM contract prices paused or slipped in 2026 while commodity DRAM rose about sixfold. In 2025, supply (about 2.8 EB) ran well ahead of the bits shipped inside accelerators. The excess went into inventory across the chain: Nvidia’s own rose from about $10B to $21B over its fiscal 2026 and reached $31.6B by July 2026.</p>
<p><strong>2027 is the crunch at announced specs.</strong> Demand jumps with TPU 8, Trainium 3–4, MI450, MTIA and the first Rubin Ultra builds. Supply grows only 54%: wafer input rises 31% and good bits per wafer 18%, less than node shrinks alone would give because the HBM4 die is larger for its doubled interface. The announced specs would need 7.3 EB against about 6.0 EB available.</p>
<p><strong>2028–30 turns on two things we can’t know yet:</strong> how fast memory per package grows again, and how many packages the power grid can absorb. New cleanrooms lift supply 33% in 2028 and about 24% a year after. If memory per package grows about 17% a year, the market stays balanced; if it follows the announced roadmaps, supply falls more than a third short.</p>
<h3>How this compares with published views</h3>
<figure><a href="https://06.ai/packs/ai-memory-balance"><img src="https://06.ai/newsletter/ai-memory-balance/fig-02.png" alt="Goldman Sachs (Jun) — 2026: −5.4% · 2027: −6.0% · 2028: −4.3%; J.P. Morgan (Sep) — −16% across 2026–28; SemiAnalysis (via Luminix) — 2026: ~−6% · 2027: ~−9%; This pack, trimmed specs — 2026: −2% · 2027: −2% · 2028: +3%; This pack, announced specs — 2026: −6% · 2027: −18% · 2028: −29%" width="712" height="289"></a></figure>
<p>Supply ÷ demand − 1. Goldman sits between our two cases; J.P. Morgan’s deeper −16% came in September, after the first trims were reported.</p>
<p><strong>The commodity DRAM squeeze is the same story one step removed.</strong> With nearly every new big-three wafer going to HBM, conventional bit growth has to come from node shrinks and CXMT. AI then takes a second bite through CPU memory: Vera’s LPDDR5X, AMD Helios’s 36 TB of DDR5 per rack, and agentic CPU servers. Counting HBM at its wafer cost, AI systems take roughly 27% of DRAM wafers in 2026 and about 43% in 2027, by our estimate from the rack configurations below. Conventional contract prices compounded roughly sixfold from 3Q25 to 4Q26, with 4Q26 still up 10–15% (TrendForce). IDC expects 2026 PC shipments to fall 11.3% and smartphones 13%. Published 2027 DRAM sufficiency ranges from −1.9% (DS Investment) to −5.9% (Goldman); Micron calls 2027–28 “tighter than 2026”.</p>
<h2>Is Nvidia really cutting memory per GPU?</h2>
<p>Partly. No shipping data-center GPU has lost HBM: Rubin keeps 288 GB, the same as Blackwell Ultra. But Nvidia is reported to be halving Vera’s CPU memory for 2027 builds and cutting Rubin Ultra’s HBM by 75–81% per package against its announced spec. HBM per GPU has stopped growing this generation, and per compute die it is set to fall in the next.</p>
<ul><li><strong>Rubin (VR200), shipping since Aug 2026</strong> — What changed: Still 288 GB of HBM4 on eight 12-Hi stacks. Listed bandwidth fell from 22 TB/s (CES) to 19.2 TB/s, no reason given · Status, Oct 2026: Nvidia’s page. A ZDNet Korea report of 8-Hi orders was contradicted by Counterpoint, which says most Rubin stays 12-Hi</li><li><strong>Vera CPU (LPDDR5X on SOCAMM2)</strong> — What changed: 1.5 TB → 768 GB per CPU; ~54 → ~28 TB per NVL72 rack · Status, Oct 2026: Reported for 2027 supply plans by TrendForce, Daishin and SemiAnalysis (June). Nvidia calls 768 GB a “flexible SKU”; its spec still says “up to 1.5 TB”</li><li><strong>Rubin Ultra (2027–28)</strong> — What changed: Announced: 1 TB of HBM4E on 4 dies. Reported: ~192–256 GB of 8-Hi HBM4 on 2 dies · Status, Oct 2026: SemiAnalysis; The Information. Nvidia: “our roadmap is intact”; no confirmation</li><li><strong>Rubin CPX (128 GB GDDR7)</strong> — What changed: Absent from Nvidia’s October Vera Rubin page; the SRAM-based Groq 3 LPX rack appears instead · Status, Oct 2026: No source ties this to the GDDR7 shortage</li><li><strong>GeForce RTX 50</strong> — What changed: Fewer cards (a 30–40% cut reported for the 16 GB RTX 5070 Ti and 5060 Ti in 1H26), not less memory per card · Status, Oct 2026: Nvidia: “memory supply is constrained”</li></ul>
<p><strong>Per compute die, the cut is real.</strong> Rubin carries 144 GB per die; the reported Rubin Ultra carries 96 GB per die, a third less, against 256 GB per die in the original plan. Half of the package-level drop comes from going from four dies to two, which sources put down to packaging and rack issues rather than memory.</p>
<p><strong>It is not just Nvidia.</strong> AMD reportedly built a 288 GB MI450 for Meta beside the 432 GB MI455X. TrendForce says GPU and ASIC makers are “considering reducing HBM capacity per device” and will evaluate 8-Hi stacks in 2027. The tell: 8-Hi stacks now carry a 10–30% per-gigabit premium over 12-Hi, the reverse of normal pricing.</p>
<p><strong>Why.</strong> SemiAnalysis gives three reasons for Rubin Ultra. There are not enough HBM wafers to build 12-Hi stacks for the logic and CoWoS capacity Nvidia has secured for 2027, and 8-Hi yields more stacks per wafer. HBM cost jumps in 2027: GF Securities estimates memory would have reached about 29% of a Vera Rubin system’s bill of materials without the Vera cut, against Nvidia’s preferred 20%. And workloads are shifting: a 576-GPU domain of 192 GB parts (about 110 TB) still dwarfs a 20.7 TB NVL72.</p>
<p>Nvidia has guided its gross margin down from 75% to 71–72% by the quarter ending January 2027 on “extreme pricing conditions in memory”. Its supply and capacity purchase commitments, mainly memory, rose from $119B to $279B in one quarter. The logic is to protect units by trimming memory per unit: TrendForce describes the Vera cut as “lowering memory per module while increasing total module shipments”, a supply-driven compromise, not a drop in demand. We model the reported trims as the base case and the announced specs as the alternative.</p>
<h2>Spec trims cut HBM demand by 1.2 EB in 2027 and 3.5 EB in 2028, mostly at Nvidia</h2>
<figure><a href="https://06.ai/packs/ai-memory-balance#trims"><img src="https://06.ai/newsletter/ai-memory-balance/fig-03.png" alt="Spec trims cut HBM demand by 1.2 EB in 2027 and 3.5 EB in 2028, mostly at Nvidia" width="694" height="313"></a><figcaption>From HBM demand at announced specs to base-case supply, EB bought in the year, base-case units. Each family moves from its announced to its trimmed specs in turn. Pink lowers the running total, green raises it.</figcaption></figure>
<h2>Supply supports about 270 GB per accelerator in 2027 and 430 GB in 2030</h2>
<figure><a href="https://06.ai/packs/ai-memory-balance#per-package"><img src="https://06.ai/newsletter/ai-memory-balance/fig-04.png" alt="Supply supports about 270 GB per accelerator in 2027 and 430 GB in 2030" width="694" height="430"></a><figcaption>Fleet-average HBM per accelerator package, GB. Supportable: each year’s HBM supply divided by the packages it serves, 65% of that year’s plus 35% of next year’s (about a third of a package’s HBM is bought the year before it ships), after assembly losses. The announced and trimmed paths are averaged the same way, so each gap is that path’s supply-vs-demand. Whiskers: low and high supply cases. Shown: Base.</figcaption></figure>
<h2>How much memory per accelerator the market can afford</h2>
<p>To balance, fleet-average HBM per accelerator can be about 270 GB in 2027, 320 GB in 2028 and 430 GB in 2030: growth of about 17% a year, against 28% on the announced roadmaps. That is feasible. 288 GB stays the flagship through 2027, next-generation parts move to 8-Hi stacks, and capacity beyond that moves off HBM. It is a fleet average, so a 432 GB part has to be offset by a 144–216 GB part elsewhere.</p>
<ul><li><strong>2027 (~270 GB).</strong> Rubin, TPU 8i and the 288 GB MI450 sit at the top of the affordable range. The 432 GB MI455X and the 384–512 GB MTIA 500 fit only if Trainium 3 (144 GB), TPU 8t (216 GB) and Maia 200 (216 GB) keep the average down.</li><li><strong>2028 (~320 GB).</strong> Rubin Ultra at 192–256 GB on two dies fits. The announced four-die, 1 TB version would absorb about a fifth of 2028 supply on about 15% of units.</li><li><strong>2029–30 (~380–430 GB).</strong> Room returns for Feynman-class parts at 400–600 GB, as long as custom ASICs stay near 300–420 GB.</li></ul>
<p>Per compute die, that means about 100–130 GB on Rubin Ultra (Rubin: 144 GB) and roughly 200 GB per die by 2030.</p>
<p><strong>Workloads can live with it.</strong> MLPerf Inference v6.1 compares GB300 (278 GB) with GB200 (186 GB) at the same 8 TB/s. A 671B model and small models that can’t be capacity-bound gained the same ~1.29×, so the gain from capacity itself was about zero. Our modelling of a one-third cut, 288 GB to 192 GB through 8-Hi stacks at unchanged bandwidth:</p>
<ul><li><strong>0–10% throughput loss</strong> for today’s dominant pattern: 0.7–2T-parameter MoE models on 72 or more GPUs with wide expert parallelism, 32K–400K context. Weights take about 20–45 GB per GPU at that width, depending on precision and replication; the rest is KV cache.</li><li><strong>15–30% loss</strong> for 10T-class models, sparse-attention long-context agents, or idle sessions parked in HBM. About 45% in the worst combination.</li><li><strong>0–12%</strong> for training.</li><li><strong>Concurrency falls more than throughput.</strong> Sessions per GPU drop 36–41%, because weights are a fixed charge, but per-user decode speeds up, so tokens per GPU fall only 5–11%. Token-priced services barely notice; session-priced agent hosting does, unless KV moves to another tier.</li></ul>
<p>The hard floor is about 155 GB per GPU, below which a 10T-parameter model at FP8 no longer fits in a 72-GPU domain (about 90 GB at FP4). These figures assume FP4 weights, FP8-or-smaller KV cache and an off-HBM KV tier, all shipping on Blackwell and Rubin.</p>
<p><strong>The economics push the same way.</strong> At 2027 HBM4 prices of $32–40/GB, 288 GB costs about $9–12K per package and 1 TB about $33–41K. Morgan Stanley puts memory at 25–30% of a Vera Rubin rack’s cost, up from 5–10% in earlier racks. A 12-Hi to 8-Hi switch cuts bits per GPU by a third; with 8-Hi’s 10–30% per-Gb premium, HBM spend per GPU falls 13–27%.</p>
<p><strong>The technology is not the limit; wafers are.</strong> Per-stack bandwidth depends on interface width and pin speed, not height, so a shorter stack keeps bandwidth: B200 to B300 proved the inverse, 192 GB to 288 GB on the same eight stacks at 8 TB/s. Packaging allows far more than the market can feed: 12 stacks × 48 GB = 576 GB in 2027, and 20 × 64 GB ≈ 1.3 TB by 2028–29 on TSMC’s roadmap.</p>
<h2>Units nearly double from 2026 to 26.5M in 2030; custom ASICs are about 60% of them from 2027</h2>
<figure><a href="https://06.ai/packs/ai-memory-balance#units"><img src="https://06.ai/newsletter/ai-memory-balance/fig-05.png" alt="Units nearly double from 2026 to 26.5M in 2030; custom ASICs are about 60% of them from 2027" width="694" height="402"></a><figcaption>AI accelerator packages shipped a year, millions, base case. Whiskers: total units in the low case (a power- and capex-limited build) and the high case (the vendors’ deal books at face value). Row below: the custom ASICs’ share.</figcaption></figure>
<h2>Demand: units, memory per unit, and each generation’s appetite</h2>
<p>Accelerator shipments roughly double from about 14M packages in 2026 to 26.5M in 2030 in our base case, and custom ASICs (TPU, Trainium, MTIA, OpenAI’s chip) make up about 60% of units from 2027. At announced specs, the HBM inside each year’s shipments would grow sixfold, from 2.9 EB in 2026 to about 17.5 EB in 2030; at the specs now shipping or reported, about fourfold, to 11.1 EB.</p>
<p><strong>Units.</strong> Our 2026–27 paths follow vendor-level evidence: Nvidia’s guide for about 70% revenue growth in its fiscal 2028 (February 2027 to January 2028) at a much higher price per package (a Vera Rubin rack costs roughly twice a GB300 rack), Morgan Stanley’s 2027 CoWoS demand (2.7M wafers, Nvidia about 45%), TPU forecasts of 5–8.8M in 2027, and the AMD, Broadcom and Anthropic deal books. After 2027 units grow 12% in 2028 and about 8% a year after. A capex- and power-constrained top-down view gives far fewer, about 15–17M packages a year through 2030; it is our low case. The <a href="https://06.ai/packs/ai-compute-balance">compute-balance</a> pack shows why it matters: see “How this squares with the <a href="https://06.ai/packs/ai-compute-balance">compute-balance</a> pack” below.</p>
<h3>Nvidia generations: memory per package, per rack, and over each generation’s life</h3>
<ul><li><strong>Hopper H100 / H200</strong> — Volume years: 2023–25 · HBM per package: 80 / 141 GB · Rack fast memory (HBM + CPU LPDDR): 0.64–1.1 TB per 8-GPU HGX · Lifetime HBM, our estimate: ~0.5 EB (2024–26)</li><li><strong>Blackwell B200 / GB200</strong> — Volume years: 2025 · HBM per package: 186–192 GB · Rack fast memory (HBM + CPU LPDDR): 30.4 TB per NVL72 (13.4 + 17) · Lifetime HBM, our estimate: ~0.6 EB</li><li><strong>Blackwell Ultra B300 / GB300</strong> — Volume years: 2H25–27 · HBM per package: 288 GB · Rack fast memory (HBM + CPU LPDDR): 37 TB per NVL72 (20 + 17) · Lifetime HBM, our estimate: ~1.8 EB</li><li><strong>Rubin VR200</strong> — Volume years: 2H26–28 · HBM per package: 288 GB · Rack fast memory (HBM + CPU LPDDR): 48–75 TB per NVL72 (20.7 + 28–54) · Lifetime HBM, our estimate: ~2.6 EB of HBM plus 3.4–6.8 EB of LPDDR</li><li><strong>Rubin Ultra</strong> — Volume years: 2H27–29 · HBM per package: 1 TB on 4 dies (announced); ~192–256 GB on 2 dies (reported) · Rack fast memory (HBM + CPU LPDDR): 147 TB of HBM per Kyber rack (announced) · Lifetime HBM, our estimate: ~2.9 EB at announced spec vs ~1.1–1.5 EB reported</li><li><strong>Feynman</strong> — Volume years: 2028– · HBM per package: “Custom HBM”, capacity undisclosed</li></ul>
<p>Lifetime totals use our unit mix (Hopper 5.2M, Blackwell 3.3M, Blackwell Ultra 6.4M, Rubin 8.9M and Rubin Ultra 5.7M packages) and are rounded. Rubin Ultra counts as two-die equivalents.</p>
<p>The Rubin row is the one to notice. Its LPDDR5X attach, even at the reported halved level, matches or exceeds its HBM in bytes, which is why the Vera cut matters for commodity DRAM.</p>
<h3>Everyone else’s current parts</h3>
<ul><li><strong>AMD</strong> — Now shipping: MI355X, 288 GB; MI455X, 432 GB HBM4 (since late Q3 2026) · Next: MI500 (2027); memory not yet disclosed · A 288 GB MI450 reportedly built for Meta</li><li><strong>Google</strong> — Now shipping: TPU v7 Ironwood, 192 GB; TPU 8i, 288 GB HBM3E (since Sep 2026) · Next: TPU 8t, 216 GB HBM3E · Anthropic: 1 GW in 2026, 5 GW of 8i in 2027</li><li><strong>AWS</strong> — Now shipping: Trainium 3, 144 GB · Next: Trainium 4, ~288 GB (HBM4-class) · Anthropic up to 5 GW; OpenAI 2 GW from 2027</li><li><strong>Meta</strong> — Now shipping: MTIA 300, 216 GB · Next: MTIA 400/450, 288 GB; MTIA 500, 384–512 GB · The largest among announced 2027 custom ASICs</li><li><strong>Microsoft</strong> — Now shipping: Maia 200, 216 GB · Next: Maia 300 in design</li><li><strong>Huawei</strong> — Now shipping: Ascend 950PR / 950DT, 128 / 144 GB · Next: Ascend 960DT, 288 GB (Q1 2027) · Domestic HBM; see China below</li></ul>
<p>By HBM bits, the 2027 split is about 37% Nvidia, 31% Google, 13% AWS, 9% AMD and 9% other ASICs in our base case, close to Morgan Stanley’s 37/36/12% for Nvidia, Google and AMD.</p>
<h2>After M15X and P4, new memory fabs add wafers from mid-2027; Terafab’s DRAM points to about 2030</h2>
<figure><a href="https://06.ai/packs/ai-memory-balance#fabs"><img src="https://06.ai/newsletter/ai-memory-balance/fig-06.png" alt="After M15X and P4, new memory fabs add wafers from mid-2027; Terafab’s DRAM points to about 2030" width="694" height="462"></a><figcaption>New DRAM and HBM capacity by first output, as reported by October 2026. Bars: ramps; diamonds: first output. HBM stacks follow wafers by about six months.</figcaption></figure>
<h2>Supply: where the HBM bits come from</h2>
<p>HBM output roughly doubles in two years, from about 3.9 EB in 2026 to 6.0 EB in 2027 and 8.0 EB in 2028 in our base case, but only by taking about 30% of the big three’s DRAM wafers. Apart from SK hynix’s M15X and Samsung’s P4 phase 2, both producing since the first half of 2026, no new fab adds volume before mid-2027.</p>
<p><strong>HBM eats wafers more than twice as fast as it makes bits.</strong> HBM takes 18%, 22% (vendor capacity figures imply about 25%) and 30% of the big three’s DRAM wafer starts in 2025, 2026 and 2027, but yields only 8%, 9% and 13% of DRAM bits (TrendForce). Micron puts HBM3E at about 3× the wafer area per bit of DDR5 and says the ratio rises each generation. We estimate 650–850 good GB per HBM wafer start in 2026, about twenty 36 GB stacks. The consequence: non-HBM wafer starts at Samsung, SK hynix and Micron stay roughly flat at 1.26–1.33M a month through 2027. Nearly every new wafer goes to HBM.</p>
<h3>Capacity by vendor, thousand HBM wafer starts a month</h3>
<figure><a href="https://06.ai/packs/ai-memory-balance"><img src="https://06.ai/newsletter/ai-memory-balance/fig-07.png" alt="Samsung — End-2025: ~170 · 2026: ~180 average · 2027: ~250 average · HBM4 status, Oct 2026: Shipping since Feb 2026; yield ~80%; 1c core dies on an in-house foundry base die; SK hynix — End-2025: 160–170 · 2026: 150–200 · 2027: ~230–260 (our estimate) · HBM4 status, Oct 2026: Volume since Q2 2026; anchor Rubin supplier, ~55–70%; Micron — End-2025: 40–55 · 2026: ~100 by year-end · 2027: ~120–150 (our estimate) · HBM4 status, Oct 2026: Ships above 11 Gbps; reportedly marginal on flagship Rubin; custom HBM4E next; Industry, ex-China — End-2025: ~370–390 · 2026: ~450 average · 2027: ~550–680 · HBM4 status, Oct 2026: HBM4E ramps in 2H 2027 at all three; CXMT (China only) — End-2025: ~10 · 2026: ~20–30 · 2027: ~55 · HBM4 status, Oct 2026: HBM3 at ~25% yield; small-scale HBM3E since Sep 2026" width="712" height="438"></a></figure>
<p>Q2 2026 HBM revenue share: SK hynix 50%, Samsung 33%, Micron 18% (Counterpoint).</p>
<p><strong>From wafers to bits.</strong> Our base case runs average HBM wafer input from about 450k a month in 2026 to 590k in 2027, 710k in 2028 and 975k in 2030, with good output per wafer rising from 0.72 TB to 1.05 TB as the industry moves to 1c/1d nodes and 32 Gb dies. That lands close to the published anchors, 1.5 EB in 2024 and about 3.0 EB in 2025 (TrendForce), about 4 EB in 2026, and TrendForce’s +50–60% for 2027: we get 1.5, 2.8 and 3.9 EB and +54%.</p>
<p>Beyond 2027 no vendor publishes HBM capacity. Our 2028–30 path assumes HBM keeps about a third of a big-three DRAM base growing toward about 3M wafer starts a month by 2030 (Citrini puts Samsung and SK hynix alone at 2.3–2.6M). The low and high cases sit about 19% below and 23% above the base in 2028, and 28% below and 33% above in 2030.</p>
<p><strong>Price is what forces the shift.</strong> 2026 was an HBM pricing pause: HBM3E sold near $1.50/Gb (about $12/GB) while conventional contract prices compounded about sixfold from 3Q25 to 4Q26, to roughly $2/Gb. By Q1 2026 a wafer of 64 GB server DIMMs earned more than a wafer of HBM. At a 3× wafer ratio, a $2/Gb commodity bit implies about $6/Gb for an HBM bit to break even per wafer. That is why 2027 HBM4 quotes sit at $4–5/Gb (about $32–40/GB, or $1,150–1,450 per 36 GB stack) and blended 2027 forecasts run from +54% (J.P. Morgan) to +121% (TrendForce). The HBM price is now hostage to the commodity price.</p>
<p><strong>Most of it is already sold.</strong> Micron has $32B of customer commitments and has contracted the “vast majority” of its 2027 HBM. Samsung has signed five-year agreements with its top five data-center customers and plans to cover 60–70% of capacity; SK hynix has long-term deals with about ten customers. A buyer without a contract in 2027 has very little to buy.</p>
<h2>At 128K tokens, 2026’s flagship models need 5–50× less KV cache than Llama 3.1 405B</h2>
<figure><a href="https://06.ai/packs/ai-memory-balance#kv"><img src="https://06.ai/newsletter/ai-memory-balance/fig-08.png" alt="At 128K tokens, 2026’s flagship models need 5–50× less KV cache than Llama 3.1 405B" width="694" height="434"></a><figcaption>GB of KV cache per 128K-token sequence at FP8, the serving default; hybrids include their fixed recurrent state. Right column: how many such sequences fit in the ~200 GB KV pool a 288 GB Rubin GPU has after weights and runtime.</figcaption></figure>
<h2>KV cache: the demand driver and the levers</h2>
<p>KV cache is what now fills HBM, and it is the most movable part of memory demand. In a typical 2026 serving setup KV takes about 60% of a 288 GB GPU and weights about 15% (LMSYS). Attention design changes KV per token by more than 50× and precision by up to 3.5×, and warm context can be parked in LPDDR, CXL memory or flash. Net: KV growth raises total memory demand but pushes the marginal byte off HBM.</p>
<p><strong>Sizing.</strong> KV per token is 2 × layers × KV heads × head size × bytes for standard attention; latent, compressed and linear-attention designs cut it further. At FP8, a 128K-token context ranges from 34 GB on Llama 3.1 405B to 0.7 GB on DeepSeek-V4-Pro, a model four times larger. FP4 KV fits 1.78× more than FP8 but is still experimental, costing 1.6 points on SWE-bench Verified in one test.</p>
<p><strong>Why it keeps growing.</strong> Google processed more than 3.2 quadrillion tokens a month in May 2026, against 480 trillion a year earlier. Agents make prompts long and repetitive: 94–98% of coding-agent prompt tokens are reused prefixes, and inputs outnumber outputs 131 to 1. DeepSeek-V4-Pro agent inputs run at a median of 88K tokens and 675K at the 99th percentile; Nvidia sizes Rubin for 128K–400K cached tokens per user.</p>
<p><strong>Where it can live.</strong> The tier ladder runs from HBM (the active batch) to CPU LPDDR, optical and CXL pools, flash and network storage (table below). Tiering is now standard in vLLM, SGLang, TensorRT-LLM and Nvidia’s Dynamo. On coding-agent traces, a shared KV store lifted throughput 3.8× and cut time-to-first-token 46× by raising cache hits from 2% to 92%.</p>
<ul><li><strong>CPU LPDDR.</strong> Vera carries 768 GB–1.5 TB per two GPUs over a 1.8 TB/s link. This is Nvidia’s main answer, and why the reported LPDDR cut matters.</li><li><strong>CXL pools.</strong> Samsung’s CMM-D pool ran KV offload at about 92% of local-DRAM throughput at TP=8; Alibaba’s CXL pool cut time-to-first-token about 90% against an RDMA store.</li><li><strong>Flash.</strong> Nvidia’s BlueField-4 context-memory tier is due in 2H 2026 with “petabytes” per pod and a claimed 5× tokens/s, still unmeasured in public. Citi estimates about 16 TB of NAND per Rubin GPU, about 115 EB in 2027 or 9% of global NAND.</li><li><strong>High Bandwidth Flash.</strong> A 512 GB-per-stack design, samples in 2027 and volume in 2028 per analysts, for read-mostly weights, not active KV.</li></ul>
<p><strong>Shrinking it.</strong> DeepSeek-V4 stores 8.7× less KV than V3.2 at 1M tokens, Kimi K3 and Qwen3.8 cut per-token KV 2–3× against their predecessors, and Nvidia credits part of GB300’s 1.6× software gain in MLPerf to lower KV precision. Hardware moves the same way: Groq 3 LPX racks run MoE decode from SRAM (500 MB per chip, 128 GB per rack, no HBM) while Rubin GPUs keep attention and KV. The GDDR7-based Rubin CPX no longer appears on Nvidia’s Vera Rubin page.</p>
<p><strong>What it does to the balance.</strong> HBM capacity sets concurrency: GB300’s 1.5× HBM over GB200 gave 1.6× the decode batch, and a third less HBM removes 36–41% of resident sequences because weights are a fixed charge. Whether that costs tokens depends on reads. With dense attention, per-user decode speeds up and tokens per GPU fall only 5–11%; with sparse attention, or agent sessions parked between tool calls, the loss approaches the full 40%. Hence the design now shipping: size HBM for the active batch and stage warm context one tier down during tool-call pauses, which last about 5 seconds at the median.</p>
<p><strong>Which tier depends on attention design.</strong> At 57 agents per GPU (Artificial Analysis’s measured rate for DeepSeek-V4-Pro on GB300) and 90th-percentile agentic context, an MLA model spills about 25 TB of warm KV per Rubin rack, which fits Vera’s LPDDR. A dense 405B-class model spills 80–280 TB, which needs flash, while DeepSeek-V4-class models fit in HBM (our estimates). So KV growth lands mostly in LPDDR, tightening commodity DRAM, and in NAND: enterprise SSD is now about 37% of NAND demand (CFM), and Citi sees a 6.1% NAND shortfall in 2027.</p>
<p>Our trimmed-spec case assumes that tiering works. If agent hosting is priced per session and insists on HBM-resident context, buyers will pay for taller stacks again, and demand moves toward the announced-roadmap case.</p>
<h2>Marvell’s Photonic Fabric and pooled memory</h2>
<p>Optical pooling slows the growth of HBM per GPU from 2028; it does not cut it before then. On an accelerator designed around a pool, it could replace 10–20% of HBM capacity in 2028 and 25–35% by 2030. Across the whole fleet that is about 2% in 2028 and 3–8% in 2030 (our estimates).</p>
<p><strong>What Marvell bought.</strong> Marvell closed its $3.25B purchase of Celestial AI on February 2, with up to $2.25B more tied to revenue. Photonic Fabric links chips over light at about 2.4–3.1 pJ/bit and adds roughly 200–300 ns of latency; on-package HBM is under 100 ns. The memory module fronts 2 TB of DDR5 with two HBM3E stacks (72 GB) acting as a cache. A 2U appliance of 8–16 modules holds about 33 TB, shared by up to 16 servers as far as 50 m away.</p>
<p><strong>Where it stands.</strong> The memory module taped out in 2026, is in bring-up, and samples to “multiple customers” in 2027. Every disclosed win so far is optical interconnect, not pooled memory: a Tier 1 hyperscaler’s XPU scale-up network and co-packaged optics added to the AWS agreement. Marvell’s revenue targets, a $500M run-rate by November 2027–January 2028 and $1B by November 2028–January 2029, are interconnect too. Amazon holds a warrant for 1,045,171 Marvell shares at $87.00 that vests with its Photonic Fabric purchases through 2030, the clearest signal of who adopts first.</p>
<p><strong>It replaces capacity, not bandwidth.</strong> The appliance switch moves about 115 Tbps (about 14 TB/s), roughly one Rubin GPU’s HBM4 bandwidth. Shared across 64–128 accelerators, that is 0.1–0.2 TB/s each, about 1% of Rubin’s 19–22 TB/s, while offering 250–500 GB of capacity each. So the pool can hold only bytes touched rarely: warm or parked KV cache, cold MoE experts and adapters, optimizer state, checkpoints. KV for sequences in the active decode batch, and hot weights, are read every token step and stay in HBM.</p>
<p><strong>The lever it unlocks is stack height.</strong> HBM capacity comes bundled with bandwidth: drop a stack and you lose both. Shortening stacks keeps bandwidth (12-Hi to 8-Hi cuts GB by a third, 16-Hi to 12-Hi by a quarter), and a pooled tier is what lets a designer take that cut. That bounds displacement near 25–35% per accelerator.</p>
<ul><li><strong>On-package HBM</strong> — Example: Rubin, 288 GB of HBM4 · Added latency: under 100 ns · Bandwidth per accelerator: 19.2–22 TB/s · Status, Oct 2026: shipping 2026</li><li><strong>CPU-attached LPDDR</strong> — Example: Vera, 768 GB–1.5 TB of LPDDR5X per 2 GPUs · Added latency: local DRAM · Bandwidth per accelerator: ~0.6 TB/s (1.2 TB/s per Vera) · Status, Oct 2026: ships with Vera Rubin</li><li><strong>Optical memory pool</strong> — Example: Marvell Photonic Fabric appliance · Added latency: ~200–300 ns · Bandwidth per accelerator: 0.1–0.2 TB/s shared; 1.8 TB/s per dedicated chiplet · Status, Oct 2026: samples 2027</li><li><strong>CXL switch pool</strong> — Example: Marvell Structera S 30260 (48 TB), Astera Leo, Samsung CMM-D · Added latency: ~460–750 ns · Bandwidth per accelerator: ~26–36 GB/s per device path · Status, Oct 2026: expanders shipping; switches sampling</li><li><strong>RDMA / flash context tier</strong> — Example: Nvidia ICMS (BlueField-4), Enfabrica EMFASYS · Added latency: microseconds · Bandwidth per accelerator: ~0.1 TB/s (800 Gb/s) · Status, Oct 2026: ICMS due 2H 2026</li></ul>
<p><strong>Who adopts, and why the effect is muted.</strong> Nvidia, the largest HBM buyer, is building its own non-optical tiers (Vera LPDDR5X, ICMS, the Enfabrica team it paid over $900M for) while its announced Rubin Ultra still lists 1 TB of HBM per package (reportedly cut to 192–256 GB). That leaves custom accelerators (AWS, Google, Meta, Microsoft) and UALink systems as the early market; optical NVLink for Feynman in 2028 is reported but unconfirmed. Vendors sell pooling as a 2–5× throughput gain, mostly from simulation, not as HBM savings, and cheaper tokens can raise total HBM demand, so the industry-level effect is smaller than the per-accelerator one and could be zero. The tell that this is changing would be a named memory-appliance customer or purchase order; none has been disclosed.</p>
<h2>China: CXMT, YMTC, Huawei and export controls</h2>
<p>China narrows the global commodity DRAM shortfall but barely touches the HBM balance. Hanwha estimates CXMT shrinks the 2026 global DRAM deficit from 15.5% to 9.4%, and 2027’s from 12.0% to 5.6%. Almost none of that reaches HBM, and most of it stays inside China.</p>
<p><strong>CXMT is big in wafers, smaller in bits, and not cheap.</strong> It runs about 300–320k wafer starts a month across two Hefei fabs and one in Beijing, heading to about 420k in 2027 and 500–600k in 2028. That is about 17% of the top five makers’ wafer starts but only 6.7–9% of bits, because its nodes lag; it took 9.5% of DRAM revenue in Q2 2026, fourth globally. It prices at parity with Samsung, earned a net margin of about 52% in the first half of 2026 (RMB77.6bn on RMB150.3bn of revenue), and is booked through 2027. Its July STAR Market IPO raised RMB57.9bn ($8.6B). So it relieves volume, not price.</p>
<p><strong>It mostly displaces imports.</strong> Goldman puts CXMT at 41% of China’s DRAM demand in 2026 and about 50% by 2028. The bits that frees for Korean and US vendors elsewhere are small next to industry bit growth of 22–24% a year, and Korean HBM conversion is driven by HBM demand regardless.</p>
<p><strong>China’s HBM is a closed, starved loop.</strong> CXMT makes HBM3 8-high at about 25% overall yield (SemiAnalysis, June) and began small-scale HBM3E in September; it has no HBM4. Huawei’s pre-control stockpile of about 13M stacks (11.4M from Samsung) runs out in 2026. Epoch AI’s model has about 1.5M Ascend chips in 2026, but only about 240k with domestic HBM, about 30–35M GB or 0.03 EB, while about 750k use non-HBM memory. Ascend prices rose 20–50% in two months; the 950DT now lists above RMB250k. DeepSeek’s order for 160k 950DTs alone would take about 70% of 2026’s domestic HBM.</p>
<p><strong>Export controls cap China’s demand on the Korean and US balance.</strong> The December 2024 rule bars HBM above 2 GB/s/mm² of bandwidth density; equipment controls cover every CXMT node. Nvidia’s H200 is licensed, with 25% to the US Treasury and at most 75k units per buyer, but Beijing’s own gatekeeping held deliveries to about 20k by August against more than 2M ordered. Even the full licensed ceiling, about 750k H200s at 141 GB, is about 0.1 EB. China’s draw on Korean and US HBM is policy-gated, small, and could swing either way.</p>
<p><strong>The Korean fabs in China are frozen at their nodes.</strong> Samsung’s and SK hynix’s China fabs lost fast-track (validated end-user) status on December 31, 2025 and run on annual licences that bar upgrades. Wuxi is about half of SK hynix’s DRAM capacity, so its HBM growth has to come from Korea and the US while Wuxi drifts toward commodity output that competes with CXMT.</p>
<p><strong>YMTC matters more for flash than DRAM.</strong> It is now third in NAND by shipments, with 14–16.5%. Omdia has it adding about 486k NAND wafers in 2027 against about 75k each for Samsung and SK hynix, useful for the SSD tier that holds offloaded KV cache. Its DRAM entry, about 50k wafers, is immaterial before 2028.</p>
<p><strong>Oversupply risk: low through 2027, rising after.</strong> China’s DRAM capacity could reach about 1.4M wafer starts a month by 2030, and domestic DUV scanners arrive in 2027–28 (Citrini). An AI air pocket would hit DDR5 and LPDDR5X first, CXMT’s core products. Citrini still models a global DRAM deficit of about 25% in 2030.</p>
<p><strong>Dates to watch.</strong> The US affiliates rule returns on November 10 unless extended; the HBM temporary licence expires on December 31; the Korean fabs in China need 2027 licences; and CXMT’s Entity List designation is approved but unpublished.</p>
<h2>Terafab and other new capacity</h2>
<p>Terafab adds essentially no memory before 2030; through this forecast it matters as a buyer, not a supplier. Our base case puts its DRAM output at about 10,000 wafer starts a month in 2030, roughly 0.1% of global DRAM bits, and no HBM.</p>
<p><strong>What is actually committed.</strong> Musk floated a “Tesla TeraFab” in November 2025, described it in January 2026 as “logic, memory and packaging”, and launched it on March 21 with talk of 100,000 wafer starts a month initially and 1,000,000 eventually. Intel joined in April, and the first full-scale phase is to run Intel’s 14A logic process. Two things are firm. A research fab at Giga Texas of “a few thousand” wafers a month, about $3B, starts operating in the first half of 2027; its first test chip is SRAM, due in May–June 2027. And a $16.8B Phase 1 at Grimes County runs from December 1, 2026 to the end of 2028, within up to $119B of phases through 2036. On October 20 Musk targeted volume logic production for Q3–Q4 2028, “with big error bars”. No DRAM capacity, node, process licensor or memory partner has been disclosed; the memory evidence is job postings since August for sub-20nm DRAM process integration and HBM/LPDDR interface design.</p>
<p><strong>Why memory comes late.</strong> Incumbents with their own process take about 3.7 years from groundbreaking to first DRAM wafers: Micron’s Idaho ID1 broke ground in October 2023 and outputs first wafers in mid-2027, with meaningful output in 2028. CXMT, the last new entrant, took about 3.5 years to first low-volume output and 8–10 years to a mid-single-digit share. Phase 1’s $16.8B buys roughly one memory fab ($15–20B) or half a leading-edge logic fab, and it is going to logic. Without a licensed DRAM process, first product wafers point to about 2030; HBM, which needs best-in-class DRAM yields plus stacking, comes after that.</p>
<p><strong>Our scenarios.</strong> Base (35%): about 2,000 DRAM wafer starts a month in 2029 and about 10,000 in 2030, about 0.1% of 2030 DRAM bits. High (10%), with a licensor signed and tools ordered by mid-2027: about 20,000 and 50,000, about 1% of bits and under 1% of HBM. Low (55%): engineering lots only. For scale, CXMT alone adds about 100,000 wafer starts a month in 2027, twice the high case for 2030.</p>
<p><strong>Net effect: neutral to tighter.</strong> Whatever Terafab makes is captive, so it displaces Tesla and SpaceX purchases rather than adding merchant supply. Meanwhile Musk’s companies plan about 13 GW of compute by end-2027, including a 440,000-GPU Vera Rubin cluster by mid-2027, roughly 0.13 EB of HBM4 on its own. Musk himself calls memory “the biggest constraint”. The one real effect before 2030 is on bargaining: a credible in-house threat can weigh on 2028–30 contract pricing before any bits exist.</p>
<p><strong>Other new capacity is incumbents, financed by customers.</strong> Outside China there is no new DRAM producer before 2030. SoftBank and Intel’s Saimemory plans Z-Angle Memory prototypes by March 2028 and commercialization in its fiscal 2029, with no capacity disclosed. Buyer money instead funds incumbents’ cleanrooms. Micron has 26 strategic customer agreements covering more than 35% of revenue through 2030, and $32B of customer commitments, mostly prepayments. SK hynix is moving to 3–5-year contracts with 10–30% upfront.</p>
<h2>Unit volumes, supply and memory per package decide the cycle</h2>
<p>Supply vs demand (supply ÷ demand − 1) in each scenario. “Low” units are the capex- and power-limited view, about 15–17M packages a year; “high” takes the vendors’ deal books at face value, about 25M in 2027 and 39M by 2030.</p>
<ul><li><strong>Base — Units: base · Specs: trimmed · HBM supply: base · 2027: −2% · 2028: +3% · 2030: −2%</strong></li><li><strong>Roadmaps hold</strong> — Units: base · Specs: announced · HBM supply: base · 2027: −18% · 2028: −29% · 2030: −38%</li><li><strong>Deal books delivered</strong> — Units: high · Specs: trimmed · HBM supply: base · 2027: −22% · 2028: −22% · 2030: −33%</li><li><strong>Supply disappoints (HBM4 yields, cleanroom slips)</strong> — Units: base · Specs: trimmed · HBM supply: low · 2027: −16% · 2028: −17% · 2030: −29%</li><li><strong>Supply surprises (Samsung, new fabs on time)</strong> — Units: base · Specs: trimmed · HBM supply: high · 2027: +12% · 2028: +27% · 2030: +30%</li><li><strong>Power- and capex-limited build</strong> — Units: low · Specs: trimmed · HBM supply: base · 2027: +37% · 2028: +52% · 2030: +52%</li><li><strong>Power-limited build, roadmaps hold</strong> — Units: low · Specs: announced · HBM supply: base · 2027: +15% · 2028: +4% · 2030: −6%</li></ul>
<h3>One input at a time, 2028</h3>
<p>2028 supply vs demand at trimmed specs when one input changes and the rest stay at base. Labels: the input’s setting and the result. Whole unit cases swing it further, −22% to +52%.</p>
<figure><a href="https://06.ai/packs/ai-memory-balance#scenarios"><img src="https://06.ai/newsletter/ai-memory-balance/fig-09.png" alt="Unit volumes, supply and memory per package decide the cycle" width="694" height="330"></a></figure>
<h2>Reading the scenarios</h2>
<p><strong>The base case sits on a knife-edge.</strong> Moving to the low or high supply case (about −19% or +23%), or ±15% on memory per package, swings 2028 anywhere from 10–17% short to 21–27% long. Three variables decide the cycle: total unit volumes, which power limits; good bits per HBM wafer (HBM4/4E yields, Samsung’s ramp); and whether 8-Hi stacks stick.</p>
<p><strong>Reading the camps.</strong> The “oversupply by 2028” view (Bloomberg Intelligence, DS Investment, Morgan Stanley’s Micron model with fiscal 2029 EPS down about 40%) needs either the power-limited build or the supply upside. The “short through 2030” view (SK hynix, Micron, Citrini’s 25% DRAM deficit in 2030) needs the announced roadmaps or the deal books. Our base is neither: within ±3% every year, and only because memory per package stops compounding at its old rate.</p>
<p><strong>How this squares with </strong><strong>the </strong><a href="https://06.ai/packs/ai-compute-balance"><strong>compute-balance</strong></a><strong> pack</strong><strong>.</strong> <a href="https://06.ai/packs/ai-compute-balance">The </a><a href="https://06.ai/packs/ai-compute-balance">compute-balance</a><a href="https://06.ai/packs/ai-compute-balance"> pack</a> found that from 2027 power and money, not chips, limit what gets switched on: about 38 GW of all-in capacity in 2027, rising to 51 GW in 2030, while the chipmakers could deliver chips for 47–80 GW. On the <a href="https://06.ai/packs/ai-compute-balance">compute-balance</a> pack’s basis (a data center draws about twice its chips’ rated power) and our estimate of 1.4–1.7 kW of chip power per package, our base units need about 55 GW in 2027 and 68–93 GW a year in 2028–30, 7–17% more than the <a href="https://06.ai/packs/ai-compute-balance">compute-balance</a> pack’s chip path. Its switched-on path sits below even our low unit case (41–61 GW). In that low case HBM runs 37–52% long at trimmed specs in 2027–30, and the industry could afford the announced roadmaps instead: supply would sit within −6% to +15% of demand. If chip purchases fall to the <a href="https://06.ai/packs/ai-compute-balance">compute-balance</a> pack’s switched-on path, the surplus is larger still.</p>
<p><strong>Asymmetry.</strong> Our base needs more new AI capacity a year after 2027 than the <a href="https://06.ai/packs/ai-compute-balance">compute-balance</a> pack expects to be switched on, so the unit risk leans down. The downside case for memory makers, a power-limited build with trimmed specs (+52% in 2028), arrives just as Yongin, P5, ID2 and Hiroshima add wafers. The upside needs the announced roadmaps or deal-book volumes, either of which leaves 2028 22–29% short.</p>
<h2>What to watch</h2>
<p>The next six months settle more of this than the last six: 2027 HBM4 contract prices, whether 8-Hi sticks on Rubin Ultra and the 2027 ASICs, and how many gigawatts actually get switched on. 2027 memory earnings are largely contracted; the open question is 2028.</p>
<ul><li><strong>Oct 29</strong> — SK hynix and Samsung Q3 results · 2027 capex, HBM wafer adds, HBM4 share at Nvidia, long-term-agreement coverage</li><li><strong>November</strong> — Nvidia fiscal Q3 2027 results · Rubin ramp, gross margin path, supply commitments; any confirmation of Rubin Ultra’s 192/256 GB configuration or Vera’s 768 GB SKU</li><li><strong>Nov 10 and Dec 31</strong> — US affiliates rule returns; HBM temporary licence expires · China’s access to HBM and tools; 2027 licences for Korean fabs in China</li><li><strong>4Q 2026</strong> — 2027 HBM4 contract settlements · HBM4 near $4–5/Gb validates the top of the +54–121% range; HBM4 well below $4/Gb, or blended pricing nearer +54%, means buyers had more leverage</li><li><strong>1Q 2027</strong> — First 2027 conventional DRAM contract print · A rise under 5% a quarter favours the balance-by-late-2027 camp</li><li><strong>March 2027</strong> — Nvidia GTC · Feynman’s memory spec and Rubin Ultra’s final configuration</li><li><strong>Mid-2027</strong> — M15X at full rate; Yongin fab 1, Micron ID1 and Tongluo first output · Whether 2028 supply arrives on time</li><li><strong>2027</strong> — Photonic Fabric memory samples; HBF samples; first named pooling customer · Whether capacity moves off HBM after 2028</li><li><strong>Rolling</strong> — Gigawatts switched on against SemiAnalysis’s ~51 GW for 2027; hyperscalers’ 2027 capex guides · Which unit path, base, power-limited or deal-book, is real</li></ul>
<p><strong>HBM makers (SK hynix, Samsung, Micron).</strong> Pricing power peaks in 2027, when HBM must reprice toward commodity-DRAM parity at about 3× the wafer cost per bit. Micron’s $32B of customer commitments and Samsung’s five-year agreements, planned to cover 60–70% of capacity, make 2027 visible; the equities already discount a 2028 turn (SK hynix about 57% off its June peak, at about 2.7× 2027 earnings per BofA). Our base case says the 2028 risk comes from the demand side, trimmed specs plus a power-limited build, more than from new fabs.</p>
<p><strong>Nvidia.</strong> Memory is the main gross-margin headwind (75% to 71–72%) and the reason for server price increases of more than 15%. Trimming memory per GPU protects unit volume; per-die HBM falls on Rubin Ultra.</p>
<p><strong>Custom-ASIC chains (Broadcom, MediaTek, Alchip, Marvell).</strong> They take most of 2027’s unit growth and are still mostly on HBM3E, which keeps HBM3E tight through 2027.</p>
<p><strong>Capacity tiers.</strong> LPDDR/SOCAMM suppliers, enterprise SSD and NAND, and CXL silicon (Astera Labs, Marvell, Montage) are where KV-cache growth shows up first. Optical pooling (Marvell) and High Bandwidth Flash (SanDisk, SK hynix) come in 2028 and later.</p>
<p><strong>China and Terafab.</strong> CXMT relieves volume, not price, and stays mostly domestic; China’s demand for Korean and US HBM is set by policy, not need. Terafab is a buyer through 2030, with bargaining value before it has any bits.</p>
<h2>What this is anchored on</h2>
<p>Supply, demand, memory per package, the balance and the scenarios are our estimates, built from the sources below as of October 25, 2026.</p>
<ul><li><strong>Nvidia</strong> — Rubin: 288 GB of HBM4 per GPU, listed bandwidth 19.2 TB/s (22 TB/s at CES); Vera “up to 1.5 TB” of LPDDR5X; sized for 128K–400K cached tokens per user · <a href="https://www.nvidia.com/en-us/data-center/vera-rubin-nvl72/">Nvidia Vera Rubin NVL72 page</a></li><li><strong>Nvidia</strong> — GB300 NVL72: 20 TB of HBM (about 278 GB per GPU) and 17 TB of LPDDR5X per rack · <a href="https://www.nvidia.com/en-us/data-center/gb300-nvl72/">Nvidia GB300 NVL72 page</a></li><li><strong>Nvidia</strong> — Rubin Ultra announced with 1 TB of HBM4E on four compute dies; Kyber NVL144 · <a href="https://technews.tw/2026/03/18/nvidia-gtc-2026-chips/">TechNews, GTC 2026, Mar 18</a></li><li><strong>Nvidia</strong> — Q2 gross margin 75.0%; Q3 guide 74% ±50 bp; inventory $31.6B · <a href="https://www.sec.gov/Archives/edgar/data/0001045810/000104581026000073/q2fy27pr.htm">Nvidia Q2 FY27 8-K</a></li><li><strong>Nvidia</strong> — CFO Colette Kress on “extreme pricing conditions in memory”: gross margin troughs at 71–72% in the quarter to January 2027; supply commitments from $119B to $279B · <a href="https://www.nasdaq.com/articles/memory-costs-rise-can-nvidia-protect-its-70-gross-margin">Nasdaq / Zacks, Sep 3</a></li><li><strong>Nvidia</strong> — Supply and capacity purchase commitments, mostly memory, from $119B to $279B in one quarter; Morgan Stanley’s 2027 HBM split of 37/36/12% for Nvidia, Google and AMD · <a href="https://xenospectrum.com/en/nvidia-hbm-demand-supply-commitments/">XenoSpectrum</a></li><li><strong>Rubin Ultra</strong> — Moves to 8-Hi stacks: not enough HBM wafers for 12-Hi, HBM cost, and a 576-GPU domain of 192 GB parts (~110 TB) · <a href="https://newsletter.semianalysis.com/p/long-live-the-short-king-why-4-hi">SemiAnalysis, Sep 13</a></li><li><strong>Rubin Ultra</strong> — Configurations tested with as little as 192 GB, and a step back to HBM4 · <a href="https://www.tomshardware.com/pc-components/gpus/nvidia-reportedly-testing-lower-memory-configs-of-rubin-ultra-as-memory-shortage-bites-back-designs-tested-include-as-little-as-192-gb-and-step-back-to-hbm4">The Information, via Tom’s Hardware</a></li><li><strong>Rubin</strong> — Most Rubin stays on 12-Hi HBM4 · <a href="https://www.sammyfans.com/2026/10/08/nvidia-rubin-ai-chips-to-retain-12-layer-hbm4-from-samsung-and-sk-hynix/">Counterpoint, via Sammy Fans, Oct 8</a></li><li><strong>Vera</strong> — LPDDR5X per Vera CPU cut from 1.5 TB to 768 GB in 2027 supply plans, “lowering memory per module while increasing total module shipments” · <a href="https://www.trendforce.com/presscenter/news/20260610-13090.html">TrendForce, Jun 10</a></li><li><strong>Vera Rubin cost</strong> — Without the Vera cut, memory would be ~29% of a ~$2.1M VR200 system’s bill of materials, against Nvidia’s preferred ~20% · <a href="https://www.trendforce.com/news/?p=61614">GF Securities, via TrendForce, Jul 28</a></li><li><strong>Vera Rubin cost</strong> — Memory up from ~5–10% to 25–30% of a rack’s bill of materials; server prices up more than 15% · <a href="https://www.vktr.com/ai-market/nvidia-tells-microsoft-google-oracle-that-ai-server-prices-are-going-up-more-than-15/">Morgan Stanley and Bloomberg, via VKTR</a></li><li><strong>GeForce</strong> — RTX 5070 Ti and 5060 Ti 16 GB supply cut 30–40% in 1H26 against 1H25 (Board Channels, Benchlife) · <a href="https://tweaktown.com/news/109394/nvidia-to-cut-supply-of-best-value-geforce-rtx-50-series-gpus-in-half-over-memory-shortages/index.html">TweakTown, Dec 17, 2025</a></li><li><strong>GeForce</strong> — Nvidia: “memory supply is constrained”; all GeForce SKUs still shipping · <a href="https://www.tomsguide.com/computing/gpus/yes-the-entire-series-effectively-nvidia-rtx-50-series-production-reportedly-on-hold-and-its-all-because-of-ai-demand">Tom’s Guide, Jan 22</a></li><li><strong>HBM per device</strong> — GPU and ASIC makers “considering reducing HBM capacity per device”, evaluating 8-Hi in 2027; 8-Hi at a 10–20% per-Gb premium; blended 2027 HBM price +121% · <a href="https://www.trendforce.com/presscenter/news/20260929-13255.html">TrendForce, Sep 29</a></li><li><strong>HBM price</strong> — HBM4 12-Hi about $2/Gb in 2026 and $4–5/Gb in 2027; 8-Hi at a 20–30% premium · <a href="https://stockhub.kr/en/insights/citi-samsung-005930-q3-2026-hbm4-20261008">Citi, via StockHub, Oct 8</a></li><li><strong>HBM price</strong> — HBM3E near $1.50/Gb under 2026 contracts · <a href="https://www.koreaherald.com/article/10895833">Korea Herald</a></li><li><strong>HBM balance</strong> — HBM sufficiency −16% across 2026–28; HBM price +54% in 2027 · <a href="https://www.webull.com/news/15626418510496768">J.P. Morgan, via Zhitong / Webull, Sep</a></li><li><strong>HBM balance</strong> — HBM sufficiency −5.4%, −6.0% and −4.3% in 2026–28; DRAM −5.0%, −5.9%, −3.9% · <a href="https://www.itiger.com/news/1117469753">Goldman Sachs, via Tiger, Jun</a></li><li><strong>HBM balance</strong> — SemiAnalysis’s HBM balance; Samsung P5 HBM output in 2028 · <a href="https://www.useluminix.com/reports/industry-analysis/dram-cycle-position-analysis-peak-timing-indicators">Luminix</a></li><li><strong>DRAM balance</strong> — DRAM sufficiency −5.0%, −1.9% and +0.5% in 2026–28 · <a href="https://file.alphasquare.co.kr/media/pdfs/market-report/%EB%B0%98%EB%8F%84%EC%B2%B4%EA%B3%B5%EA%B8%8920260720DS%ED%88%AC%EC%9E%90%EC%A6%9D%EA%B6%8C.pdf">DS Investment &amp; Securities, Jul 20 (PDF)</a></li><li><strong>HBM supply</strong> — HBM 18%, 22% and 30% of the big three’s DRAM wafer starts in 2025–27, and 8%, 9% and 13% of bits · <a href="https://iconnect007.com/article/150239/tight-dram-supply-boosts-pricing-power-hbm-contract-prices-set-to-surge-in-2027/150236/pcb">TrendForce, via I-Connect007</a></li><li><strong>HBM supply</strong> — HBM3E takes about 3× DDR5’s wafer area per bit, and the ratio rises each generation · <a href="https://xenospectrum.com/en/micron-hbm-ddr-silicon-gap/">Micron, via XenoSpectrum</a></li><li><strong>HBM supply</strong> — Samsung HBM wafer input ~180k a month on average in 2026 and ~250k in 2027 · <a href="https://www.trendforce.com/news/?p=64643">TrendForce, Sep 29</a></li><li><strong>HBM supply</strong> — Micron adds 60k HBM wafers a month, reaching ~100k by year-end · <a href="https://www.trendforce.com/news/2026/09/04/news-micron-reportedly-to-add-60k-hbm-wafersm-reaching-100k-by-year-end-to-narrow-gap-with-samsung-sk-hynix/">TrendForce, Sep 4</a></li><li><strong>HBM supply</strong> — DRAM and TSV (HBM) wafer capacity by vendor, 2025–26 · <a href="https://files.futurememorystorage.com/proceedings/2025/20250805_BMKT-101-1_Avril-Wu.pdf">TrendForce, FMS 2025</a></li><li><strong>HBM supply</strong> — HBM bit shipments: 1.5 EB in 2024, about 3.0 EB in 2025 · <a href="https://files.futurememorystorage.com/proceedings/2025/20250805_BMKT-102-1_Ellie-Wang.pdf">TrendForce, FMS 2025</a></li><li><strong>HBM supply</strong> — Q2 2026 HBM revenue share: SK hynix 50%, Samsung 33%, Micron 18% · <a href="https://www.mt.co.kr/en/industry/2026/10/19/2026101913371799572">Counterpoint, via MT, Oct 19</a></li><li><strong>DRAM capacity</strong> — 2030 capacity: Samsung 1.14–1.45M and SK hynix 1.18M wafer starts a month; China ~1.4M; domestic DUV in 2027–28 · <a href="https://www.tomshardware.com/pc-components/dram/cxmt-close-to-matching-microns-memory-capacity-in-2026-research-claims-would-put-china-on-track-to-become-worlds-second-largest-dram-producer">Citrini, via Tom’s Hardware, Jul 15</a></li><li><strong>DRAM price</strong> — Conventional contract prices up about sixfold from 3Q25 to 4Q26; 4Q26 +10–15% · <a href="https://www.trendforce.com/presscenter/news/20260930-13258.html">TrendForce, Sep 30</a></li><li><strong>Contracts</strong> — Micron: 26 strategic customer agreements covering more than 35% of revenue through 2030, $32B of customer commitments; ID1 and Tongluo output mid-2027, ID2 and Japan late 2028, New York 2030 · <a href="https://www.benzinga.com/node/62095451">Micron Q4 FY26 call, via Benzinga, Sep 30</a></li><li><strong>Contracts</strong> — SK hynix moves to 3–5-year agreements with about 10–30% upfront; Samsung requires agreements of at least three years · <a href="https://www.trendforce.com/news/2026/04/09/news-from-annual-deals-to-3-5-year-ltas-samsung-and-sk-hynix-reportedly-reset-big-tech-memory-contracts/">TrendForce, Apr 9</a></li><li><strong>New fabs</strong> — SK hynix: M15X mass production accelerated; Yongin phase-1 cleanroom opens early 2027 · <a href="https://en.edaily.co.kr/news/eda202607295226/">Edaily, Jul 29</a></li><li><strong>New fabs</strong> — SK hynix Yongin fab 2: groundbreaking Jul 2027, first cleanroom Jun 2029 · <a href="https://www.fool.com/investing/2026/08/15/sk-hynix-approved-38-billion-of-new-memory-fabs-th/">The Motley Fool, Aug 15</a></li><li><strong>New fabs</strong> — SK hynix’s Yongin fab 1 at ~150k wafer starts a month by end-2028 · <a href="https://www.csis.org/analysis/beyond-memory-cycle-ai-hbm-and-new-semiconductor-shortage">CSIS</a></li><li><strong>New fabs</strong> — Samsung P4 adds about 60k wafer starts a month through Q2 2026; P5 operating in 2028 · <a href="https://www.koreajoongangdaily.com/business/samsung-sk-hynix-dramatically-scaling-up-hbm-production-to-meet-rising-demand-recent-nvidia-deals/12391367">KB Securities, via JoongAng Daily, Jan 4</a></li><li><strong>Units</strong> — Anthropic: 1 GW of Ironwood TPUs in 2026 and 5 GW of TPU 8i in 2027; TPU 8i production shipments started · <a href="https://www.webull.com/news/15513120997254144">Broadcom fiscal Q3 call, via Webull, Sep 3</a></li><li><strong>Units</strong> — OpenAI: 2 GW of Trainium from 2027 · <a href="https://www.247wallst.com/investing/2026/07/21/the-3-ways-amazons-trainium-chips-erode-nvidias-dominance-into-early-2027/">247 Wall St., Jul 21</a></li><li><strong>Units</strong> — 2027 CoWoS demand of about 2.7M wafers, Nvidia ~45% · <a href="https://www.kucoin.com/news/flash/morgan-stanley-report-cowos-demand-to-double-by-2027-amd-and-asics-drive-ai-chip-growth">Morgan Stanley, via KuCoin</a></li><li><strong>Power</strong> — New data-center IT capacity outside China ~25 GW in 2026 and ~51 GW in 2027 · <a href="https://superpowerdaily.com/posts/semianalysis-forecasts-51-gw-of-new-data-center-capacity-outside-china-in-2027">SemiAnalysis, via Superpower Daily</a></li><li><strong>Power</strong> — From 2027 power and money cap what is switched on: about 38 GW all-in in 2027 rising to 51 GW in 2030 · <a href="https://06.ai/packs/ai-compute-balance">06.ai </a><a href="https://06.ai/packs/ai-compute-balance">compute-balance</a><a href="https://06.ai/packs/ai-compute-balance"> pack</a></li><li><strong>Other chips</strong> — MTIA 300, 216 GB; MTIA 400/450, 288 GB; MTIA 500, 384–512 GB · <a href="https://www.tomshardware.com/tech-industry/semiconductors/meta-reveals-four-new-mtia-chips-built-for-ai-inference">Tom’s Hardware</a></li><li><strong>Other chips</strong> — TPU 8t, 216 GB of HBM3E; TPU 8i, 288 GB · <a href="https://www.storagereview.com/news/google-announces-tpu-8t-sunfish-and-tpu-8i-zebrafish">StorageReview</a></li><li><strong>Other chips</strong> — MI455X with 432 GB of HBM4 in the Helios rack · <a href="https://rocm.blogs.amd.com/ecosystems-and-partners/cdna5-helios/README.html">AMD ROCm blog</a></li><li><strong>Other chips</strong> — Maia 200, 216 GB · <a href="https://hothardware.com/news/microsoft-unveils-maia-200-ai-chips">HotHardware</a></li><li><strong>Other chips</strong> — Trainium 3, 144 GB of HBM3E · <a href="https://www.electronicspecifier.com/?p=180729">Electronic Specifier</a></li><li><strong>Other chips</strong> — Trainium 4 doubles Trainium 3’s memory and quadruples its bandwidth · <a href="https://www.nextplatform.com/2025/12/03/with-trainium4-aws-will-crank-up-everything-but-the-clocks/">The Next Platform</a></li><li><strong>Workloads</strong> — MLPerf Inference v6.1: GB300 and GB200 gain the same ~1.29× on models that can and can’t be capacity-bound · <a href="https://developer.nvidia.com/deep-learning-performance-training-inference/ai-inference">Nvidia MLPerf results</a></li><li><strong>Packaging</strong> — TSMC: 9.5-reticle CoWoS in volume in 2027 with 12 or more HBM stacks · <a href="https://www.trendforce.com/news/?p=38440">TrendForce, Apr 24, 2025</a></li><li><strong>Packaging</strong> — TSMC: CoWoS with 20 HBM stacks in 2028 and up to 24 in 2029 · <a href="https://www.trendforce.com/news/2026/05/14/news-tsmc-sees-ai-wafer-demand-rising-11x-from-2022-2026-targets-cowos-with-24-hbm-stacks-in-2029/">TrendForce, May 14</a></li><li><strong>KV cache</strong> — GB300 serving DeepSeek-R1: KV pool ~61% of HBM, weights ~14%; 1.5× the HBM gave 1.6× the decode batch · <a href="https://github.com/lm-sys/lm-sys.github.io/blob/main/blog/2026-02-19-gb300-longctx.md">LMSYS, Feb 19</a></li><li><strong>KV cache</strong> — FP4 KV holds 1.78× more than FP8; −1.6 points on SWE-bench Verified in one test · <a href="https://github.com/lm-sys/lm-sys.github.io/blob/main/blog/2026-09-16-nvfp4-kv-cache.md">LMSYS, Sep 16</a></li><li><strong>KV cache</strong> — DeepSeek-V4 stores 8.7× less KV than V3.2 at 1M tokens · <a href="https://github.com/vllm-project/vllm-project.github.io/blob/main/_posts/2026-04-24-deepseek-v4.md">vLLM, Apr 24</a></li><li><strong>KV cache</strong> — Coding-agent traces: 94% prefix-cache hits, 131:1 input to output, 5.2 s median pause; a shared KV store lifted throughput 3.8× and cut time-to-first-token 46× · <a href="https://github.com/vllm-project/vllm-project.github.io/blob/main/_posts/2026-05-06-mooncake-store.md">vLLM × Mooncake, May 6</a></li><li><strong>KV cache</strong> — Agent calls hit 85–97% cache after the first; agent teams 97.2% · <a href="https://github.com/ai-dynamo/dynamo/blob/main/docs/fern/pages/blog/2026/agentic-inference-optimizations.mdx">Nvidia Dynamo blog</a></li><li><strong>KV cache</strong> — DeepSeek-V4-Pro agent inputs: median 88K tokens, 675K at the 99th percentile · <a href="https://newsletter.semianalysis.com/p/agentx-inferencexv3-does-cuda-moat">SemiAnalysis AgentX, Aug 24</a></li><li><strong>KV cache</strong> — 57.5 concurrent agents per GPU for DeepSeek-V4-Pro on GB300 (Artificial Analysis AgentPerf) · <a href="https://github.com/NVIDIA/TensorRT-LLM/blob/main/docs/source/blogs/tech_blog/blog26_DeepSeek_V4_on_NVIDIA_Blackwell_Model_Specific_and_Agentic_Workload_Optimizations_in_TensorRT-LLM.md">TensorRT-LLM tech blog 26</a></li><li><strong>KV cache</strong> — Model configs: Llama 3.1 layers and KV heads · <a href="https://raw.githubusercontent.com/meta-llama/llama-models/main/models/sku_list.py">Meta llama-models</a></li><li><strong>KV cache</strong> — Kimi K3: ~27 KB of MLA KV per token, ~434 MB of recurrent state per request · <a href="https://github.com/lm-sys/lm-sys.github.io/blob/main/blog/2026-07-27-kimi-k3-day0-support.md">LMSYS, Jul 27</a></li><li><strong>KV cache</strong> — Qwen3.8: 23 grouped-query layers with 4 KV heads of 256, plus linear-attention state · <a href="https://github.com/sgl-project/sglang/blob/main/docs/cookbook/autoregressive/Qwen/Qwen3.8.mdx">SGLang cookbook</a></li><li><strong>KV cache</strong> — GLM-5.2: a measured FP8 KV page of 2.95 MiB per 64 tokens · <a href="https://github.com/NVIDIA/TensorRT-LLM/blob/main/docs/source/blogs/tech_blog/blog30_KV_Cache_Compression_in_TensorRT-LLM.md">TensorRT-LLM tech blog 30</a></li><li><strong>KV cache</strong> — Google: more than 3.2 quadrillion tokens a month in May 2026, against 480 trillion a year earlier · <a href="https://www.shacknews.com/article/149205/google-3-2-quadrillion-monthly-ai-tokens">Shacknews</a></li><li><strong>Lower precision</strong> — Part of GB300’s 1.6× software gain credited to lower KV-cache precision · <a href="https://blogs.nvidia.com/blog/vera-rubin-nvl72-mlperf-inference">Nvidia blog</a></li><li><strong>Groq 3 LPX</strong> — SRAM decode racks: 500 MB per chip, 128 GB per rack, paired with Rubin GPUs · <a href="https://developer.nvidia.com/blog/inside-nvidia-groq-3-lpx-the-low-latency-inference-accelerator-for-the-nvidia-vera-rubin-platform">Nvidia technical blog</a></li><li><strong>Tiers</strong> — KV offload to a CXL pool at ~92% of local-DRAM throughput · <a href="https://download.semiconductor.samsung.com/resources/white-paper/Optimizing_KV_Cache_Offloading_to_CMM-D_in_a_CXL_Switch-based_Memory_Pool.pdf">Samsung CMM-D white paper</a></li><li><strong>Tiers</strong> — Alibaba’s CXL pool cut time-to-first-token ~90% against an RDMA store · <a href="https://arxiv.org/html/2511.20172v2">arXiv</a></li><li><strong>Tiers</strong> — BlueField-4 context-memory storage, 2H 2026, “up to 5×” tokens/s · <a href="https://nand-research.com/research-note-improving-inference-nvidias-inference-context-memory-storage-platform/">NAND Research</a></li><li><strong>Tiers</strong> — ~16 TB of NAND per Rubin GPU; ~115 EB in 2027, ~9% of global NAND · <a href="https://wccftech.com/nvidia-next-generation-of-ai-systems-could-gobble-up-millions-of-terabytes-of-nand-ssd/amp/">Citi, via Wccftech</a></li><li><strong>Tiers</strong> — High Bandwidth Flash: samples 2027, production 2028 · <a href="https://www.trendforce.com/news/2026/08/14/news-sandisk-reportedly-tapes-out-first-hbf-product-targets-2027-samples-and-2028-production/">TrendForce, Aug 14</a></li><li><strong>Tiers</strong> — Enterprise SSD 37% of 2026 NAND demand · <a href="https://cdn.technews.tw/2026/04/07/cfm-kv-cache/">CFM, via TechNews, Apr 7</a></li><li><strong>Tiers</strong> — NAND shortfall of 6.1% in 2027 · <a href="https://wallstreetcn.com/articles/3782525">Citi, via Wallstreetcn, Sep 21</a></li><li><strong>Tiers</strong> — Structera: 48 TB behind one CXL switch · <a href="https://futurumgroup.com/insights/marvell-scales-ai-memory-to-48tb-behind-a-single-cxl-switch-at-fms-2026/">Futurum</a></li><li><strong>Marvell</strong> — Celestial AI acquisition closed Feb 2: $3.25B plus up to $2.25B tied to revenue · <a href="https://www.marvell.com/company/newsroom/marvell-completes-acquisition-of-celestial-ai.html">Marvell</a></li><li><strong>Marvell</strong> — Photonic Fabric memory module: 2 TB of DDR5 behind 72 GB of HBM3E; a 2U appliance of ~33 TB for up to 16 servers · <a href="https://hoti.org/assets/slides/2025_08_21_day2_Afternoon_CelestialAI_sponsor_talk.pdf">Celestial AI, HOTI 2025</a></li><li><strong>Marvell</strong> — Memory module samples to “multiple customers” in 2027; interconnect revenue targets · <a href="https://stockanalysis.com/stocks/mrvl/transcripts/724088-investor-day-2026/">Marvell Investor Day transcript</a></li><li><strong>Marvell</strong> — Amazon warrant for 1,045,171 shares at $87.00, vesting with Photonic Fabric purchases through 2030 · <a href="https://www.sec.gov/Archives/edgar/data/1835632/000119312525305271/d81371d8k.htm">Marvell 8-K</a></li><li><strong>Marvell</strong> — Photonic Fabric link energy and latency · <a href="https://www.hc2025.hotchips.org/assets/program/conference/day2/76-Hotchips_25_Celestial_AI_v4.pdf">Celestial AI, Hot Chips 2025</a></li><li><strong>China</strong> — CXMT cuts the global DRAM deficit from 15.5% to 9.4% in 2026 and from 12.0% to 5.6% in 2027 · <a href="https://xenospectrum.com/en/cxmt-dram-supply-growth-2027/">Hanwha, via XenoSpectrum</a></li><li><strong>China</strong> — CXMT capacity and output · <a href="https://www.trendforce.com/news/?p=64555">TrendForce</a></li><li><strong>China</strong> — CXMT to double DRAM output · <a href="https://www.digitaltoday.co.kr/en/view/116690/china-cxmt-to-double-dram-output-but-easing-memory-shortage-seen-unlikely">DigitalToday</a></li><li><strong>China</strong> — CXMT begins HBM3E production · <a href="https://www.koreajoongangdaily.com/business/chinas-cxmt-begins-hbm3e-production-as-chip-battle-with-korean-powerhouses-widens/12854260">JoongAng Daily</a></li><li><strong>China</strong> — CXMT capacity sold out through 2027 · <a href="https://www.guru3d.com/story/cxmt-dram-capacity-reportedly-sold-out-through-2027-as-major-oems-fight-for-supply/">Guru3D</a></li><li><strong>China</strong> — CXMT 41% of China’s DRAM demand in 2026, ~50% by 2028 (Goldman) · <a href="https://www.digitaltoday.co.kr/en/view/96664/cxmt-hits-production-ceiling-dimming-hopes-of-easing-dram-shortage">DigitalToday</a></li><li><strong>China</strong> — ~1.5M Ascend chips in 2026, ~240k with domestic HBM · <a href="https://epoch.ai/publications/huaweis-roadmap-to-2031">Epoch AI</a></li><li><strong>China</strong> — Ascend prices up 20–50%; 950DT above RMB250k · <a href="https://www.trendforce.com/news/?p=63912">TrendForce</a></li><li><strong>China</strong> — Huawei’s pre-control stockpile of ~13M HBM stacks · <a href="https://the-decoder.com/huaweis-ai-chip-production-boom-reportedly-faces-a-critical-shortage-of-high-bandwidth-memory/">The Decoder</a></li><li><strong>China</strong> — H200 licence terms: 25% to the US Treasury, at most 75k units per buyer · <a href="https://introl.com/de/blog/bis-export-policy-h200-mi325x-china-case-by-case-2026">Introl</a></li><li><strong>China</strong> — Chinese firms ordered more than 2M H200s (Reuters) · <a href="https://fintool.com/news/china-blocks-nvidia-h200-customs">Fintool, Jan 18</a></li><li><strong>China</strong> — About 20k H200s delivered, to ByteDance and Tencent, by August · <a href="https://www.spheron.network/blog/nvidia-h200-china-shipment-2026/">Spheron</a></li><li><strong>China</strong> — Samsung and SK hynix China fabs lose validated end-user status on Dec 31, 2025 · <a href="https://www.govinfo.gov/content/pkg/FR-2025-09-02/html/2025-16735.htm">Federal Register</a></li><li><strong>China</strong> — YMTC adds ~486k NAND wafers in 2027 (Omdia) · <a href="https://en.fnnews.com/news/202608301656577940">FNNews</a></li><li><strong>Terafab</strong> — Launched Mar 21 with talk of 100,000 wafer starts a month initially, 1,000,000 eventually; Texas filings put Phase 1 construction from Dec 1, 2026 to end-2028 · <a href="https://en.wikipedia.org/wiki/Terafab">Wikipedia, citing filings and the Austin American-Statesman</a></li><li><strong>Terafab</strong> — $16.8B first phase at Grimes County, to run Intel’s 14A process · <a href="https://www.tomshardware.com/tech-industry/semiconductors/terafab-starts-to-take-shape-100-million-square-feet-of-manufacturing-space-and-usd16-8b-initial-capital-investment">Tom’s Hardware, Aug 7</a></li><li><strong>Terafab</strong> — Volume logic production targeted for Q3–Q4 2028, “with big error bars” · <a href="https://www.thestar.com.my/tech/tech-news/2026/10/21/spacex-and-tesla039s-terafab-targets-2028-for-volume-chip-production-musk-says">Reuters, via The Star, Oct 21</a></li><li><strong>Terafab</strong> — Hiring a memory design interface architect with HBM and LPDDR experience; research fab to start operating in 1H 2027 · <a href="https://en.sedaily.com/international/2026/10/23/musks-hiring-spree-maps-out-teslas-ai-and-chip-ambitions">Seoul Economic Daily, Oct 23</a></li><li><strong>Terafab</strong> — Hiring for sub-20nm DRAM process integration · <a href="https://www.trendforce.com/news/2026/08/07/news-teslas-terafab-takes-shape-in-texas-with-16-8b-first-phase-memory-engineering-hiring-signals-dram-push/">TrendForce, Aug 7</a></li><li><strong>Terafab</strong> — CXMT adds about 100,000 wafer starts a month in 2027 · <a href="https://www.trendforce.com/news/2026/09/24/news-cxmt-ymtc-ramp-memory-capacity-but-chinas-ai-cloud-boom-could-soak-up-new-supply-through-2027/">TrendForce, Sep 24</a></li><li><strong>Terafab</strong> — About 13 GW of compute by end-2027, including a 440,000-GPU Vera Rubin cluster · <a href="https://www.diariobitcoin.com/tecnologia/musk-acelera-terafab-con-entregas-anticipadas-de-asml-y-apunta-a-producir-chips-en-2028/">DiarioBitcoin</a></li><li><strong>Terafab</strong> — Musk calls memory “the biggest constraint” · <a href="https://en.sedaily.com/technology/2026/08/08/musk-says-memory-is-biggest-constraint-as-terafab-advances">Seoul Economic Daily, Aug 8</a></li><li><strong>New entrants</strong> — Saimemory: Z-Angle Memory prototypes by Mar 2028, commercialization in FY2029 · <a href="https://www.softbank.jp/en/corp/news/press/sbkk/2026/20260203_01/">SoftBank</a></li><li><strong>Markets</strong> — SK hynix down ~57% from its June peak at ~2.7× 2027 earnings · <a href="https://www.itiger.com/news/1158341353">BofA, via Tiger, Oct 16</a></li><li><strong>Markets</strong> — Micron EPS of $225 in fiscal 2028 and $134 in fiscal 2029 · <a href="https://www.itiger.com/news/1129623218">Morgan Stanley, via Tiger, Oct 2</a></li><li><strong>Markets</strong> — IDC expects 2026 PC shipments to fall 11.3% · <a href="https://www.idc.com/resource-center/press-releases/idc-cuts-2026-pc-outlook-to-11-3-as-memory-shortages-and-supply-chain-disruptions-persist-into-2027/">IDC, Mar 12</a></li></ul>
<h2>Method</h2>
<p><strong>The balance.</strong> A calendar-year model of HBM from SK hynix, Samsung and Micron against the HBM inside the accelerators that use it. Supply = average HBM wafer starts a month × 12 × good TB per wafer start. Demand = K × [(1 − L) × C(y) + L × C(y + 1)], where C(y) is the HBM inside the year’s shipments (units × GB per package, summed over families), L = 0.35 is the share of a package’s HBM bought in the calendar year before it ships (stacks move through CoWoS, boards and buffers a quarter or more ahead), and K = 1.06 covers assembly losses and physical capacity above the usable figure. Supply vs demand = supply ÷ demand − 1. Supportable GB per package = supply ÷ (K × [(1 − L) × U(y) + L × U(y + 1)]), with U the packages shipped; the announced and trimmed averages use the same weights. China’s domestic HBM and accelerators sit outside the balance; licensed H200s are inside Nvidia’s units. LPDDR, GDDR and flash effects are discussed, not modelled.</p>
<ul><li><strong>HBM wafer starts, thousand a month</strong> — 2026: 450 (420–480) · 2027: 590 (540–640) · 2028: 710 (620–820) · 2029: 840 (700–1,000) · 2030: 975 (780–1,180)</li><li><strong>Good TB per wafer start</strong> — 2026: 0.72 (0.70–0.78) · 2027: 0.85 (0.80–0.90) · 2028: 0.94 (0.87–1.00) · 2029: 1.00 (0.91–1.08) · 2030: 1.05 (0.95–1.15)</li></ul>
<p><strong>Supply inputs.</strong> Vendor wafer figures run to 2027 (TrendForce, ETNews); after that HBM keeps about a third of a big-three DRAM base growing toward ~3M wafer starts a month by 2030. Good TB per wafer is calibrated to TrendForce’s bit totals, then rises 5–10% a year with 1c/1d nodes and 32 Gb dies, less than node gains alone because HBM4E and HBM5 dies are larger.</p>
<p><strong>Units.</strong> 2026–27 bottom-up from Nvidia’s fiscal 2028 guide, Morgan Stanley’s 2027 CoWoS demand, TPU forecasts and the AMD, Broadcom and Anthropic deal books; then about 12% in 2028 and 8% a year after. Low is the capex- and power-limited top-down view; high takes the deal books at face value. We count packages; Rubin Ultra counts in two-die equivalents so units compare across generations (Nvidia’s GTC 2025 names counted dies as GPUs; its current names, like ours, count packages).</p>
<p><strong>Memory per package.</strong> “Announced” uses vendor specs where they exist (Rubin Ultra 1 TB on four dies, MI455X 432 GB, MTIA 500 up to 512 GB) and extends their trend where they don’t, putting Feynman-class parts at about 0.8–0.9 TB per two-die package in 2029–30. “Trimmed” uses shipping or reported configurations: Rubin 288 GB, Rubin Ultra 192–256 GB on two dies, a 288 GB MI450, 8-Hi stacks on the 2027 ASICs, and Feynman near 400–460 GB. Per-family values are in the data under the memory-per-package chart.</p>
<p><strong>KV cache.</strong> Bytes per token from each model’s configuration at FP8 (2 × layers × KV heads × head size for standard attention; 576 bytes per layer for latent attention), times 131,072 tokens, plus the fixed recurrent state of linear-attention hybrids. GLM-5.2 uses a measured FP8 page; DeepSeek-V4-Pro is as deployed (FP8 attention, FP4 indexer, vLLM’s formula). The throughput, concurrency and spill figures are our estimates from published serving configurations.</p>
<p><strong>The </strong><a href="https://06.ai/packs/ai-compute-balance"><strong>compute-balance</strong></a><strong> pack</strong><strong> cross-check.</strong> Chip power per package is our estimate by family and year (Nvidia 1.4 kW in 2026 rising to 2.6 kW by 2030; TPUs, Trainium and other ASICs about 0.8–1.3 kW; AMD 1.4–2.2 kW), times units, times two for the whole facility, as in the <a href="https://06.ai/packs/ai-compute-balance">compute-balance</a> pack.</p>
<p><strong>Known limits.</strong> Supply after 2027 is our estimate: no vendor publishes HBM capacity that far out, and the low and high cases are the honest range. HBM4 and HBM4E bits per wafer are undisclosed; a 10% miss in good TB per wafer moves the balance about 10 points. Units after 2027 are set by power, not chips. 2025 doesn’t reconcile: supply of about 2.8 EB exceeds the 2.0 EB our lead implies was bought; we read the gap as inventory across the chain, and some may be units we undercount. Units and specs don’t respond to price in the model; in practice price clears the market, so a shortfall here is pressure on price and specs, not a shortage that will occur.</p>
<hr>
<p>Every chart above links to its interactive version on 06.ai, with the data behind it. The daily trackers are at <a href="https://06.ai/trackers">06.ai/trackers</a>.</p>
<p><em>Not investment advice. Estimates are labeled as estimates.</em></p>]]></content:encoded>
</item>
<item>
  <title>Can ads alone pay for the AI capex?</title>
  <link>https://inthesingularity.ai/packs/ai-capex-payback</link>
  <guid isPermaLink="true">https://inthesingularity.ai/packs/ai-capex-payback</guid>
  <pubDate>Sun, 25 Oct 2026 12:11:00 GMT</pubDate>
  <description>Meta’s revenue is up 26% on a year ago and Alphabet’s 24%, faster than before the AI buildout. How much of that is AI, and what Meta’s and Alphabet’s AI capex earns if superintelligence, Waymo and every other moonshot never return a cent.</description>
  <category>AI returns</category>
  <enclosure url="https://inthesingularity.ai/newsletter/ai-capex-payback/cover.png" length="58054" type="image/png" />
  <media:content url="https://inthesingularity.ai/newsletter/ai-capex-payback/cover.png" medium="image" type="image/png" />
  <content:encoded><![CDATA[<figure><a href="https://06.ai/packs/ai-capex-payback"><img src="https://06.ai/newsletter/ai-capex-payback/thumb.png" alt="Can ads alone pay for the AI capex?" width="421" height="432"></a><figcaption><a href="https://06.ai/packs/ai-capex-payback">Interactive version and data on 06.ai →</a></figcaption></figure>
<p>Before the AI buildout Meta’s revenue grew about 17% a year and Google’s ads about 15%; in 2022–23 both averaged single digits. Meta now grows <strong>26%</strong> on more than three times its 2019 revenue. We estimate AI accounts for about <strong>15 of Meta’s 27 points</strong> of ad growth this year, about <strong>$55B</strong>, and about 5 of Google’s 14. Count only those gains, charge every dollar of AI capex, superintelligence labs included, and Meta’s spending through 2026 earns about <strong>35% a year</strong>: core ads pay for the moonshot’s capex. The 2027 step-up is the bet. If it earns nothing new, Meta’s return falls to about <strong>18%</strong>, or 12% once spending outside capex is counted. Alphabet’s, with only Cloud’s backlog to show for 2027, falls to about <strong>13%</strong>. These are the middle of our ranges.</p>
<ul><li><strong>15 of 27 pts</strong> — AI’s share of Meta’s 2026 ad growth about $55B of revenue, 22% of the year’s ads, against a no-AI trend of 11% a year</li><li><strong>35%</strong> — Meta’s return on AI capex through 2026 a year, from ads alone with every moonshot at zero. Paid back in 2029; the ad gains could carry 2.5× that capex at a 10% hurdle</li><li><strong>18%</strong> — Add 2027’s ~$235B, no extra payoff. 12% counting cloud leases, AI talent and the moonshot fleet’s running costs; 8% if about $250B more goes into 2028</li><li><strong>13%</strong> — Alphabet, ads and Cloud, through 2027 ads alone 11%, Cloud alone 15%; 23% on capex through 2026</li></ul>
<h2>Meta grows half again as fast as before the buildout; Google’s ads are back to their old pace</h2>
<figure><a href="https://06.ai/packs/ai-capex-payback#growth"><img src="https://06.ai/newsletter/ai-capex-payback/fig-01.png" alt="Meta grows half again as fast as before the buildout; Google’s ads are back to their old pace" width="694" height="320"></a><figcaption>Revenue growth a year. 2019–23: compound growth through the pandemic boom, Apple’s tracking change and the 2022 slump, before AI capex took off. 2022–23: the average of those two years. Latest: the most recent quarter against a year earlier (Meta Q3 2026, Alphabet Q2 2026; Alphabet reports its third quarter on October 28).</figcaption></figure>
<h2>What changed</h2>
<p><strong>Meta.</strong> Revenue grew 17.5% a year from 2019 to 2023 and averaged about 7% in 2022–23. It grew 22% in 2025 and 26% in the third quarter of 2026, on more than three times its 2019 revenue. In constant currency, growth peaked at 29% in the first quarter.</p>
<p><strong>Alphabet.</strong> Outside Cloud it grows about 15%, its 2019–23 pace and twice its 2022–23 rate. Search accelerated from 10% in early 2025 to 19% in the first quarter of 2026 and 17% in the second; about 3 points of the first quarter’s company-wide growth was currency, against 1 in the second. YouTube ads grow about 13%, well below their 2019–23 pace. Cloud is the outlier: 28% growth in early 2025, 82% now, with a $514B backlog.</p>
<h2>Capex went from a fifth of Meta’s revenue to over half</h2>
<figure><a href="https://06.ai/packs/ai-capex-payback#capex"><img src="https://06.ai/newsletter/ai-capex-payback/fig-02.png" alt="Capex went from a fifth of Meta’s revenue to over half" width="694" height="302"></a><figcaption>Capex, including principal payments on finance leases for Meta, as a share of revenue. 2026: the midpoints of capex guidance ($135–145B and $195–205B) against Meta’s guided and Alphabet’s estimated revenue (third quarter at consensus).</figcaption></figure>
<h2>Google Cloud is still accelerating; Meta peaked in the first quarter</h2>
<figure><a href="https://06.ai/packs/ai-capex-payback#quarters"><img src="https://06.ai/newsletter/ai-capex-payback/fig-03.png" alt="Google Cloud is still accelerating; Meta peaked in the first quarter" width="694" height="658"></a><figcaption>Revenue growth against the same quarter a year earlier. Meta in constant currency; its fourth quarter is the midpoint of guidance with about 1 point of currency headwind added back (open dot). Alphabet as reported.</figcaption></figure>
<h2>Is growth accelerating?</h2>
<p><strong>Meta: no, it has leveled off high.</strong> Constant-currency growth rose from 19% in early 2025 to 29% in the first quarter of 2026 and has eased by about a point a quarter since; the fourth-quarter guidance midpoint is about 25%. Against two years earlier Meta is steady at 24–26% a year, so most of the slowdown is tougher comparisons.</p>
<p><strong>Alphabet: only Cloud.</strong> Its ad businesses accelerated through 2025 and have leveled off at about 15% outside Cloud. Cloud has accelerated every quarter, from 28% to 82%. The Street expects Alphabet’s total growth to ease to about 22% in the third quarter, on consensus revenue of $124.6B.</p>
<h2>At Meta, price is doing the work</h2>
<figure><a href="https://06.ai/packs/ai-capex-payback#price"><img src="https://06.ai/newsletter/ai-capex-payback/fig-04.png" alt="At Meta, price is doing the work" width="694" height="322"></a><figcaption>Growth in ad impressions and in the average price per ad, Meta’s Family of Apps, against a year earlier.</figcaption></figure>
<p>Impressions grew 19% in the first quarter of 2026 and 10% in the third, as daily users grew 3%. The average price per ad went the other way, from 6% in the fourth quarter of 2025 to 14%, with no currency effect in the third quarter. Higher prices on slowing volume are consistent with better targeting and conversion, though tighter ad supply would push prices up too. Meta credits ad performance, along with macro conditions and currency, and said on its second-quarter call that ranking advances lifted Facebook ad clicks 8.3% and conversions 15.7%. Volume is the risk: WPP Media expects social ad growth to slow from 2027 as engagement plateaus and chatbots compete for time.</p>
<h2>By our estimate AI adds about $55B to Meta’s ads this year, $30B to Google’s and $38B to Cloud</h2>
<p>Meta ads, 2026: $250B of revenue, $55B of it (22%) from AI at a 11% no-AI trend. AI adds about 15 of this year’s 27 points of growth.</p>
<figure><a href="https://06.ai/packs/ai-capex-payback#uplift"><img src="https://06.ai/newsletter/ai-capex-payback/fig-05.png" alt="By our estimate AI adds about $55B to Meta’s ads this year, $30B to Google’s and $38B to Cloud" width="694" height="362"></a><figcaption>Revenue against a no-AI trend that starts from 2023, the last year before the buildout. The gap is above-trend revenue; we credit AI with 80% of it at Meta, 75% at Google’s ads and 90% at Cloud. 2026 and 2027 are estimates (dashed). Choose the business and the no-AI trend.</figcaption></figure>
<h3>By business</h3>
<figure><a href="https://06.ai/packs/ai-capex-payback#uplift"><img src="https://06.ai/newsletter/ai-capex-payback/fig-06.png" alt="By our estimate AI adds about $55B to Meta’s ads this year, $30B to Google’s and $38B to Cloud — Google ads" width="694" height="393"></a><figcaption><strong>Google ads</strong>, 2026: $335B of revenue, $30B of it (9%) from AI at a 7.5% no-AI trend. AI adds about 5 of this year’s 14 points of growth.</figcaption></figure>
<figure><a href="https://06.ai/packs/ai-capex-payback#uplift"><img src="https://06.ai/newsletter/ai-capex-payback/fig-07.png" alt="By our estimate AI adds about $55B to Meta’s ads this year, $30B to Google’s and $38B to Cloud — Google Cloud" width="694" height="393"></a><figcaption><strong>Google Cloud</strong>, 2026: $102B of revenue, $38B of it (37%) from AI at a 22% no-AI trend. AI adds about 50 of this year’s 74 points of growth.</figcaption></figure>
<h2>How we split AI from trend</h2>
<p>No business has a clean control group, so the trend is a judgment. <strong>Meta, 11% a year:</strong> user growth of 3–5%, more ad load and new surfaces such as WhatsApp and Threads, and ad prices rising with the market. WPP Media puts global ad growth at 8.9% this year excluding US political advertising, and the US at 11.9%. Meta’s own 2019–23 pace, 17.5%, includes a pandemic boom on half today’s base. <strong>Google’s ads, 7.5%:</strong> their 2022–23 pace was 6.6%, Search’s 8.4%. Without AI, Search would arguably have lost queries to chatbots, so this may understate what AI is worth to Google. <strong>Cloud, 22%:</strong> it grew 26% in 2023 and was slowing with scale. The rest of each gap we leave to macro, currency and products that aren’t AI.</p>
<p>At these settings AI adds about 9 and 10 points to Meta’s growth in 2024 and 2025 and 15 in 2026, 22% of this year’s ad revenue; about 3–5 points a year to Google’s ads, 9% of this year’s; and 37% of Cloud’s revenue. For 2027 we assume Meta’s ads grow 18%, Google’s 11% and Cloud 55%. AI’s points are the growth in AI’s revenue over last year’s revenue, so they and the rest add up to the year’s growth.</p>
<h2>Ads alone pay for Meta’s AI capex through 2027, even if superintelligence earns nothing</h2>
<p>At a 11% no-AI trend with 80% of the gap credited to AI: 35% on capex through 2026, 18% adding 2027 and 8% adding 2028; all-in 27%, 12% and 3%.</p>
<figure><a href="https://06.ai/packs/ai-capex-payback#meta-irr"><img src="https://06.ai/newsletter/ai-capex-payback/fig-08.png" alt="Ads alone pay for Meta’s AI capex through 2027, even if superintelligence earns nothing" width="694" height="306"></a><figcaption>Internal rate of return on Meta’s AI capex from 2022, counting only AI’s share of above-trend ad revenue. Every dollar of capex above 17% of no-AI ad revenue is charged, Meta Superintelligence Labs included. All-in adds AI spending outside capex, $15B a year in 2026–31 and $7B after (cloud contracts, data-center joint ventures, AI talent), and running costs for the moonshot share of the fleet, 3% of its capex a year: our assumptions. The line is a 10% hurdle.</figcaption></figure>
<h2>How the return is counted</h2>
<p><strong>Returns.</strong> AI’s share of above-trend ad revenue at a 75% cash margin, after the cost of serving ads and before depreciation, taxed at 15%. Capex earns from the year after it’s spent, so 2027’s ad gains count against capex through 2026. After 2027 the gain is held as a share of the business, since a better ad system multiplies whatever advertisers spend, and fades 1.5% a year: no further share gains from AI.</p>
<p><strong>Costs.</strong> All capex above 17% of no-AI ad revenue: $187B over 2022–26, and $201B more if 2027 comes in at $235B. Meta hasn’t guided 2027; in July the buy-side expected $225–250B. The 17% baseline is close to Meta’s 2021 capex (16% of revenue) and below its 2019–23 average (about 21%), so it charges more to AI than a higher baseline would; at 21% the return through 2026 would be 46%. We put servers at 60% of capex, Alphabet’s split. They are replaced every six years at 75% of their cost for half the fleet, the part the ad gains need; the rest belongs to the moonshots and earns nothing. Servers are expensed at once for tax, data centers over 20 years. Nothing is counted after 2040.</p>
<p><strong>Results.</strong> Through 2026: 35%, paid back in 2029. At a 10% hurdle the ad gains could carry 2.5× that capex. With 2027 at $235B and nothing extra to show: 18%, paid back in 2032; each further $10B in 2027 takes about 0.6 points off. With another $250B in 2028: 8%, paid back in 2036. All-in: 27%, 12% and 3%. In our base case another year of moonshot spending, 2028, is what takes the return below 10%; all-in, 2027 alone brings it close. At the low end of our ranges 2027 can do it on its own, as the grid above and the simulation below show.</p>
<h2>Spend through 2026 pays back in 2029; each further year of moonshot capex pushes payback out three to four years</h2>
<figure><a href="https://06.ai/packs/ai-capex-payback#jcurve"><img src="https://06.ai/newsletter/ai-capex-payback/fig-09.png" alt="Spend through 2026 pays back in 2029; each further year of moonshot capex pushes payback out three to four years" width="694" height="362"></a><figcaption>Cumulative after-tax cash flow of Meta’s AI program, $B: AI’s share of above-trend ad profit, less incremental capex, server replacements and tax. Base case, no off-capex spending. The dots mark payback.</figcaption></figure>
<h2>At Alphabet, Cloud carries the 2027 build; the ad case turns on what Search would have done without AI</h2>
<p>At a 7.5% no-AI trend for Google’s ads and a 50% Cloud margin: ads and Cloud together return 23% on capex through 2026 and 13% through 2027; ads alone 11% and Cloud alone 15% through 2027.</p>
<figure><a href="https://06.ai/packs/ai-capex-payback#goog-irr"><img src="https://06.ai/newsletter/ai-capex-payback/fig-10.png" alt="At Alphabet, Cloud carries the 2027 build; the ad case turns on what Search would have done without AI" width="694" height="264"></a><figcaption>Internal rate of return on Alphabet’s AI capex with every moonshot at zero, for capex through 2026 and through 2027 (about $300B in 2027, on analysts’ projections). Ads and Cloud each carry half of the incremental capex, as about half of 2026’s ML compute goes to Cloud; together they carry all of it. Cloud’s revenue is capacity-bound: with capex through 2026 it stays at its end-2026 run-rate ($122B a year); with 2027’s build it converts backlog in 2028; after that its AI revenue is held in dollars and fades.</figcaption></figure>
<h2>Alphabet’s two legs</h2>
<p><strong>Ads.</strong> Charged half the capex, Google’s AI ad gains return 33% on spend through 2026 and 11% through 2027. That rests on something nobody can observe: at a 9.5% no-AI trend the 2027 case returns −1%, at 5.5% 21%. At a 10% hurdle the ad gains alone could carry 105% of all Alphabet’s incremental capex through 2026, but only 53% through 2027.</p>
<p><strong>Cloud.</strong> Revenue grew 82% to $24.8B in the second quarter, with an operating margin of 35.6%. We use a 50% cash margin on its AI revenue, before depreciation; Cloud’s figures include Gemini’s API, compute rented to AI labs and TPU sales. Charged the other half of the capex, Cloud returns 11% through 2026, when its revenue stops at the end-2026 run-rate, and 15% through 2027, as the 2027 build converts backlog: compute-rental economics, steady but not spectacular. Alphabet says renting third-party capacity as a bridge while it builds its own will weigh on Cloud’s margins.</p>
<p><strong>Together</strong> 23% through 2026 and 13% through 2027, paid back in 2034. The 2027 build pays well only if Cloud keeps compounding beyond its backlog or the models earn money of their own.</p>
<h2>Spend through 2026 very likely pays; the 2027 build is likely at Meta, closer to a coin flip at Alphabet</h2>
<figure><a href="https://06.ai/packs/ai-capex-payback#odds"><img src="https://06.ai/newsletter/ai-capex-payback/fig-11.png" alt="Spend through 2026 very likely pays; the 2027 build is likely at Meta, closer to a coin flip at Alphabet" width="694" height="288"></a><figcaption>Internal rate of return across 8,000 draws, each assumption uniform across its range, centered on the base case: the no-AI trend (Meta 9–13%, Google’s ads 5.5–9.5%, Cloud 18–26%), AI’s share of the gap (Meta 65–95%, Google’s ads 55–95%, Cloud 80–100%), cash margins (Meta 65–85%, Google’s ads 60–80%, Cloud 35–65%), 2027 growth, the gain fading 0–3% a year (Cloud 0–5%), the capex baseline (Meta 15–19%, Alphabet 10–12%), 2027 capex ($205–265B Meta, $260–340B Alphabet), the share of the fleet replaced (35–65%), server life (5–7 years), replacement cost (60–90%), Alphabet’s split between Cloud and the rest (40–60%) and, all-in, $5–25B a year of AI spending outside capex in 2026–31 ($3–11B after) and running costs of 2–4% a year. Fixed: tax rates, the horizon, 2026 capex and the server share of capex. Bars run from the 10th to the 90th percentile; the dot is the median.</figcaption></figure>
<h2>When does the cash come back?</h2>
<p>Spend through 2026 clears a 10% return in all of our 8,000 draws for Meta and 93% for Alphabet. Add 2027 with nothing extra from the moonshots and it clears in 93% of draws for Meta, 61% all-in, and 66% for Alphabet. The medians, 34%, 17% and 11% for Meta and 22% and 13% for Alphabet, sit within a point of the base case.</p>
<p>A vintage that earns its cost of capital still doesn’t reach shareholders while the race is on. Meta bought back $26.2B of stock in the first nine months of 2025 and none so far in 2026; its free cash flow was −$2.1B in the third quarter, and it is creating a non-voting Class C share, partly to raise capital. Alphabet bought back no stock in the first half of 2026 and raised about $106B: $56B of debt, $30B of common stock and $19B of preferred. Cash comes back when capex plateaus, and that runs against success: Zuckerberg’s stated fallback if superintelligence takes longer is to put the compute into the core business and, at worst, slow new building for a while. In our judgment, buybacks back at 2025’s pace by the end of 2028 are about a 30% chance at each company, likelier if the moonshots disappoint than if they work.</p>
<h2>What to watch</h2>
<p><strong>Alphabet’s third quarter, October 28.</strong> Search against 17% and Cloud against 82%. A Cloud slowdown with capex still rising is the case where the 2027 build doesn’t pay.</p>
<p><strong>Meta’s 2027 capex.</strong> Meta hasn’t guided yet. Each $10B above $235B with nothing extra to show costs about 0.6 points of return on the program.</p>
<p><strong>Price against volume at Meta.</strong> Price per ad growing faster than impressions is the ranking gains showing up. Both slowing together would mean the gains are fading.</p>
<p><strong>Legal costs.</strong> Meta’s settlement with 52 state attorneys general provides for payments of up to about $18.6B, and more youth-related trials are scheduled. None of this is in the model.</p>
<h2>What this is anchored on</h2>
<p>No-AI trends, AI’s share of above-trend revenue, margins, 2026–27 estimates beyond guidance, capex baselines, replacement, off-capex spending and the returns are our estimates, built from the sources below.</p>
<ul><li><strong>Meta Q3 2026</strong> — Revenue $64.60B, up 26% reported and in constant currency; ad impressions +10%, average price per ad +14%; daily users 3.64B, +3%; capex $34.91B; free cash flow −$2.12B; no buybacks in 2026; Q4 guidance $73–76B with about 1 point of currency headwind; 2026 capex $135–145B; a new non-voting Class C share, partly for raising capital · <a href="https://www.sec.gov/Archives/edgar/data/0001326801/000162828026066700/meta-09302026xexhibit991.htm">Meta Q3 2026 results (8-K), Oct 21</a></li><li><strong>Meta Q3 2026</strong> — AI investments already deliver substantial incremental ad revenue; MSL investments will take longer to return than the core business; buybacks of $26.2B in the first nine months of 2025 · <a href="https://www.sec.gov/Archives/edgar/data/0001326801/000162828026066700/meta-09302026xexhibit991.htm">Meta Q3 2026 results (8-K), Oct 21</a></li><li><strong>Meta legal</strong> — An agreement with 52 state attorneys general providing for payments of up to about $18.59B; youth-related trials scheduled in the coming months · <a href="https://www.sec.gov/Archives/edgar/data/0001326801/000162828026066700/meta-09302026xexhibit991.htm">Meta Q3 2026 results (8-K), Oct 21</a></li><li><strong>Meta Q2 2026</strong> — Revenue $60.80B, +28% (+27% in constant currency); impressions +14%, price per ad +12%; capex $31.08B · <a href="https://www.sec.gov/Archives/edgar/data/0001326801/000162828026050596/meta-06302026xexhibit991.htm">Meta Q2 2026 results (8-K), Jul 29</a></li><li><strong>Meta Q2 2026 call</strong> — Price per ad up on ad performance, better macro conditions and currency; ranking advances lifted Facebook ad clicks 8.3% and conversions 15.7% (no period given) · <a href="https://www.marketbeat.com/instant-alerts/meta-platforms-q2-earnings-call-highlights-2026-07-29/">MarketBeat, Jul 29</a></li><li><strong>Meta Q1 2026</strong> — Revenue $56.31B, +33% (+29% in constant currency); impressions +19%, price per ad +12%; ad revenue $55.02B; capex outlook raised to $125–145B · <a href="https://app.edgar.tools/filing/1326801/0001628280-26-028364/meta-03312026xexhibit991.htm">Meta Q1 2026 results (8-K), Apr 29</a></li><li><strong>Meta 2025</strong> — Q4 revenue $59.89B, +24% (+23% in constant currency); ad impressions +18%, price per ad +6% in Q4; 2025 revenue $200.97B; 2025 capex $72.22B · <a href="https://s21.q4cdn.com/399680738/files/doc_financials/2025/q4/Meta-12-31-2025-Exhibit-99-1-FINAL.pdf">Meta Q4 2025 results, Jan 28</a></li><li><strong>Meta 2025</strong> — Q3 revenue $51.24B, +26% (+25% in constant currency); impressions +14%, price per ad +10% · <a href="https://www.sec.gov/Archives/edgar/data/1326801/000162828025047114/meta-09302025xexhibit991.htm">Meta Q3 2025 results (8-K)</a></li><li><strong>Meta 2025</strong> — Q2 revenue $47.52B, +22% reported and in constant currency; impressions +11%, price per ad +9% · <a href="https://sec.gov/Archives/edgar/data/1326801/000162828025036719/meta-06302025xexhibit991.htm">Meta Q2 2025 results (8-K)</a></li><li><strong>Meta 2025</strong> — Q1 revenue $42.31B, +16% (+19% in constant currency); impressions +5%, price per ad +10% · <a href="https://www.sec.gov/Archives/edgar/data/1326801/000132680125000050/meta-03312025xexhibit991.htm">Meta Q1 2025 results (8-K)</a></li><li><strong>Meta 2027 capex</strong> — No 2027 guidance; the buy-side expected $225–250B · <a href="https://longportapp.cn/topics/43054018">Dolphin Research, Jul 29</a></li><li><strong>Meta off-capex</strong> — Most of 2026’s expense growth from infrastructure, including third-party cloud spend · <a href="https://datacenterdynamics.com/en/news/meta-estimates-2026-capex-to-be-between-115-135bn/">DCD, Jan 29</a></li><li><strong>Meta off-capex</strong> — Meta funds its 20% share of Hyperion’s remaining construction as an investment, not capex · <a href="https://www.datacenterdynamics.com/en/news/meta-plans-notably-larger-capex-spend-on-ai-data-centers-in-2026-compute-expectations-already-higher-than-last-quarter-prediction/">DCD, Oct 30, 2025</a></li><li><strong>Meta strategy</strong> — If superintelligence takes longer, the extra compute goes to the core business; at worst Meta slows new building for a while · <a href="https://thefinancialexpress.com.bd/trade/meta-forecasts-bigger-capital-costs-in-2026">Reuters, via The Financial Express</a></li><li><strong>Alphabet Q2 2026</strong> — Revenue $119.8B, +24% (+23% in constant currency); Search &amp; other $63.27B, +17%; YouTube ads $11.06B, +13%; Network $7.30B; Cloud $24.77B, +82%, operating income $8.81B; capex $44.9B; no buybacks in the first half; first-half proceeds of $56.2B of debt, $30.5B of common stock and $19.1B of mandatory convertible preferred · <a href="https://www.sec.gov/Archives/edgar/data/0001652044/000165204426000066/googexhibit991q22026.htm">Alphabet Q2 2026 results (8-K), Jul 22</a></li><li><strong>Alphabet Q2 2026</strong> — Cloud backlog $514B · <a href="https://www.cnhinews.com/news/article_e3220c1b-dca0-5369-ae1d-54f65c6ea076.html">Zacks, via CNHI</a></li><li><strong>Alphabet capex</strong> — 2026 capex guidance raised to $195–205B; Q2 capex $44.9B, about 60% of technical infrastructure in servers · <a href="https://w.media/ai-infrastructure-demand-pushes-alphabets-2026-capex-guidance-to-us-205-billion/">W.Media, Jul 23</a></li><li><strong>Alphabet capex</strong> — Another meaningful step-up in capex in 2027 · <a href="https://www.nasdaq.com/articles/alphabet-heavy-etfs-consider-capex-woes-offset-cloud-strength">Zacks, via Nasdaq</a></li><li><strong>Alphabet capex</strong> — Analysts’ 2027 capex projections above $300B · <a href="https://pepperstone.com/en-ae/insights/market-analysis/equities/alphabet-q2-earnings-preview-cloud-and-capex-take-center-stage">Pepperstone, Jul 21</a></li><li><strong>Alphabet compute</strong> — About half, or just over half, of 2026’s ML compute goes to Cloud · <a href="https://abc.xyz/investor/events/event-details/2026/Anat-Ashkenazi-SVP-and-CFO-of-Alphabet-and-Google-at-the-Morgan-Stanley-Technology-Media-and-Telecom-Conference-on-March-3-2026-2026-r9dUUoVl6o/default.aspx">Alphabet CFO, Morgan Stanley TMT conference, Mar 3</a></li><li><strong>Alphabet Q1 2026</strong> — Revenue $109.9B, +22% (+19% in constant currency); Search $60.40B, +19%; YouTube ads $9.88B, +11%; Network $6.97B; Cloud $20.03B, +63% · <a href="https://www.sec.gov/Archives/edgar/data/0001652044/000165204426000043/googexhibit991q12026.htm">Alphabet Q1 2026 results (8-K), Apr 29</a></li><li><strong>Alphabet 2025</strong> — Q4: Search $63.07B, +17%; YouTube ads $11.38B, +9%; Cloud $17.66B, +48%. Full year: revenue $402.8B; capex $91.4B · <a href="https://www.sec.gov/Archives/edgar/data/1652044/000165204426000012/googexhibit991q42025.htm">Alphabet Q4 2025 results (8-K), Feb 4</a></li><li><strong>Alphabet Q3</strong> — Reports October 28; consensus revenue $124.6B · <a href="https://finviz.com/news/405729/2-magnificent-7-stocks-in-focus-ahead-of-earnings">Finviz, Oct 21</a></li><li><strong>Alphabet Cloud</strong> — Alphabet plans to rent more third-party capacity in Q3 as a bridge while it builds its own, and says the cost will weigh on Cloud’s operating margin · <a href="https://elnuevodiario.com.do/tecnologica-alphabet-eleva-su-prevision-de-gastos-de-capital-en-ia-hasta-205-000-millones/">EFE, via El Nuevo Diario, Jul 22</a></li><li><strong>Ad market</strong> — Global ad revenue +8.9% in 2026 excluding US political advertising, US +11.9%; social media growth slowing from 2027 as engagement plateaus and AI chatbots compete · <a href="https://www.exchange4media.com/advertising-news/global-ad-revenue-to-grow-89-by-end-of-2026-wpp-media-155474.html">WPP Media, via exchange4media, Jun 16</a></li><li><strong>History</strong> — Annual revenue by segment, capex and quarterly revenue, 2017–2025 · <a href="https://investor.atmeta.com/">Meta and Alphabet annual reports (10-K) and earnings releases</a></li><li><strong>History</strong> — Alphabet segment revenue and capex, 2017–2025 · <a href="https://abc.xyz/investor/">Alphabet investor relations</a></li></ul>
<h2>Method</h2>
<p><strong>Clock.</strong> Reported through Meta’s third quarter of 2026 and Alphabet’s second. Meta’s 2026 is nine months reported plus the midpoint of its fourth-quarter guidance, $74.5B, of which about $72.3B ads. Alphabet’s second half is our estimate: Search up 15% and 14%, YouTube ads 12% and 11%, Network down 2%, Cloud $27.0B and $30.5B. 2027: Meta ads +18%, Google ads +11%, Cloud +55% and +30% in 2028 as the backlog converts, both only when 2027’s capex is charged; otherwise Cloud stays at its end-2026 run-rate, $122B. Capex earns from the year after it’s spent.</p>
<p><strong>AI’s share.</strong> Each business against a trend from 2023: Meta’s ads 11% a year, Google’s ads (Search, YouTube, Network) 7.5%, Cloud 22%, held through 2027 and fading to 4% (Cloud 6%) by 2035. AI is credited with 80%, 75% and 90% of the gap. AI points of growth are the growth in AI’s revenue over the year before’s revenue. After 2027 the ad gains are held as a share of the business and fade 1.5% a year; Cloud’s are held in dollars from its last estimated year (2028 with 2027’s build, 2027 without) and fade 2.5% a year. Holding Meta’s gain flat in dollars instead would cut the return through 2026 to 29%.</p>
<p><strong>Returns.</strong> Cash flows from 2022 to 2040, no terminal value. In: AI’s share of above-trend revenue times a cash margin before depreciation (Meta ads 75%, Google ads 70% after traffic acquisition and YouTube content costs, a judgment, Cloud 50%). Out: capex above a pre-buildout intensity (17% of Meta’s no-AI ad revenue; 11% of Alphabet’s no-AI revenue), counted through 2026 or 2027 (2028 in one Meta case); servers, 60% of capex, replaced every 6 years at 75% of their cost for the share of the fleet the gains need (half of Meta’s, half of Alphabet’s outside Cloud, all of Cloud’s); cash tax at 15% (Alphabet 16%) with servers expensed at once and data centers over 20 years, losses offsetting other income. Alphabet’s incremental capex is split half to Cloud. All-in Meta cases add $15B a year of AI spending outside capex in 2026–31 and $7B after, plus running costs for the moonshot share of the fleet at 3% of its capex a year while its servers last. The base case is the middle of every range in the simulation.</p>
<p><strong>What’s left out.</strong> Any revenue from the moonshots: Meta AI and business agents, Meta selling compute, frontier models beyond what Cloud sells, Gemini subscriptions, Waymo. The value of what AI defended rather than added, likely large for Search. Running costs of Alphabet’s non-Cloud moonshot compute, spending on AI researchers beyond Meta’s all-in cases, legal costs, and currency beyond what reported growth shows. The returns are on the AI program alone, not the companies. Monte Carlo: 8,000 draws per case with a fixed seed; the ranges are in the chart’s note. Model and data: scripts/packs/ai_payback.py in the site’s repository.</p>
<hr>
<p>Every chart above links to its interactive version on 06.ai, with the data behind it. The daily trackers are at <a href="https://06.ai/trackers">06.ai/trackers</a>.</p>
<p><em>Not investment advice. Estimates are labeled as estimates.</em></p>]]></content:encoded>
</item>
<item>
  <title>Will AI compute run short or long through 2030?</title>
  <link>https://inthesingularity.ai/packs/ai-compute-balance</link>
  <guid isPermaLink="true">https://inthesingularity.ai/packs/ai-compute-balance</guid>
  <pubDate>Tue, 20 Oct 2026 12:10:00 GMT</pubDate>
  <description>The compute-demand pack sized the compute AI demand needs to the end of 2027. Here we add the supply side, the chips, power and money to switch on new capacity, carry demand to 2030 and compare the two at the end of each year.</description>
  <category>AI compute</category>
  <enclosure url="https://inthesingularity.ai/newsletter/ai-compute-balance/cover.png" length="68331" type="image/png" />
  <media:content url="https://inthesingularity.ai/newsletter/ai-compute-balance/cover.png" medium="image" type="image/png" />
  <content:encoded><![CDATA[<figure><a href="https://06.ai/packs/ai-compute-balance"><img src="https://06.ai/newsletter/ai-compute-balance/thumb.png" alt="Will AI compute run short or long through 2030?" width="421" height="432"></a><figcaption><a href="https://06.ai/packs/ai-compute-balance">Interactive version and data on 06.ai →</a></figcaption></figure>
<p>The <a href="https://06.ai/packs/ai-compute-demand">compute-demand</a> pack sized the compute AI demand needs to the end of 2027. Here we add the supply side, carry demand to 2030 and compare the two at the end of each year. In our base case the shortage all but closes during 2027, when about <strong>38 GW</strong> of new capacity comes online, if power keeps pace with capex. From 2028 demand pulls ahead again: what customers and labs plan to spend on compute grows faster than the power and money for new capacity, and the market ends 2030 <strong>21% short</strong>. Power and money, not chips, limit what is switched on from 2027; training budgets and capex decide whether, and how far, the market runs short.</p>
<ul><li><strong>+38 GW</strong> — Switched on in 2027 of AI data-center capacity, all-in, limited by power and money. The chips the industry could deliver would fill about 47 GW</li><li><strong>−1%</strong> — End of 2027 supply about level with demand, from 19% short at the end of 2026, if power keeps pace with capex (15% short at our power research’s high case)</li><li><strong>−21%</strong> — End of 2030 short again. Demand grows 1.9–2.4× a year in 2028–30, supply 1.8–2.2×, in compute terms</li><li><strong>−69% to +43%</strong> — Biggest swing, 2030 if training and research grow at our high or low path instead of the base</li></ul>
<h2>Short in 2026, about even at the end of 2027, short again by the end of 2028</h2>
<p>Base demand, base supply, demand 20% above supply today: 1% short at the end of 2027 and 21% short at the end of 2030.</p>
<figure><a href="https://06.ai/packs/ai-compute-balance#balance"><img src="https://06.ai/newsletter/ai-compute-balance/fig-01.png" alt="Short in 2026, about even at the end of 2027, short again by the end of 2028" width="694" height="702"></a><figcaption>Compute installed at the end of each year against compute demanded, as a share of demand. Both are counted in gigawatts of today’s hardware, so a gigawatt installed later counts for more: it does more work. The first bar covers October to December 2026. Choose the demand and supply cases and how far demand runs above supply today: 20% above means supply meets 83% of demand, 17% short.</figcaption></figure>
<p>Needed: gigawatts of that year’s hardware it would take, on top of everything installed before, for supply to meet demand at year-end. A surplus or shortfall carries into the next year.</p>
<h2>What the base case says</h2>
<p><strong>Short through 2026.</strong> Today’s market is short. One-year H100 rental contracts rose about 40% between October and March and the B200 rental index is up about 28% this year. Every large cloud says demand exceeds what it can supply, and NVIDIA says its supply supports about 70% revenue growth in its year to January 2028, against demand growth nearer 100%. We assume demand at today’s prices runs 20% above what is being served. The fourth quarter adds about 7 GW, mostly Blackwell racks and new TPUs, while inference demand grows about 56% in the quarter, so the year ends 19% short.</p>
<p><strong>2027 brings supply about level with demand.</strong> About 38 GW comes online, 60% more than the 24 GW switched on in 2026. Most of it is Vera Rubin racks and Google’s eighth-generation TPUs, plus Blackwell racks delivered late in 2026. Each gigawatt does about 4.6× the inference work of today’s average and 2.2× the training work. Supply grows about 3.1× during the year in compute terms and demand about 2.6×, so by December supply is about level with demand (−1%): 37.9 GW switched on against 38.3 GW needed. Month by month, with last year’s late deliveries switched on first, the market comes within 3 points of balance in September 2027 and stays there until August 2028. That fits Goldman’s view that supply and demand rebalance in the second half of 2027 at the earliest and more likely in the first half of 2028; Amazon expects to be short through 2027. It needs about 38 GW of power in 2027, above the top of our power research’s range: its high case is 30 GW in 2027 and 40 GW in 2028. At that high case, 2027 ends 15% short.</p>
<p><strong>Then demand pulls ahead.</strong> In compute terms supply grows 2.2×, 1.9× and 1.8× in 2028–30, and demand 2.4×, 2.1× and 1.9×. Capacity switched on rises only from 43 GW to 51 GW a year, because AI capex grows 11–18% a year on street estimates. Meanwhile what customers spend on inference grows 35–70% a year in our base case, and training budgets keep growing. The market ends 2028 9% short and 2030 21% short. By 2030, 86 GW of new capacity would be needed against 51 GW switched on, because each year’s shortfall carries into the next. Either capex keeps rising faster than analysts expect, or prices rise until demand fits. The demand and supply cases move this a long way; see the controls above and “What tips the balance” below.</p>
<h2>From 2027 the chips could outrun power and money</h2>
<figure><a href="https://06.ai/packs/ai-compute-balance#ceilings"><img src="https://06.ai/newsletter/ai-compute-balance/fig-02.png" alt="From 2027 the chips could outrun power and money" width="694" height="352"></a><figcaption>New AI data-center capacity each year that chips, power and money could each support, GW all-in. Chips: what the chipmakers could deliver from the fourth quarter of the year before through the third quarter, at twice their rated power, Epoch AI’s rule for a data center’s total draw. Power: capacity that can be switched on. Money: AI capex divided by the all-in cost of a gigawatt, chips included. The lowest bar is what comes online. Our estimates from the sources below.</figcaption></figure>
<h2>Supply: chips, power and money</h2>
<p><strong>The chips keep coming.</strong> Epoch AI counts 11.3M H100-equivalents delivered in the first half of 2026. We estimate 33M for the year, above the <a href="https://06.ai/packs/ai-compute-demand">compute-demand</a> pack’s 24–27M, which extended the first half’s pace: NVIDIA’s guidance implies a much larger second half, with Vera Rubin, rated at 8.8 H100-equivalents a chip, shipping from the fourth quarter. In 2027 the chipmakers could deliver about 100M as Rubin ramps, Google ships its eighth-generation TPUs and AMD its MI455X racks; NVIDIA would be about 56% of that and Google 27%. NVIDIA says supply limits its revenue at least through January 2028, Broadcom expects its AI revenue to double to about $115B in fiscal 2027 and again in 2028, and TSMC says AI demand still exceeds what it can make. Rated at about 26 GW, 2027’s chips need about 52 GW to run; allowing a quarter to install them, about 47 GW could be switched on during 2027. After 2027 we let accelerator spending rise from about $1T to $1.25T a year and each generation’s gains do the rest: 170M H100-equivalents in 2028 and 350M in 2030.</p>
<p><strong>Power doesn’t keep up.</strong> About 24 GW came online for AI in 2026, by our estimate from chip deliveries; we assume power for about 27 GW was ready. For 2027 Morgan Stanley expects hyperscalers and neoclouds, 60–95% of new capacity, to add 32 GW. Goldman counts 2.7 times as much US capacity scheduled for 2027 as for 2026 but expects only half to 60% of it on time. Neither says whether it counts IT power or the whole facility. Trackers found most US projects due in 2027 not yet under construction in April, new gas turbines are being booked for delivery as late as 2031 and large transformers take more than three years. We expect about 38 GW of all-in capacity to come online in 2027, rising to 58 GW by 2030. That is within the range those forecasts imply once they are put on an all-in, worldwide basis, and above the top of our own power research’s range through 2029. Its central estimate, which applies project trackers’ delay rates, is about 23 GW in 2027 rising to 35 GW in 2030, with a high case rising from 30 GW in 2027 to 60 GW in 2030. It starts from a lower 2026, about 16 GW, partly because it takes Epoch’s end-2025 installed base at 2.5× chip power where we use 2×. Our path leaves chips for about 9 GW more than can be switched on in 2027; a consensus estimate cited by Elon Musk in August put that at about 15 GW.</p>
<p><strong>Money binds first.</strong> AI capex is about $1T in 2026 (Goldman’s estimate). On street estimates for the hyperscalers, Oracle, SpaceX and the neoclouds, plus China and other buyers, it reaches about $1.4T in 2027 and $2.05T by 2030, growing 11–18% a year after 2027. All-in, a gigawatt of IT power costs about $45B to build today (Epoch puts a GB200 site at $38B, NVIDIA says $50–60B), rising to about $52B by 2030 as chips carry more value per watt. That pays for about 38 GW in 2027 and 51 GW in 2030, at or below our power path, so money is the limit in the base case from 2027. Both sit well below what the chipmakers could ship. The money includes the chips, so when it binds, fewer chips get bought than could be made, or chips wait for power. If these estimates are right, from 2027 new AI capacity is limited less by the fabs than by power and the money to build.</p>
<h2>A gigawatt switched on in 2030 could do 19× the inference work of today’s average</h2>
<figure><a href="https://06.ai/packs/ai-compute-balance#per-gw"><img src="https://06.ai/newsletter/ai-compute-balance/fig-03.png" alt="A gigawatt switched on in 2030 could do 19× the inference work of today’s average" width="694" height="332"></a><figcaption>Work per GW of the capacity switched on each year, as a multiple of the fleet operating today. Inference: tokens per MW on SemiAnalysis’s AgentX benchmark (DeepSeek V4 Pro, 100 tokens/s per user) for each year’s mix of hardware, our estimate from measured results and roadmaps. Training: peak 8-bit FLOP/s per GW, from the chips’ rated power per H100-equivalent. Capacity switched on in a year mixes the previous year’s fourth-quarter deliveries with its own first three quarters; 2026 is the fourth quarter’s.</figcaption></figure>
<h2>Why gigawatt totals mislead</h2>
<p>Today’s fleet is still about a fifth Hopper. On agentic inference a GB300 rack does about 12.6× an H200’s work per MW, and Vera Rubin about 2.1× a GB300’s (SemiAnalysis, measured on pre-release software). So the capacity switched on in 2027 does about 4.6× today’s average per gigawatt at inference, and 2030’s about 19× if Rubin Ultra and Feynman deliver what their roadmaps suggest. Training gains ride on FLOPs per watt, which improve about 30% a year: 2.2× today’s average for 2027’s capacity and 5.0× for 2030’s. These are peak 8-bit FLOPs, the way Epoch counts them. The training throughput our efficiency research expects newer chips to achieve is lower for capacity switched on through 2029; counted that way, 2027 would end 11% short and 2030 25% short. That research labels its tables by install year. We read them as each year’s deliveries, because its 2026 mix includes Rubin, which only ships from the fourth quarter; read as install years, 2027 would end 3% long and 2030 17% short.</p>
<p>Cheaper inference also raises demand. Capacity switched on in 2028 does about 58% more inference work per dollar than 2027’s, and about 60% and 52% more again in 2029 and 2030. If prices follow costs, each dollar customers spend buys that much more, so after 2027 inference demand rises with hardware gains just as supply does. Faster hardware therefore moves the balance mainly through training, which we count in FLOPs, and through the years to 2027, when demand follows usage.</p>
<p>Who gets the new hardware matters too. We split each year’s hardware between inference and training in the same proportion. Newer hardware’s edge is larger at inference (4.6× today’s average in 2027, against 2.2× at training). If it went to inference first and older chips did the training, the market would serve about 6–7 points more of demand each year. If the newest went to training first, as frontier labs train on their newest clusters, it would serve about 8–12 points less.</p>
<h2>After 2027 demand grows faster than supply</h2>
<figure><a href="https://06.ai/packs/ai-compute-balance#growth"><img src="https://06.ai/newsletter/ai-compute-balance/fig-04.png" alt="After 2027 demand grows faster than supply" width="694" height="322"></a><figcaption>Growth during each year in the base case, in compute terms. Supply: the compute installed, measured at that year’s mix of inference and training demand. Demand: inference plus training and research.</figcaption></figure>
<h2>Demand to 2030</h2>
<p><strong>Inference to 2027</strong> is the <a href="https://06.ai/packs/ai-compute-demand">compute-demand</a> pack’s: OpenRouter’s compute on the <a href="https://06.ai/packs/ai-demand">AI-demand</a> pack’s three paths. To scale it to the whole market we compared OpenRouter’s growth with six disclosed token counts over the same windows. The market grew at 0.32–0.86 of OpenRouter’s pace, measured in doublings, with a median of 0.63; we use 0.6. We apply it to OpenRouter’s compute rather than its tokens, because OpenRouter’s mix is probably moving to small, cheap models faster than the market’s; applied to tokens, demand at the end of 2027 would be about 28% lower. That puts the market’s inference demand at 4.3× today’s by the end of 2027, about 2.8× during 2027 alone. The <a href="https://06.ai/packs/ai-compute-demand">compute-demand</a> pack’s headline, about 47 GW of new GB300 racks by December 2027, assumed the whole market grew at OpenRouter’s full pace. At 0.6 of it, measured as the <a href="https://06.ai/packs/ai-compute-demand">compute-demand</a> pack does from October’s average to December’s, inference needs about 15 GW of GB300s, or about 10 GW of the capacity switched on in 2027, before counting today’s shortage.</p>
<p><strong>After 2027, spending decides.</strong> The <a href="https://06.ai/packs/ai-demand">AI-demand</a> pack’s paths have OpenRouter’s growth halving every three, six or twelve months. Carried past 2027, the base path has OpenRouter’s compute growing 65% in 2028, 14% in 2029 and 3% in 2030, which describes one marketplace maturing rather than the market. Instead we let inference demand follow what customers spend on it: 1.7× in 2028, 1.5× in 2029 and 1.35× in 2030 in the base case. That is slower than the leading labs plan to grow. OpenAI’s plan takes revenue from $36B in 2026 to $350B in 2030, about 1.8× a year, and Anthropic’s internal forecast of $190–200B in 2028 implies about 1.9× a year from an estimated $55B in 2026. Each dollar buys as much more inference as the capacity switched on that year delivers per dollar, 1.5–1.6× more each year. So inference demand grows about 2.7× in 2028, 2.4× in 2029 and 2.1× in 2030, reaching 57× today’s level by 2030.</p>
<p><strong>Training and research</strong> use about 60% of today’s chips, and the <a href="https://06.ai/packs/ai-compute-demand">compute-demand</a> pack didn’t project them. We let them grow 2.4× in 2027, slowing to 1.6× in 2030, measured in 8-bit FLOPs: 19× today’s by 2030. The anchors: OpenAI’s compute roughly quadrupled in each of 2024 and 2025 (Epoch AI); it ended 2025 with 1.9 GW and targets 30 GW by 2030. Anthropic, at about 2 GW early in 2026, has placed about 5 GW of demand with Broadcom for 2027 and 10 GW for 2028 and could pass 30 GW by 2030 (22V Research). The power behind frontier training runs grows 2.2–2.9× a year (Epoch AI). We assume market-wide training grows more slowly than these labs’ plans: they use less than half of the world’s AI compute.</p>
<h2>The market grows at about 0.6 of OpenRouter’s pace</h2>
<figure><a href="https://06.ai/packs/ai-compute-balance#pace"><img src="https://06.ai/newsletter/ai-compute-balance/fig-05.png" alt="The market grows at about 0.6 of OpenRouter’s pace" width="694" height="266"></a><figcaption>Growth of each disclosed token count, in doublings, divided by OpenRouter’s over the same window (its average for the 30 days before each date). 1 would be OpenRouter’s pace. Disclosures use different definitions, so they compare within a series, not with each other. Two rows are Google’s all-products total over overlapping windows; dropping either moves the median to 0.57 or 0.69.</figcaption></figure>
<h2>What tips the balance</h2>
<figure><a href="https://06.ai/packs/ai-compute-balance#tips"><img src="https://06.ai/newsletter/ai-compute-balance/fig-06.png" alt="What tips the balance" width="694" height="380"></a><figcaption>Supply against demand at the end of the year with one input at its low or high setting and everything else at the base case (demand 20% above supply today). The dot is the base case. Shown: 2030.</figcaption></figure>
<figure><a href="https://06.ai/packs/ai-compute-balance#tips"><img src="https://06.ai/newsletter/ai-compute-balance/fig-07.png" alt="What tips the balance — 2027" width="694" height="411"></a><figcaption><strong>2027.</strong></figcaption></figure>
<h2>What would change the numbers</h2>
<p><strong>Training and research budgets.</strong> The biggest lever. If they grow only 4.5× by 2030, supply ends 2030 43% ahead of demand, a glut. If they grow 66×, the market is 69% short. Much of this is spent by a few labs that are still raising the money.</p>
<p><strong>Power and money.</strong> If capacity added stays near 30 GW a year, 2030 is 43% short. At our power research’s central estimate it is 47% short, and at its high case 24% short. If power follows SemiAnalysis’s path, about 90 GW a year by 2029, with capex to match, supply runs 25% ahead by 2030. Chips aren’t what holds supply back in our base case: if power and money kept up with every chip the industry could deliver, supply would run 17% ahead by 2030.</p>
<p><strong>Inference spending.</strong> If what customers spend on inference grows only 1.2×, 1.1× and 1.0× in 2028–30, the market stays near balance (+5% in 2030). If it grows 2.35×, 2.0× and 1.7×, faster than even OpenAI’s plan, 2030 is 49% short.</p>
<p><strong>Hardware.</strong> If Rubin Ultra slips (SemiAnalysis reported in July that its 144-GPU rack moves to 2028), Feynman disappoints and the 2026–27 mix underdelivers too, 2030 is 43% short. If they beat their roadmaps, it is 11% long. Because cheaper inference also raises demand, hardware moves the balance mostly through training and the years to 2027. Software gains on existing chips help a little: the efficiency research’s central case, software gains of 30% a year against workloads 20% heavier, would leave 2030 18% short instead of 21%.</p>
<p><strong>Who gets the new hardware.</strong> If the newest hardware went to inference first, where its edge is largest, 2030 would be 14% short; if it went to training first, 30% short.</p>
<p><strong>Today’s shortage.</strong> With no shortage today, supply leads by 19% at the end of 2027 and 2030 ends 6% short. If demand runs 30% above supply today, 2027 ends 8% short.</p>
<p><strong>Inference to 2027.</strong> The pace against OpenRouter and inference’s share of today’s chips move 2027 between 12 points up and 13 points down. Letting OpenRouter’s growth fade, as the <a href="https://06.ai/packs/ai-demand">AI-demand</a> pack’s base path does if carried to 2030, would leave 22% more compute than demand by 2030.</p>
<h2>What to watch</h2>
<p><strong>Gigawatts switched on in 2027.</strong> Our base case has about 38 GW. Hyperscalers report capacity added each quarter, and turbine and transformer deliveries lead them. At our power research’s high case, 30 GW, the market ends 2027 15% short.</p>
<p><strong>Capex plans for 2028–30.</strong> Street estimates have AI capex growing about 18% in 2028 and 11% by 2030. If lab revenue grows as planned, that is too little; capex rising faster than that is how the gap would close.</p>
<p><strong>Lab revenue against plan.</strong> OpenAI’s and Anthropic’s revenue, funding rounds and IPOs decide both what customers spend on inference and what labs spend on training, the two largest demand inputs after 2027.</p>
<p><strong>Rubin in the field.</strong> SemiAnalysis measured Rubin at about 2× GB300 per MW on pre-release software. Production results, and when Rubin Ultra ships, set how much each 2027–28 gigawatt is worth.</p>
<p><strong>Rents.</strong> In September the B200 rental curve priced the 12-month forward about 12% below spot: the market expects some easing within a year, consistent with supply catching up during 2027.</p>
<h2>What this is anchored on</h2>
<p>Supply, power, capex and demand paths, efficiency by year, the shortage today and the balance are our estimates, built from the sources below. OpenRouter data is licensed under CC BY 4.0.</p>
<ul><li><strong>Today’s fleet and demand</strong> — AI chips operating, their power, the inference share, AgentX throughput by system and OpenRouter’s compute paths to December 2027 · <a href="https://06.ai/packs/ai-compute-demand">06.ai </a><a href="https://06.ai/packs/ai-compute-demand">compute-demand</a><a href="https://06.ai/packs/ai-compute-demand"> pack</a></li><li><strong>OpenRouter paths</strong> — Tokens and spend to December 2027, with growth halving every 3, 6 or 12 months · <a href="https://06.ai/packs/ai-demand">06.ai </a><a href="https://06.ai/packs/ai-demand">AI-demand</a><a href="https://06.ai/packs/ai-demand"> pack</a></li><li><strong>OpenRouter tokens</strong> — Daily tokens by model, January 2025 to October 2026, for the pace comparison · <a href="https://openrouter.ai/rankings">OpenRouter rankings dataset (CC BY 4.0)</a></li><li><strong>Chips delivered</strong> — 20.3M H100-equivalents delivered by end-2025; 4.98M and 6.33M in the first two quarters of 2026, rated at 2.7 and 3.2 GW · <a href="https://epoch.ai/data/ai-chip-sales">Epoch AI chip-sales data (via </a><a href="https://epoch.ai/data/ai-chip-sales">the </a><a href="https://epoch.ai/data/ai-chip-sales">compute-demand</a><a href="https://epoch.ai/data/ai-chip-sales"> pack</a><a href="https://epoch.ai/data/ai-chip-sales">)</a></li><li><strong>Chips by designer</strong> — Q2 2026 share of compute sold: NVIDIA 63%, Google 25%, Amazon 5%, AMD 4%, Huawei 3% · <a href="https://epoch.ai/graphs/ai-chip-market-share-by-designer">Epoch AI, updated Oct 19</a></li><li><strong>Chip ratings</strong> — Vera Rubin 8.84 H100-equivalents per chip; TPU v8i 5.1; B300 2.53 · <a href="https://epoch.ai/data/ai_chip_sales_chip_types.csv">Epoch AI chip-types data, Oct 14</a></li><li><strong>Facility power</strong> — A data center draws about twice its AI chips’ rated power · <a href="https://epoch.ai/data/ai-chip-sales?tab=power">Epoch AI chip-sales data</a></li><li><strong>NVIDIA</strong> — A supply-constrained outlook of about 70% revenue growth in the year to January 2028; supply the bottleneck at least through then · <a href="https://www.cfodive.com/news/nvidia-cfo-predicts-70-percent-revenue-jump-fiscal-2028-ai/828989/">CFO Dive, Aug 27</a></li><li><strong>NVIDIA</strong> — At least $1T of Blackwell and Rubin demand and orders through 2027 · <a href="https://www.datacenterdynamics.com/en/news/nvidia-ceo-jensen-huang-says-company-has-one-trillion-dollars-in-orders-through-2027">DCD, GTC, Mar 16</a></li><li><strong>NVIDIA</strong> — A GW costs about $50–60B; NVIDIA’s content per GW about $18B for Hopper, $25B for Blackwell, $40B for Vera Rubin; supply supports about 70% revenue growth in fiscal 2028, against demand growth nearer 100%; top-5 hyperscaler capex near $800B in 2026 and $1.3T in 2027 · <a href="https://www.webull.com/news/15497781742806016">NVIDIA Q2 FY27 call, Aug 26</a></li><li><strong>Broadcom</strong> — AI revenue outlook about $115B in fiscal 2027 and $230B in fiscal 2028 · <a href="https://www.thestar.com.my/business/business-news/2026/09/04/broadcom-forecasts-explosive-ai-chip-growth">Bloomberg, via The Star, Sep 4</a></li><li><strong>TSMC</strong> — AI demand still exceeds what TSMC can supply · <a href="https://letsdatascience.com/news/tsmc-says-ai-demand-still-exceeds-its-supply-capability-as-q-2fed75bf">TSMC Q3 call, via AP, Oct 15</a></li><li><strong>Packaging</strong> — TSMC CoWoS capacity about 130K wafers a month at end-2026, doubling by 2028 · <a href="https://www.trendforce.com/news/2026/09/14/news-tsmc-reportedly-targets-22-2nm-16-3nm-capacity-boost-by-mid-2027-cowos-to-double-by-2028/">TrendForce, Sep 14</a></li><li><strong>Memory</strong> — HBM bit shipments up 50–60% in 2027, still short of demand · <a href="https://trendforce.com/presscenter/news/20260804-13166.html">TrendForce, Aug 4</a></li><li><strong>Google TPUs</strong> — TPU shipments estimated at 4.5M in 2026 and 8.8M in 2027 · <a href="https://247wallst.com/investing/2026/08/25/google-tpu-volume-could-triple-to-8-8-million-by-2027/">GF Securities, via 24/7 Wall St, Aug 25</a></li><li><strong>AMD</strong> — Helios racks shipping from late Q3 2026; 6 GW for OpenAI, 6 GW for Meta and up to 2 GW for Anthropic · <a href="https://convergedigest.com/?p=113861">Converge Digest, Jul 24</a></li><li><strong>Power</strong> — Hyperscalers and neoclouds add 32 GW in 2027 and 39 GW in 2028, 60–95% of new capacity worldwide · <a href="https://www.webull.com/news/15695876726391808">Morgan Stanley, via Zhitong, Oct 5</a></li><li><strong>Power</strong> — US data-center capacity scheduled for 2026 13.6 GW and 2027 36.3 GW; only 50–60% assumed on time · <a href="https://oilprice.com/Energy/Energy-General/The-Gas-Turbine-Shortage-Just-Became-AIs-Biggest-Constraint.amp.html">Goldman Sachs, via OilPrice, Aug 22</a></li><li><strong>Power</strong> — About 30 GW of AI compute added in 2026, 50 GW in 2027, 70 GW in 2028 and 90–100 GW in 2029 · <a href="https://www.theneuron.ai/news/ais-11t-bet-the-bull-and-bear-case-for-the-compute-boom/">Dylan Patel (SemiAnalysis), via The Neuron, Aug 31</a></li><li><strong>Power</strong> — A consensus estimate that about 15 GW of AI compute made in 2027 can’t be switched on in 2027 · <a href="https://247wallst.com/cards/xpost-01m17pydcft8z890gj36hmt3t9">Elon Musk, via 24/7 Wall St, Aug 29</a></li><li><strong>Power</strong> — At least 16 GW of large projects slated for 2026 worldwide, about 5 GW under construction in February; 30–50% unlikely to finish on time · <a href="https://www.latitudemedia.com/news/up-to-half-of-the-worlds-data-centers-may-be-delayed-this-year/">Sightline Climate, via Latitude Media, Feb 24</a></li><li><strong>Power</strong> — More than 60% of US projects due in 2027 not under construction in April · <a href="https://longbridge.com/news/283103823">SynMax, via the FT (Longbridge), Apr 17</a></li><li><strong>Equipment</strong> — A heavy-duty GE Vernova gas turbine ordered now arrives in 2031; Mitsubishi Heavy’s new orders deliver in 2028–30 · <a href="https://oilprice.com/Energy/Energy-General/The-Gas-Turbine-Shortage-Just-Became-AIs-Biggest-Constraint.amp.html">OilPrice, Aug 22</a></li><li><strong>Equipment</strong> — Generator step-up transformers take more than 160 weeks; one utility is buying about five years ahead · <a href="https://www.kitco.com/news/off-the-wire/2026-07-09/us-power-companies-scramble-secure-equipment-surging-data-center">Wood Mackenzie, via Reuters, Jul 9</a></li><li><strong>Capex</strong> — Global AI investment about $1T in 2026 · <a href="https://www.goldmansachs.com/insights/articles/global-investment-is-forecast-to-exceed-1-trillion-in-2026">Goldman Sachs, Aug 7</a></li><li><strong>Capex</strong> — Alphabet, Microsoft, Amazon, Meta, Oracle and SpaceX: about $800B in 2026, $1.2T in 2027 and $1.4T in 2028 · <a href="https://cryptobriefing.com/goldman-sachs-hyperscaler-capex-ai-revenue/">Goldman Sachs, via Crypto Briefing, Sep 25</a></li><li><strong>Capex</strong> — Hyperscaler, Oracle, SpaceX and neocloud capex to 2030 · <a href="https://06.ai/packs/ai-funding-scale">06.ai </a><a href="https://06.ai/packs/ai-funding-scale">hyperscaler</a><a href="https://06.ai/packs/ai-funding-scale">, </a><a href="https://06.ai/packs/ai-funding-scale">neocloud</a><a href="https://06.ai/packs/ai-funding-scale"> and </a><a href="https://06.ai/packs/ai-funding-scale">leveraged-middle</a><a href="https://06.ai/packs/ai-funding-scale"> packs</a></li><li><strong>Cost per GW</strong> — $38B per GW of IT power for a GB200 NVL72 site · <a href="https://epoch.ai/data-insights/ai-datacenter-cost-breakdown">Epoch AI, May 14</a></li><li><strong>Cost per GW</strong> — $34.6–39.5B per GW by architecture, from OpenAI’s custom chip to Vera Rubin · <a href="https://phemex.com/news/article/bernstein-estimates-ai-data-center-costs-at-346b395b-per-gigawatt-99408">Bernstein, via Phemex, Oct 10</a></li><li><strong>Inference per MW</strong> — Vera Rubin NVL72 59.4M tokens/s per MW on AgentX at 100 tokens/s per user, 2.09× GB300 NVL72 (pre-release software) · <a href="https://inferencex.semianalysis.com/blog/vera-rubin-nvl72-agentic-inference">SemiAnalysis InferenceX, Sep 14</a></li><li><strong>Inference per MW</strong> — Vera Rubin about 2× GB300 per MW up to 100 tokens/s per user on DeepSeek-R1 · <a href="https://newsletter.semianalysis.com/p/vera-rubin-nvl72-vs-gb200-nvl72-inference">SemiAnalysis, Jul 23</a></li><li><strong>Rubin</strong> — 35 PF dense NVFP4 and 17.5 PF dense FP8 per GPU · <a href="https://www.nvidia.com/en-us/data-center/vera-rubin-nvl72/">NVIDIA Vera Rubin NVL72 page</a></li><li><strong>Rubin Ultra</strong> — Kyber NVL144 rack reported slipping more than a year, to 2028 · <a href="https://mlq.ai/news/nvidias-kyber-nvl144-rack-system-delayed-over-12-months-to-2028-on-pcb-manufacturing-snags/">SemiAnalysis, via MLQ, Jul 6</a></li><li><strong>Hardware trends</strong> — Chip performance per dollar up about 49% a year since 2023 · <a href="https://epoch.ai/data-insights/chip-performance-per-dollar">Epoch AI, Aug 13</a></li><li><strong>Hardware trends</strong> — Peak FLOP/s per watt of ML GPUs up about 1.28× a year · <a href="https://epoch.ai/publications/can-ai-scaling-continue-through-2030">Epoch AI, Aug 2024</a></li><li><strong>Tokens</strong> — Google: 480T tokens a month in May 2025, more than 3.2 quadrillion in May 2026 · <a href="https://blog.google/innovation-and-ai/sundar-pichai-io-2026/">Google I/O, May 19</a></li><li><strong>Tokens</strong> — Google: about 1.3 quadrillion tokens a month in October 2025 · <a href="https://robonomics.substack.com/p/token-tracker-and-implications">Google, via Robonomics</a></li><li><strong>Tokens</strong> — Google’s models: 7B tokens a minute via direct API use in October 2025 · <a href="https://s206.q4cdn.com/479360582/files/doc_events/2025/Oct/29/2025_Q3_Earnings_Transcript.pdf">Alphabet Q3 2025 call, Oct 29, 2025</a></li><li><strong>Tokens</strong> — Google’s model APIs: about 22B tokens a minute in July 2026 · <a href="https://www.marketbeat.com/instant-alerts/alphabet-q2-earnings-call-highlights-2026-07-22/">Alphabet Q2 call, via MarketBeat, Jul 22</a></li><li><strong>Tokens</strong> — OpenAI’s API: 6B tokens a minute in October 2025 · <a href="https://techcrunch.com/2025/10/06/sam-altman-says-chatgpt-has-hit-800m-weekly-active-users">TechCrunch, Oct 6, 2025</a></li><li><strong>Tokens</strong> — OpenAI’s APIs: more than 15B tokens a minute as of March 31 · <a href="https://finance.yahoo.com/markets/stocks/articles/openai-closes-122-billion-funding-010940572.html">Quartz, via Yahoo Finance, Mar 31</a></li><li><strong>Tokens</strong> — Doubao: more than 30T tokens a day in September 2025 · <a href="https://kr-asia.com/pulses/157843">Volcano Engine, via KrASIA, Oct 21, 2025</a></li><li><strong>Tokens</strong> — Doubao: more than 180T tokens a day in June 2026 · <a href="https://www.diariobitcoin.com/china/volcano-engine-lanzara-seedance-2-5-en-julio-y-presume-liderazgo-en-maas-en-china/">Volcano Engine, via Diario Bitcoin, Jun 23</a></li><li><strong>Tokens</strong> — China: about 100T tokens a day at end-2025, more than 140T in March 2026 · <a href="https://www.chinadaily.com.cn/a/202603/25/WS69c2bcd9a310d6866eb3fa66.html">National Data Administration, via China Daily, Mar 25</a></li><li><strong>Lab revenue</strong> — OpenAI plans revenue of $36B in 2026 rising to $350B in 2030, and about $856B on compute and infrastructure · <a href="https://kfgo.com/2026/09/18/openai-expects-to-burn-through-almost-280-billion-by-2030-ft-reports/">FT, via Reuters, Sep 18</a></li><li><strong>Lab revenue</strong> — Anthropic’s internal forecast of about $190–200B of revenue in 2028 · <a href="https://thenextweb.com/news/anthropic-ipo-190-200bn-2028-revenue-forecast">Reuters, via The Next Web, Aug 17</a></li><li><strong>Lab compute</strong> — OpenAI ended 2025 with about 1.9 GW of compute, more than triple a year earlier · <a href="https://datacenterdynamics.com/en/news/openai-cfo-says-company-ended-2025-with-19gw-of-compute-scaled-revenue-at-same-speed/">DCD, Jan 21</a></li><li><strong>Lab compute</strong> — OpenAI targets 30 GW by 2030 · <a href="https://yourstory.com/ai-story/openai-30gw-ai-compute-plan-2030">YourStory, Apr 23</a></li><li><strong>Lab compute</strong> — Anthropic at about 2 GW early in 2026; about 5 GW of demand placed with Broadcom for 2027 and 10 GW for 2028; over 30 GW by 2030 · <a href="https://22vresearch.com/2026/09/04/anthropic-leadership-and-scale-thoughts-ahead-of-the-s-1-and-ipo/">22V Research, Sep 4</a></li><li><strong>Lab compute</strong> — OpenAI’s compute about 0.39M H100-equivalents in 2024 and 1.74M in 2025 · <a href="https://epoch.ai/latest/introducing-the-ai-chip-users-explorer">Epoch AI, Sep 9</a></li><li><strong>Training</strong> — The power drawn by frontier training runs grows 2.2–2.9× a year; the largest could draw 4–16 GW by 2030 · <a href="https://epoch.ai/publications/power-demands-of-frontier-ai-training">Epoch AI and EPRI, Aug 2025</a></li><li><strong>Shortage</strong> — H100 one-year rental contracts from about $1.70 an hour in October 2025 to $2.35 in March 2026 · <a href="https://newsletter.semianalysis.com/p/the-great-gpu-shortage-rental-capacity">SemiAnalysis, Apr 2</a></li><li><strong>Shortage</strong> — B200 rental index up about 28% this year to September 1; 12-month forward about 12% below spot in September · <a href="https://www.silicondata.com/blog/b200-rental-price-august-2026-update">Silicon Data</a></li><li><strong>Shortage</strong> — Not enough capacity to meet demand in 2026, likely the same in 2027 · <a href="https://www.fool.com/earnings/call-transcripts/2026/08/07/amazon-amzn-q2-2026-earnings-call-transcript/">Amazon Q2 call, via The Motley Fool, Jul 30</a></li><li><strong>Shortage</strong> — Supply and demand rebalance in the second half of 2027 at the earliest, the first half of 2028 more likely · <a href="https://www.kucoin.com/blog/th-ai-infrastructure-shortage-may-last-until-2028-as-hyperscaler-capex-keeps-rising">Goldman Sachs, via KuCoin, Sep 15</a></li></ul>
<h2>Method</h2>
<p><strong>Units and clock.</strong> Supply and demand are both counted in gigawatts of today’s hardware: the work one gigawatt of the fleet operating in September 2026 does, for inference (AgentX tokens, the <a href="https://06.ai/packs/ai-compute-demand">compute-demand</a> pack’s measure) or for training (peak 8-bit FLOP/s, Epoch AI’s H100-equivalents). Today’s fleet is 31.6M H100-equivalents, about 40 GW all-in, 40% of it serving inference. A gigawatt installed later counts for its multiple of today’s average. Today is October 1, 2026: the fleet is the chips delivered to June 2026, demand is measured from October 1 (OpenRouter’s October figure covers October 1–18), and every balance is at December 31. The balance is the largest share of both demands the installed capacity can serve at once, with every year’s hardware split between inference and training in the same proportion. All power figures are all-in, at the facility, about 1.3× IT power; several forecasts we cite don’t say which they count.</p>
<p><strong>Chips.</strong> Our estimate of what the designers Epoch tracks could deliver: 33, 100, 170, 250, 350M H100-equivalents in 2026–30 (range 185–550M in 2030), rated at 15, 26, 34, 38, 41 GW. 2026 builds on Epoch’s first half and NVIDIA’s guided second half; 2027 on NVIDIA’s guidance, Broadcom’s outlook, TPU and AMD unit estimates and CoWoS and HBM capacity; 2028–30 on accelerator spending of $1T, $1.15T, $1.25T and about 30% more H100-equivalents per dollar each year. Chips switch on a quarter after delivery, at twice their rated power: 30% of a year’s deliveries come in its fourth quarter and switch on the next year. We split 2026’s second half between the third and fourth quarters 9:13, as for its H100-equivalents; the fourth quarter’s heavier Rubin mix would put a little more power in the third.</p>
<p><strong>Power and money.</strong> Power: 38, 46, 52, 58 GW in 2027–30, our estimate from analysts’ forecasts put on an all-in, worldwide basis (Morgan Stanley, Goldman). Our power research’s own synthesis is lower: 23, 28, 31, 35 GW, high case 30, 40, 50, 60 (sensitivities). Low 30, 33, 35, 37; high 50, 70, 95, 105 (SemiAnalysis to 2029; 2030 ours). Money: AI capex of $1T, $1.4T, $1.65T, $1.85T, $2.05T in 2026–30, from the street capex in the <a href="https://06.ai/packs/hyperscaler-capex-funding">hyperscaler</a>, <a href="https://06.ai/packs/neocloud-capex-funding">neocloud</a> and <a href="https://06.ai/packs/leveraged-middle">leveraged-middle</a> packs plus China’s largest platforms, sovereigns and others (86–88% AI), divided by $45, 48, 50, 51, 52B per GW of IT power and multiplied by 1.3 for facility power. The money includes the chips. What comes online each year is the lowest of chips, power and money; chips beyond that aren’t counted.</p>
<p><strong>Each year’s hardware.</strong> Inference: AgentX tokens per MW for each year’s deliveries, 0.8, 1.7, 2.7, 4.5, 6.9× a GB300 NVL72 rack in 2026–30, our estimate anchored on SemiAnalysis’s measurements (Vera Rubin 2.09×) and roadmaps; low 0.55, 1.2, 1.8, 2.7, 3.8; high 1, 2.5, 4.4, 7.8, 13. Divided by today’s fleet average, 0.33× GB300. Training: H100-equivalents per GW from the chips’ rated power, against today’s 625 W per H100-equivalent; low and high from our range for training gains per MW. Capacity switched on in a year blends the previous year’s fourth-quarter deliveries with its own first three quarters, by power. Our efficiency research labels its tables by install year; we read them as deliveries (a sensitivity). Software stays at today’s level: we assume its gains offset heavier models and longer contexts (the efficiency research’s central case nets about 8% a year; a sensitivity).</p>
<p><strong>Demand.</strong> Inference to December 2027: the <a href="https://06.ai/packs/ai-compute-demand">compute-demand</a> pack’s OpenRouter compute paths raised to the power 0.6, the market’s pace (calibrated above; 0.45–0.75 as a sensitivity), read at December 31. After 2027: what customers spend on inference grows 1.7, 1.5, 1.35× a year in the base case (1.2, 1.1, 1 low; 2.35, 2, 1.7 high), times the gain in inference work per dollar of the capacity switched on that year (1.58, 1.60, 1.52× in the base case, from the same hardware and supply cases). Training and research: 2.7, 2.4, 2.1, 1.8, 1.6× a year in 2026–30 (2.2, 1.7, 1.4, 1.3, 1.2 low; 3.2, 3.2, 2.8, 2.5, 2.2 high), the first applied to October–December only. Today’s shortage, demand 20% above supply (0–30%), scales demand in every year.</p>
<p><strong>Needed GW</strong> is the capacity of that year’s hardware that, added to everything installed before, makes supply equal demand at year-end. Utilization stays at 60% throughout, as in the <a href="https://06.ai/packs/ai-compute-demand">compute-demand</a> pack; the table gives the utilization at which supply would just meet demand. We keep today’s chips in service and don’t model refreshes: replacing Hopper inside existing buildings would add supply without new power, though not without money. Disclosed token counts use different definitions; each is compared only with itself over time.</p>
<hr>
<p>Every chart above links to its interactive version on 06.ai, with the data behind it. The daily trackers are at <a href="https://06.ai/trackers">06.ai/trackers</a>.</p>
<p><em>Not investment advice. Estimates are labeled as estimates.</em></p>]]></content:encoded>
</item>
<item>
  <title>How much compute does AI demand need by 2027?</title>
  <link>https://inthesingularity.ai/packs/ai-compute-demand</link>
  <guid isPermaLink="true">https://inthesingularity.ai/packs/ai-compute-demand</guid>
  <pubDate>Mon, 19 Oct 2026 12:09:00 GMT</pubDate>
  <description>The AI-demand pack projects OpenRouter’s tokens to the end of 2027. We turn them into GPUs and gigawatts with benchmarks for each chip and model, then scale the answer up to the whole market.</description>
  <category>AI compute</category>
  <enclosure url="https://inthesingularity.ai/newsletter/ai-compute-demand/cover.png" length="71641" type="image/png" />
  <media:content url="https://inthesingularity.ai/newsletter/ai-compute-demand/cover.png" medium="image" type="image/png" />
  <content:encoded><![CDATA[<figure><a href="https://06.ai/packs/ai-compute-demand"><img src="https://06.ai/newsletter/ai-compute-demand/thumb.png" alt="How much compute does AI demand need by 2027?" width="421" height="432"></a><figcaption><a href="https://06.ai/packs/ai-compute-demand">Interactive version and data on 06.ai →</a></figcaption></figure>
<p>OpenRouter’s 23.8T tokens a day would need about <strong>60 MW</strong> of today’s hardware at benchmark throughput. In <a href="https://06.ai/packs/ai-demand">the </a><a href="https://06.ai/packs/ai-demand">AI-demand</a><a href="https://06.ai/packs/ai-demand"> pack</a>’s base case its tokens grow 22× by December 2027, but at today’s prices the compute they need grows only about <strong>10×</strong>, because cheap models keep taking share. If the whole market’s inference demand grew at that pace, about 7.2× a year, it would need about <strong>47 GW of new GB300-class capacity</strong> by the end of 2027, or 22 GW if it were all Vera Rubin. The AI chips sold in all of 2026 add roughly 10–11 GW for inference.</p>
<ul><li><strong>60 MW</strong> — OpenRouter today all-in, on today’s mix of hardware at benchmark throughput. On GB300 racks: 20 MW, about 9,300 GPUs</li><li><strong>10×</strong> — OpenRouter’s compute, Dec 2027 today’s, for 22× the tokens and 4.2× the dollars in the <a href="https://06.ai/packs/ai-demand">AI-demand</a> pack’s base case. About 195 MW of GB300s</li><li><strong>+47 GW</strong> — The market at OpenRouter’s pace of new GB300-class inference capacity by Dec 2027, about 22 million GPUs, 43 GW of it during 2027. +11 GW at half the pace</li><li><strong>10–11 GW</strong> — 2026’s new chips, for inference of 25–28 GW added in all, on mixed hardware: 24–27M H100-equivalents, more than the 20.3M delivered in all years through 2025 (our estimate from Epoch AI data)</li></ul>
<h2>A token isn’t a token</h2>
<figure><a href="https://06.ai/packs/ai-compute-demand#tokens"><img src="https://06.ai/newsletter/ai-compute-demand/fig-01.png" alt="A token isn’t a token" width="694" height="610"></a><figcaption>Left: tokens a day from one megawatt of all-in power for each system serving DeepSeek V4 Pro on SemiAnalysis’s AgentX benchmark, which replays real coding-agent sessions, at 100 tokens/s per user. As on OpenRouter, most of the tokens are cached input. Values at 100 tokens/s are SemiAnalysis’s interpolations between measured points. GB200 is our estimate. B300 ran vLLM, Vera Rubin pre-release TensorRT-LLM and the others SGLang. Right: tokens a day from one B200 by model, with 8K-token prompts and 1K-token answers at about 50 tokens/s per user.</figcaption></figure>
<h2>Four things set what a token costs</h2>
<p><strong>The model.</strong> Compute per token rises with the parameters a model uses for each token more than with its total size, though total size and the length of the conversation matter too. On one B200, gpt-oss-120b, which uses 5.1B parameters per token, makes 3.8B tokens a day. DeepSeek-R1, at 37B, makes 0.40B. MiniMax-M2.5 uses a quarter of GLM-5’s active parameters and makes 3.9× its tokens. The rule is rough: Llama 3.3 70B, which uses all 70B of its parameters for every token, makes more than DeepSeek-R1. We size every model by its active parameters.</p>
<p><strong>The hardware.</strong> Serving DeepSeek V4 Pro to coding agents at 100 tokens/s per user, a megawatt of GB300 NVL72 racks serves 13× the tokens of a megawatt of H200s running at 8-bit precision, and 4.1× those of B200s. Vera Rubin, now shipping, doubles GB300 again. Racks of 72 GPUs win on large models because they spread the model’s experts and its memory of the conversation across many chips. At full load, a gigawatt of GB300s would get through about 100 days of OpenRouter’s October volume each day, if every token were a DeepSeek V4 Pro agent token.</p>
<p><strong>The kind of token.</strong> Output is 2.4% of OpenRouter’s tokens but about 34% of the GPU time behind them. New input is 19% of the tokens and 55% of the time. Cached input, the context agents send again at every step, is 79% of the tokens and about 11% of the time. These weights follow the labs’ own pricing: output costs about five times new input, and a cache read a tenth to a twentieth of it; we use a twentieth.</p>
<p><strong>Speed and memory.</strong> Serving each user faster means smaller batches. DeepSeek-R1 on a B200 makes 6,799 tokens/s per GPU at 30 tokens/s per user and 2,685 at 100. A model that doesn’t fit on one server is worse: at 50 tokens/s per user, DeepSeek-R1’s best H100 setup made 74 tokens/s per GPU, against 4,610 on a B200.</p>
<h2>On OpenRouter, 12% of the tokens use about half the compute</h2>
<figure><a href="https://06.ai/packs/ai-compute-demand#mix"><img src="https://06.ai/newsletter/ai-compute-demand/fig-02.png" alt="On OpenRouter, 12% of the tokens use about half the compute" width="694" height="208"></a><figcaption>Share of OpenRouter’s tokens, of the compute they need (our estimate) and of spend (the <a href="https://06.ai/packs/ai-demand">AI-demand</a> pack’s estimate), October 1–18. Models are grouped by whether their weights are published and by parameters used per token. Closed models’ sizes are our estimates from their prices; see Method. Free versions count in tokens and compute but not in dollars.</figcaption></figure>
<h2>OpenRouter in GPUs</h2>
<p>At SemiAnalysis’s measured throughput and 60% utilization, October’s 23.8T tokens a day would need about 20 MW of GB300 racks, about 9,300 GPUs. On H200s they would need 250 MW. On a mix of hardware like the industry’s installed base today, part Hopper, part Blackwell, part TPU and other chips, they would need about 60 MW. Every megawatt here is all-in: chips, servers, networking and cooling. Real deployments run older software than a tuned benchmark, which pushes the need up; how closely the benchmark’s traffic matches OpenRouter’s moves it either way (see Method).</p>
<p>The compute isn’t where the tokens are. Open models that use 20B parameters or fewer per token, led by DeepSeek V4.1 Flash with 21% of all tokens, are 50% of the tokens, 17% of the compute and 4% of the dollars. Closed models above that size, among them Claude Opus 5.5 and Sonnet 5.5, GPT-6 Sol and Astra and Gemini 3.8 Flash, are 12% of the tokens, 64% of the dollars and, by our price-based estimate, 49% of the compute. Opus 5.5 alone is 3% of the tokens and about 18% of the compute. Those compute shares rest on what we assume about the labs’ margins: at gross margins of 45–75% instead of 60%, the larger closed models are 38–56% of the compute.</p>
<h2>Tokens 22×, compute 10×, dollars 4.2×</h2>
<figure><a href="https://06.ai/packs/ai-compute-demand#path"><img src="https://06.ai/newsletter/ai-compute-demand/fig-03.png" alt="Tokens 22×, compute 10×, dollars 4.2×" width="694" height="402"></a><figcaption>OpenRouter by month, October 2026 = 1, log scale. Tokens and dollars are the <a href="https://06.ai/packs/ai-demand">AI-demand</a> pack’s paths; growth halves every 3, 6 or 12 months. Compute is our estimate at today’s efficiency, holding each model’s price fixed: the model mix that reproduces the <a href="https://06.ai/packs/ai-demand">AI-demand</a> pack’s tokens and dollars each month, times each model’s compute per token. Shown: Base.</figcaption></figure>
<h2>OpenRouter in 2027</h2>
<p>In the base case OpenRouter handles 513T tokens a day by December 2027, 22× October’s, while spend grows only 4.2×, so the average price of a token falls 5.1×. Holding each model’s price fixed, as we do, that requires cheaper models to keep taking share: in the mix that reproduces the <a href="https://06.ai/packs/ai-demand">AI-demand</a> pack’s numbers, DeepSeek V4.1 Flash’s tokens grow 48× and Claude Opus 5.5’s 1.5×. Because the cheap tokens need little compute each, compute grows 10×, from 20 MW of GB300s to 195 MW, about 84,000 more GPUs. On today’s mix of hardware it would be 600 MW. If prices fall model by model instead, as GPT-5.6 Luna’s did when OpenAI cut them 80% this summer, compute grows closer to tokens.</p>
<p>The low path needs 4.1× today’s compute (80 MW of GB300s) and the high path 23× (460 MW). Even the high case is under half a gigawatt of GB300s. OpenRouter on its own is a rounding error in a year that adds tens of gigawatts. What matters is what it says about everyone else.</p>
<h2>If the whole market grew at OpenRouter’s pace</h2>
<p>New inference capacity the whole market needs, cumulative from October 2026, in GW of all-in power, if its inference demand grows at a share of OpenRouter’s projected compute growth in the <a href="https://06.ai/packs/ai-demand">AI-demand</a> pack’s base case. It starts from about 16 GW of chips doing inference today and assumes the new capacity is all GB300 or all Vera Rubin. The band is our estimate of the inference share of the chips added in 2026, a mix of hardware that serves fewer tokens per MW than GB300 racks.</p>
<p><em>At the default setting (Market grows at of OpenRouter’s projected monthly rate 100%), from the </em><a href="https://06.ai/packs/ai-compute-demand#market"><em>interactive version</em></a><em>:</em></p>
<ul><li>New capacity by Dec 2027: <strong>+47 GW</strong></li><li>GPUs: <strong>22 million</strong></li><li>In 2027 alone: <strong>+43 GW</strong></li><li>Market inference demand vs today: <strong>10×</strong></li></ul>
<figure><a href="https://06.ai/packs/ai-compute-demand#market"><img src="https://06.ai/newsletter/ai-compute-demand/fig-04.png" alt="If the whole market grew at OpenRouter’s pace" width="694" height="402"></a><figcaption>Shown: GB300.</figcaption></figure>
<figure><a href="https://06.ai/packs/ai-compute-demand#market"><img src="https://06.ai/newsletter/ai-compute-demand/fig-05.png" alt="If the whole market grew at OpenRouter’s pace — Vera Rubin" width="694" height="433"></a><figcaption><strong>Vera Rubin.</strong></figcaption></figure>
<h2>From OpenRouter to the market</h2>
<p>OpenRouter is a small slice. On matching dates its tokens were about 3% of Google’s across all its products, 1.3% of a national total for China that TechNode reported for June without a source or method, and 0.6–2.2% of published estimates for the whole world. Its share of the compute is smaller still, about 0.4% of the chips doing inference by our estimate. Partly that is because its tokens are cheap: half come from small open models that coding agents call over and over. Partly it is because benchmarks flatter real deployments: at benchmark throughput, today’s inference chips could serve about 6.2 quadrillion OpenRouter-like tokens a day, 6.3× the highest estimate for the world.</p>
<p>So we scale up from the hardware, not the tokens, and use the benchmarks only to compare one kind of hardware with another. Epoch AI estimates 31.6M H100-equivalents of AI chips operating by September: chips delivered through June, allowing about three months to install. At about 1.25 kW each all-in, that is 39.5 GW. If 40% of it serves inference, that is 16 GW. If the market’s inference demand grows the way OpenRouter’s compute does in the base case, 10× by December 2027, it needs about 160 GW of hardware like today’s. Built as GB300s, which serve about 3.1× as many tokens per MW as today’s average, the increase is about <strong>47 GW, or 22 million GPUs</strong>. Built as Vera Rubin, it is about 22 GW. Almost all of it falls in 2027: 43 GW of GB300s between December 2026 and December 2027.</p>
<p>That takes the premise at face value, and the premise is aggressive. At full pace the market’s inference compute grows about 7.2× a year, roughly what Google’s tokens did in the year to May; over the same year OpenRouter’s tokens grew 12.5×. The installed base of AI chips grew 3.0× in 2025. At half OpenRouter’s growth rate, about 2.7× a year, close to the chip stock’s growth, market demand rises 3.2× by December 2027 and needs 11 GW of new GB300s. Whether the answer is closer to 11 GW or 47 GW depends on whether the rest of the market speeds up toward OpenRouter’s pace.</p>
<p>Three inputs we can’t pin down move the 47 GW: the share of today’s chips doing inference (30–50% gives 35–58 GW), how efficient today’s fleet is (0.75–1.3× our estimate gives 35–60 GW) and how many chips are installed (Epoch’s range, roughly 1.5× either way, gives 31–70 GW). Moving all three together gives 17–113 GW.</p>
<h2>2026 is on track to add more AI compute than all earlier years combined</h2>
<figure><a href="https://06.ai/packs/ai-compute-demand#added"><img src="https://06.ai/newsletter/ai-compute-demand/fig-06.png" alt="2026 is on track to add more AI compute than all earlier years combined" width="694" height="342"></a><figcaption>AI chips sold, millions of H100-equivalents (Epoch AI’s unit, scaled by 8-bit compute). The second half of 2026 is our range: 10–40% above the first half. Epoch counts chips when they are delivered; installing them takes longer.</figcaption></figure>
<h2>What 2026 added</h2>
<p>Epoch AI’s chip-sales data counts 11.3M H100-equivalents sold in the first half of 2026, against 13.6M in all of 2025. If the second half runs 10–40% above the first, 2026 adds 24–27M, more than the 20.3M delivered in all the years through 2025. Those chips are rated at 12–14 GW. By Epoch’s rule of thumb a data center needs about twice its chips’ rated power once servers, networking and cooling are counted, so 2026’s chips need 25–28 GW all-in. NVIDIA’s data-center revenue was $164B in the first half of its fiscal 2027, against $194B in all of fiscal 2026, and Vera Rubin began production shipments in August. NVIDIA expects supply to stay the bottleneck at least through January 2028.</p>
<p>Not all of it serves inference. At 40%, 2026 gives inference about 10–11 GW, on a mix of hardware that serves fewer tokens per MW than GB300 racks. Demand at OpenRouter’s pace would need about 4.1× that in 2027 alone if it is built as GB300s, or 1.9× as Vera Rubin. At half the pace, 9.7 GW of GB300s in 2027 is about the inference share of a 2026-sized year, and more if built from 2026’s mix of chips. This pack is about demand; a later one will look at how much capacity 2027 can actually deliver.</p>
<h2>What would change the numbers</h2>
<p><strong>The pace.</strong> It is the biggest lever. From a quarter of OpenRouter’s growth rate to all of it, the need by December 2027 runs from 4.0 GW to 47 GW of GB300s.</p>
<p><strong>The mix.</strong> Our compute grows less than tokens because, at fixed prices, cheap models keep taking share. If instead the mix froze and every model’s price fell, compute would grow with tokens, 22×, and the market would need about 106 GW. Two other choices matter less. If the larger closed models’ tokens grew 3.5×, as fast as OpenAI’s and Anthropic’s revenue run-rates in <a href="https://06.ai/packs/ai-lab-revenue">the </a><a href="https://06.ai/packs/ai-lab-revenue">lab-revenue</a><a href="https://06.ai/packs/ai-lab-revenue"> pack</a>, the market would need 45 GW. If free tokens stayed flat instead of growing with the rest, 42 GW.</p>
<p><strong>How big the closed models are.</strong> Nobody discloses them, so we infer their size from price, assuming the labs keep 60% of it as gross margin. At 45% the models are bigger, take more of today’s compute and grow more slowly: 41 GW. At 75%, 55 GW.</p>
<p><strong>How compute scales with size.</strong> We make compute per token proportional to active parameters. Across InferenceX’s six B200 results it rises more slowly, roughly with active parameters to the power 0.89. On that basis small models cost relatively more, the shift toward them saves less, and the market would need 53 GW.</p>
<p><strong>Software.</strong> Benchmarks move every month. InferenceX has measured gains of 1.5–2× from software alone on the same chips within weeks, on DeepSeek-R1 and on DeepSeek V4 Pro. Each doubling roughly halves the need; we hold today’s software fixed.</p>
<p><strong>Speed.</strong> We measure at 100 tokens/s per user. Serving agents faster costs throughput steeply: on DeepSeek V4 Pro, a GB300 rack serves about a fifth as many tokens per MW at 150 tokens/s per user as at 100, while Vera Rubin holds up better. Faster agents would raise every figure here and widen Rubin’s lead.</p>
<p><strong>Output.</strong> Agents write about 1% of their tokens. An output token costs about five times a new input token, so agents that write more would raise compute faster than tokens.</p>
<h2>What to watch</h2>
<p><strong>Google’s next token count.</strong> Its third-quarter results, due in late October, are the cleanest read on whether the rest of the market is growing near OpenRouter’s pace.</p>
<p><strong>Rubin’s share of 2027 deliveries.</strong> Each GW of Vera Rubin does the work of about two of GB300s, so the mix of what ships in 2027 roughly halves or doubles the gigawatts needed.</p>
<p><strong>Cache rates.</strong> OpenRouter’s agent traffic went from about 60% cached in January to nearly 90% in September, which kept compute growth below token growth. It can’t rise much further, so from here more of token growth turns into compute growth.</p>
<h2>What this is anchored on</h2>
<p>Compute, power, GPU counts, tier shares, the 2027 mix and the market estimates are ours, built from the sources below. OpenRouter data is licensed under CC BY 4.0.</p>
<ul><li><strong>OpenRouter tokens</strong> — Daily tokens for the 50 biggest models plus one row for the rest, October 1–18, 2026; list prices from its model catalog · <a href="https://openrouter.ai/rankings">OpenRouter rankings-daily dataset and models API. Licensed under CC BY 4.0</a></li><li><strong>Token shape</strong> — Agent API keys used about 7× the tokens of human keys in late September; agent tokens about 87% cached input, 12% new input and 1% output; human tokens about 12% output, mostly new input (the 20% cached is our assumption) · <a href="https://openrouter.ai/data/newsletter/cache-is-crucial">OpenRouter, “Cache is Crucial,” Oct 8</a></li><li><strong>Demand forecast</strong> — OpenRouter tokens and spend to December 2027: low, base and high paths · <a href="https://06.ai/packs/ai-demand">06.ai </a><a href="https://06.ai/packs/ai-demand">AI-demand</a><a href="https://06.ai/packs/ai-demand"> pack</a></li><li><strong>Lab revenue</strong> — OpenAI’s and Anthropic’s annualized revenue to December 2027 · <a href="https://06.ai/packs/ai-lab-revenue">06.ai </a><a href="https://06.ai/packs/ai-lab-revenue">lab-revenue</a><a href="https://06.ai/packs/ai-lab-revenue"> pack</a></li><li><strong>Price cuts</strong> — OpenAI cut GPT-5.6 Luna’s prices 80%, to $0.20 per million input tokens and $1.20 output · <a href="https://devtk.ai/en/blog/ai-api-price-war-august-2026/">DevTk.AI, August 2026</a></li><li><strong>Agent throughput</strong> — Total tokens/s per utility MW serving DeepSeek V4 Pro at 100 tokens/s per user on the AgentX benchmark, interpolated between measured points: Vera Rubin NVL72 59.4M (pre-release TensorRT-LLM), GB300 NVL72 28.5M, B200 6.95M, B300 5.56M, H200 2.26M, MI355X 2.01M · <a href="https://inferencex.semianalysis.com/blog/vera-rubin-nvl72-agentic-inference">SemiAnalysis InferenceX, Sep 14</a></li><li><strong>Agent traffic</strong> — AgentX replays Claude Code sessions; on DeepSeek V4 median 88K tokens in and 413 out per request, 73–96% served from cache · <a href="https://inferencex.semianalysis.com/blog/agentx-inferencexv3-does-cuda-moat">SemiAnalysis InferenceX, Aug 24</a></li><li><strong>Token accounting</strong> — AgentX totals include reused (cached) input tokens · <a href="https://inferencex.semianalysis.com/glossary/throughput">InferenceX glossary</a></li><li><strong>Power basis</strong> — InferenceX’s power per GPU is an all-in utility envelope that includes cooling and data-center overhead (near 2.17 kW for a B200 in this post) · <a href="https://inferencex.semianalysis.com/blog/deepseekv4-16t-day-0-to-day-43-performance">InferenceX, Jun 9</a></li><li><strong>Power per GPU</strong> — H100 and H200 1.37 kW, B200 1.71 kW, B300 1.90 kW, MI355X 2.09 kW · <a href="https://inferencex.semianalysis.com/chips/b300">InferenceX chip pages, Oct 19</a></li><li><strong>Power per GPU</strong> — GB300 NVL72 12,350 tokens/s per chip and 5.83M per MW, so 2.12 kW each · <a href="https://inferencex.semianalysis.com/compare/deepseek-r1-gb300-vs-mi355x">InferenceX, DeepSeek-R1 GB300 vs MI355X</a></li><li><strong>Model benchmarks</strong> — Tokens/s per GPU at 8K in / 1K out and 50 tokens/s per user: gpt-oss-120b, Llama 3.3 70B and DeepSeek-R1 on H100, H200 and B200 · <a href="https://inferencex.semianalysis.com/run">InferenceX run pages</a></li><li><strong>Model benchmarks</strong> — MiniMax-M2.5 on H100 and B200, 8K in / 1K out · <a href="https://inferencex.semianalysis.com/blog/b200-minimax-m2-5-vllm-nvfp4-vs-h100-fp8-perf-per-dollar">InferenceX, May 26</a></li><li><strong>Model benchmarks</strong> — GLM-5 on H200 and B200, 8K in / 1K out · <a href="https://inferencex.semianalysis.com/blog/b200-glm5-nvfp4-vs-h200-fp8-3-6x-perf-per-dollar">InferenceX, May 26</a></li><li><strong>Model benchmarks</strong> — Kimi K2.5 on H200 and B200, 8K in / 1K out · <a href="https://inferencex.semianalysis.com/blog/b200-nvfp4-vs-h200-int4-kimi-k2-vllm-perf-per-dollar">InferenceX, May 26</a></li><li><strong>GB200 vs GB300</strong> — GB300 NVL72 made 1.2–2.8× GB200 NVL72’s tokens per GPU on DeepSeek V4 Pro at matched speeds (8K in / 1K out, runs a month after the model’s release) · <a href="https://inferencex.semianalysis.com/blog/gb300-nvl72-vs-gb200-nvl72-dsv4-pro-vllm-fp4">InferenceX, May 27</a></li><li><strong>Model sizes</strong> — DeepSeek V4 Pro: 1.6T parameters, 49B active · <a href="https://huggingface.co/blog/deepseekv4">Hugging Face blog</a></li><li><strong>Model sizes</strong> — DeepSeek V4.1 Flash: 552B parameters, 8B active on input and 16B on output; other open models from their model cards, except the estimates named in Method · <a href="https://openrouter.ai/deepseek/deepseek-v4.1-flash">OpenRouter model pages; Hugging Face</a></li><li><strong>AI chips delivered</strong> — 20.3M H100-equivalents delivered by end-2025; 31.6M operating by September 2026, from deliveries through June with a three-month lag to install · <a href="https://epoch.ai/latest/scaling-ai-data-centers-research">Epoch AI, Sep 16</a></li><li><strong>Power per chip</strong> — 94 tracked AI data centers: 14.8M H100-equivalents, 14.2 GW of IT power, 18.5 GW at the facility; about 44% of global AI compute · <a href="https://epoch.ai/data/ai-data-centers?view=table">Epoch AI data-center hub, Oct 19</a></li><li><strong>Chips added</strong> — H100-equivalents and chip power by quarter, from estimated sales by chip · <a href="https://epoch.ai/data/ai_chip_sales_timelines_by_chip.csv">Epoch AI chip-sales data, Oct 14 (our sums)</a></li><li><strong>Facility power</strong> — A data center draws about twice its AI chips’ rated power · <a href="https://epoch.ai/data/ai-chip-sales">Epoch AI chip-sales data, FAQ</a></li><li><strong>Inference share</strong> — OpenAI’s 2025 compute split roughly evenly between research and inference · <a href="https://epoch.ai/latest/introducing-the-ai-chip-users-explorer">Epoch AI, Sep 9</a></li><li><strong>NVIDIA</strong> — Data-center revenue $193.7B in fiscal 2026; $75.2B and $89.0B in the first two quarters of fiscal 2027 · <a href="https://s201.q4cdn.com/141608511/files/doc_financials/2027/Q227/Q2FY27-CFO-Commentary.pdf">NVIDIA filings and CFO commentary</a></li><li><strong>NVIDIA</strong> — Production shipments of Vera Rubin began in August; supply expected to remain a bottleneck at least through the end of fiscal 2028, in January 2028 · <a href="https://s201.q4cdn.com/141608511/files/content_files/TRANSCRIPT_-NVIDIA-Corp-NVDA-US-Q2-2027-Earnings-Call-26-August-2026-5_00-PM-ET.pdf">NVIDIA Q2 FY27 earnings call, Aug 26</a></li><li><strong>Software gains</strong> — DeepSeek V4 Pro on B200: about 300K to nearly 500K tokens/s per MW in six weeks · <a href="https://inferencex.semianalysis.com/blog/deepseekv4-16t-day-0-to-day-43-performance">InferenceX, Jun 9</a></li><li><strong>Software gains</strong> — DeepSeek-R1 on B200: up to 1.8× from one SGLang release · <a href="https://inferencex.semianalysis.com/blog/sglang-0-5-6-b200-deepseek-r1-fp4-up-to-1-8x">InferenceX, May 2</a></li><li><strong>Google</strong> — More than 3.2 quadrillion tokens a month across its products, May 2026, 7× a year earlier · <a href="https://blog.google/innovation-and-ai/sundar-pichai-io-2026/">Google, May 19</a></li><li><strong>Google API</strong> — About 22 billion tokens a minute through customers’ direct API use, Q2 2026 · <a href="https://blog.google/company-news/inside-google/message-ceo/alphabet-earnings-q2-2026/">Google, Jul 22</a></li><li><strong>OpenAI API</strong> — More than 15 billion tokens a minute, as of its March 31 funding round · <a href="https://www.coindesk.com/tech/2026/04/01/openai-raises-a-record-usd122-billion-at-as-revenue-crosses-usd2-billion-per-month">CoinDesk, Apr 1</a></li><li><strong>ByteDance</strong> — Doubao over 180T tokens a day, June 23 · <a href="https://news.10jqka.com.cn/20260623/c677639006.shtml">10jqka, Jun 23</a></li><li><strong>China</strong> — National use over 500T tokens a day in June, reported without a source or method · <a href="https://technode.com/2026/08/28/chinas-daily-ai-token-usage-tops-500-trillion-as-compute-demand-grows/">TechNode, Aug 28</a></li><li><strong>World estimate</strong> — About 370T tokens a day, July · <a href="https://io-fund.com/ai-stocks/ai-token-demand-shattering-forecasts">I/O Fund, Jul 30</a></li><li><strong>World estimate</strong> — More than 30 quadrillion tokens a month, late June · <a href="https://connect.cfauk.org/discussion/the-state-of-the-ai-economy">Exponential View, as relayed by CFA UK, Jul 5</a></li></ul>
<h2>Method</h2>
<p><strong>Tokens.</strong> OpenRouter’s rankings for October 1–18, the window of the <a href="https://06.ai/packs/ai-demand">AI-demand</a> pack’s October level: 23.8T tokens a day across the 50 biggest models and a row for the rest. Following OpenRouter’s October 8 data, we take seven-eighths of tokens as agent traffic (87% cached input, 12% new input, 1% output) and one-eighth as human traffic (12% output, as OpenRouter reports, and by our assumption 20% cached and 68% new input). That is 79% cached, 19% new and 2.4% output. The <a href="https://06.ai/packs/ai-demand">AI-demand</a> pack assumed 10% output and half of input cached to price tokens; its dollar total is pinned to Sacra’s revenue estimate, so we keep its dollars and use OpenRouter’s own split for compute.</p>
<p><strong>Compute per token.</strong> The reference is DeepSeek V4 Pro, 1.6T parameters with 49B active, on AgentX at 100 tokens/s per user, which counts cached input in its totals. Each model’s compute per token is its active parameters over 49B. DeepSeek V4.1 Flash, which uses 8B on input and 16B on output, is weighted by where the GPU time goes. GPU time per token is 5 for output, 1 for new input and 0.05 for cached input. Taking AgentX’s traffic to be shaped like OpenRouter’s agent traffic, an average OpenRouter token takes 1.63× an AgentX token. With AgentX’s own mix, 0.5% output and 73–96% of input from cache, it would take 1.1–3.1×, putting OpenRouter at 39–116 MW on today’s mix of hardware. That moves OpenRouter’s megawatts but not the market figure, which scales the fleet. Capacity runs at 60% utilization, the figure SemiAnalysis uses for revenue per GW.</p>
<p><strong>Closed models</strong> are sized from their input price: the median open model charges about $0.010 per million input tokens for each billion active parameters, and we assume closed labs keep a 60% gross margin against 30% for the hosts of open models. That puts Claude Opus 5.5 at about 221B active-equivalent parameters, a pricing proxy, not a disclosed size. Discounted versions take their model’s standard price: Muse Spark 1.3 Contributor is sized as Muse Spark 1.3. Open models’ sizes come from their model cards, except Space Bunny, GLM-5.3, GLM-5.2, MiMo-V2.5 and Nemotron 3.5 Super-VL, which are our estimates. The long tail takes the average of the rest.</p>
<p><strong>Hardware and power.</strong> AgentX’s tokens per MW for each system, with InferenceX’s power per GPU from its chip pages (1.37 kW for H200, 1.71 kW for B200, 1.90 kW for B300) and its GB300 comparison (2.12 kW). InferenceX’s power is all-in, including cooling and other data-center overhead, so every megawatt and gigawatt here is all-in. A June InferenceX post used about 2.17 kW for a B200; we use the chip pages throughout. GB200 NVL72 is our estimate at 70% of GB300 per MW and 1.87 kW per GPU, from rack power of about 135 kW. Today’s fleet is our estimate of the installed base by share of power: Hopper 21%, Blackwell 8-GPU servers 16%, NVL72 racks 32%, TPUs 15%, other 16%. On AgentX that averages 9.3M tokens/s per MW, against 28.5M for GB300 and 59.4M for Vera Rubin.</p>
<p><strong>The 2027 mix.</strong> Each month, every paid model’s tokens grow in proportion to its price raised to a negative power, with free tokens growing with the total. Each model’s price stays at its October level. We solve for the power and the scale that reproduce the <a href="https://06.ai/packs/ai-demand">AI-demand</a> pack’s token and spend paths, then re-weight by compute per token. In the base case the power reaches 0.68 by December 2027: a model ten times cheaper grows about 4.8× faster.</p>
<p><strong>The market.</strong> Epoch’s 31.6M H100-equivalents operating by September, at 1.25 kW each all-in (its tracked data centers: 18.5 GW at the facility for 14.8M), gives 39.5 GW; 40% serves inference (30–50%). The market’s demand grows at a set share of OpenRouter’s monthly compute growth, in logs: at 50%, a path that grows 10× becomes about 3×. New capacity needed is today’s inference fleet × (growth − 1) × today’s average tokens per MW ÷ the new hardware’s. The existing fleet stays in service. The range on today’s fleet efficiency also covers the gap between Epoch’s power basis and InferenceX’s. OpenRouter’s token shares use its average over the 30 days before each disclosure.</p>
<p><strong>2026.</strong> Our sums of Epoch AI’s estimated chip sales by quarter, which reproduce its published total of 20.3M H100-equivalents at the end of 2025 and, through June 2026, the 31.6M it counts as operating by September. The second half is 10–40% above the first. All-in power is twice the chips’ rated power, Epoch’s rule of thumb for a data center’s total draw. InferenceX’s power per GPU runs 1.5–1.7× the rating for Blackwell chips; on that basis 2026 adds 19–24 GW all-in and 7–10 GW for inference. Epoch’s estimates span roughly 1.5× either way, and it counts chips when delivered, not when running.</p>
<hr>
<p>Every chart above links to its interactive version on 06.ai, with the data behind it. The daily trackers are at <a href="https://06.ai/trackers">06.ai/trackers</a>.</p>
<p><em>Not investment advice. Estimates are labeled as estimates.</em></p>]]></content:encoded>
</item>
<item>
  <title>How big do OpenAI and Anthropic get by 2027?</title>
  <link>https://inthesingularity.ai/packs/ai-lab-revenue</link>
  <guid isPermaLink="true">https://inthesingularity.ai/packs/ai-lab-revenue</guid>
  <pubDate>Mon, 19 Oct 2026 12:08:00 GMT</pubDate>
  <description>We gathered 145 reported figures on the two labs’ revenue, rebuilt their run-rates month by month, tested simple forecasting rules against what was later reported, and projected to the end of 2027.</description>
  <category>AI labs</category>
  <enclosure url="https://inthesingularity.ai/newsletter/ai-lab-revenue/cover.png" length="65230" type="image/png" />
  <media:content url="https://inthesingularity.ai/newsletter/ai-lab-revenue/cover.png" medium="image" type="image/png" />
  <content:encoded><![CDATA[<figure><a href="https://06.ai/packs/ai-lab-revenue"><img src="https://06.ai/newsletter/ai-lab-revenue/thumb.png" alt="How big do OpenAI and Anthropic get by 2027?" width="421" height="391"></a><figcaption><a href="https://06.ai/packs/ai-lab-revenue">Interactive version and data on 06.ai →</a></figcaption></figure>
<p>Our base case has <strong>Anthropic at about $265B</strong> and <strong>OpenAI at about $230B</strong> of annualized revenue in December 2027, up from $71B and $50B at the end of September. The ranges overlap almost completely. History says the bigger risk is that both numbers are too low.</p>
<ul><li><strong>$72B</strong> — OpenAI, end of 2026. Range $66–79B. OpenAI expects at least $70B</li><li><strong>$100B</strong> — Anthropic, end of 2026. Range $88–114B. About $93B on OpenAI’s accounting</li><li><strong>$230B</strong> — OpenAI, end of 2027. Range $150–360B</li><li><strong>$265B</strong> — Anthropic, end of 2027. Range $185–400B. About $245B on OpenAI’s accounting</li></ul>
<h2>Rebuilt from the record, projected to 2027</h2>
<figure><a href="https://06.ai/packs/ai-lab-revenue#runrates"><img src="https://06.ai/newsletter/ai-lab-revenue/fig-01.png" alt="Rebuilt from the record, projected to 2027" width="694" height="520"></a><figcaption>Annualized revenue (run-rate), log scale. Lines are one smooth path per company fitted to 40 reported run-rate and revenue figures. Dashed lines are our base case; shading is the 80% range. Anthropic counts sales through cloud marketplaces in full; OpenAI counts only its share.</figcaption></figure>
<h2>The leaks agree with each other</h2>
<p>Neither company reports its run-rate on a schedule. The figures arrive through funding announcements, investor updates and leaks, on different dates and sometimes on different definitions. We collected 145 of them. 80 give a run-rate or revenue for a specific period; dropping repeats of the same figure, later restatements and figures on a different basis leaves the 40 we fit. One smooth path through each company’s figures misses the typical Anthropic figure by 3% and the typical OpenAI figure by 1%; the largest miss is 17%. The run-rates, the quarterly revenue and Anthropic’s draft prospectus tell one story.</p>
<p>The OpenAI path implies about $10.1B of revenue in the third quarter, up 50% from the second. OpenAI hasn’t reported it yet.</p>
<h2>Anthropic slowed as OpenAI sped up</h2>
<figure><a href="https://06.ai/packs/ai-lab-revenue#growth"><img src="https://06.ai/newsletter/ai-lab-revenue/fig-02.png" alt="Anthropic slowed as OpenAI sped up" width="694" height="362"></a><figcaption>Monthly growth of the run-rate, from the fitted paths. The fit smooths launch months, so single months were lumpier than this.</figcaption></figure>
<h2>The curves crossed in the summer</h2>
<p>On the fitted path, Anthropic’s run-rate grew about 44% a month in February and 49% in March and was down to about 3% by September; only the third-quarter total pins down how August and September split. Its revenue rose 45% in the third quarter, after 145% in the second. OpenAI went the other way: about 3–4% a month in March and April, 17–19% from July to September. The Information reported that OpenAI’s run-rate grew 70% from the end of June to the end of September, and Bloomberg that enterprise sales were driving most of the growth.</p>
<h2>What a simple rule would have said</h2>
<figure><a href="https://06.ai/packs/ai-lab-revenue#backtest"><img src="https://06.ai/newsletter/ai-lab-revenue/fig-03.png" alt="What a simple rule would have said" width="694" height="337"></a><figcaption>At past half-year ends, we projected the next 12 months at the growth of the previous 12, using only figures public at the time (dashed, starting from the latest figure then available; Anthropic’s figures are too sparse before mid-2025). Solid lines and dots are what was reported later. Log scale.</figcaption></figure>
<figure><a href="https://06.ai/packs/ai-lab-revenue#backtest"><img src="https://06.ai/newsletter/ai-lab-revenue/fig-04.png" alt="OpenAI, trend — Up to 3 months out: 18% mixed · 4–7: 8% mixed · 8–12: 19% high · 13–18: 24% high; OpenAI, fading — Up to 3 months out: 18% low · 4–7: 13% low · 8–12: 23% low · 13–18: 35% low; Anthropic, trend — Up to 3 months out: 30% low · 4–7: 34% low · 8–12: 32% low · 13–18: 74% low; Anthropic, fading — Up to 3 months out: 29% low · 4–7: 39% low · 8–12: 48% low · 13–18: 86% low" width="694" height="230"></a></figure>
<p>Trend: the last 12 months’ growth continues. Fading: that growth halves over 18 months. Each rule was run at every month-end since March 2024 (January 2025 for Anthropic) and scored against every fitted run-rate and quarterly revenue figure dated up to 18 months later, as forecast ÷ outcome − 1. The table shows the median size of the miss and its usual direction.</p>
<h2>OpenAI followed the rule; Anthropic broke it twice</h2>
<p>The simplest forecast is that the next 12 months grow like the last 12. Run at every month-end with only the figures public at the time, it called OpenAI well: its typical miss was 8% four to seven months out and 19–24% further out, on the high side. On Anthropic it missed by about a third four to 12 months out and by 74% beyond a year, on the low side, because Anthropic kept accelerating. Then it broke the other way. Run today, it puts Anthropic at about $174B at the end of 2026, far above the roughly $100–110B in recent reporting.</p>
<p>The version that has growth halving over 18 months did worse than the plain rule beyond three months for both companies, because neither has slowed on a twelve-month view yet. Anthropic has slowed for a few months before: its monthly growth fell from about 40% to about 12% during 2025, then picked up again in early 2026.</p>
<h2>OpenAI’s plans land; Anthropic’s haven’t</h2>
<figure><a href="https://06.ai/packs/ai-lab-revenue#plans"><img src="https://06.ai/newsletter/ai-lab-revenue/fig-05.png" alt="OpenAI’s plans land; Anthropic’s haven’t" width="694" height="500"></a><figcaption>What happened, divided by what the company projected, log scale. Open circles compare a plan with our 2026 estimates. Two projections were made partway through the period they cover: September 2024 for 2024 revenue and May 2026 for the second quarter.</figcaption></figure>
<p>OpenAI hit its 2024 revenue plan exactly and its 2025 plans within 3–13%. On its own latest forecast of about $35B, it will beat the 2026 plans it had in 2025 by 17–19% (about 25% on our estimate). Only its earliest pitch was far off: in late 2022 it told investors it would make $1B in 2024, and it made $3.7B. Anthropic’s projections a year out have been too low, mostly by a lot: its 2025 base case was less than half what it booked, and in October 2025 its internal base case for its 2026 run-rate was $20B. Its short-range calls held up: it projected $9B for the end of 2025 and got $9B, and its second-quarter projection was within 6%. Its one miss on the high side was its first: a $200M pace by the end of 2023, against $87M at the start of 2024.</p>
<h2>OpenRouter tracks Anthropic, with big swings in share</h2>
<figure><a href="https://06.ai/packs/ai-lab-revenue#openrouter"><img src="https://06.ai/newsletter/ai-lab-revenue/fig-06.png" alt="OpenRouter tracks Anthropic, with big swings in share" width="694" height="307"></a><figcaption>Left: Anthropic’s share of estimated spend through OpenRouter, by month; October is the 1st–18th. Right: Anthropic’s run-rate if its OpenRouter spend follows the <a href="https://06.ai/packs/ai-demand">AI-demand</a> pack’s base case for the whole market at three different shares, against our base case.</figcaption></figure>
<figure><a href="https://06.ai/packs/ai-lab-revenue#openrouter"><img src="https://06.ai/newsletter/ai-lab-revenue/fig-07.png" alt="Holds at October’s level — Share: 37% · Run-rate: $350B; Settles at 30% — Share: 30% · Run-rate: $285B; Keeps falling at 2026’s pace — Share: 19% · Run-rate: $180B; Our base case implies — Share: 28% · Run-rate: $265B" width="694" height="229"></a></figure>
<h2>A read on Anthropic, not on OpenAI</h2>
<p>Spend on Anthropic’s models through OpenRouter has run at 1.5–2.1% of its revenue each quarter since the end of 2025. Scaling each quarter’s OpenRouter spend by the previous quarter’s ratio would have called Anthropic’s revenue within 12–17% in four of the last five quarters and been 29% off in the fifth (the two 2025 quarters are our estimates from annual revenue). It can’t do the same for OpenAI, whose revenue comes mostly from ChatGPT and from businesses buying directly, which OpenRouter doesn’t see.</p>
<p>The swing factor is Anthropic’s share. It held at 54–62% of OpenRouter’s dollars from January to July, fell to 42% in August as its own spend there dropped, and to 28% in September as OpenAI’s GPT-5.6 and GPT-6 models took share. It recovered to 37% in October after Anthropic released Opus 5.5 and Sonnet 5.5 in late September and Haiku 5.5 in early October. Opus 5.5 alone is drawing more spend than Opus 5 ever did, at lower prices: $4 and $20 per million input and output tokens, against $5 and $25. Revenue moves less than this share does: in August and September, Anthropic’s spend through OpenRouter ran about 30% below July’s while its run-rate kept rising.</p>
<p><a href="https://06.ai/packs/ai-demand">The </a><a href="https://06.ai/packs/ai-demand">AI-demand</a><a href="https://06.ai/packs/ai-demand"> pack</a>’s base case has spend through OpenRouter growing 4.2× from October to December 2027. If Anthropic held its October share, that would put its run-rate near $350B; if its share settled at 30%, about $285B; if it kept falling at this year’s pace, about $180B. Our base case of $265B implies a share near 28%, about where it was at its September low.</p>
<h2>How fast growth fades decides 2027</h2>
<p>Run-rate in December 2027 if each company’s growth rate entering 2027 in our base case (about 13% a month for OpenAI and 11% for Anthropic) halves at different speeds.</p>
<figure><a href="https://06.ai/packs/ai-lab-revenue#forecast"><img src="https://06.ai/newsletter/ai-lab-revenue/fig-08.png" alt="Every 6 months — OpenAI: $155B · Anthropic: $195B; Every 12 months — OpenAI: $205B · Anthropic: $240B; Every 18 months (base) — OpenAI: $230B · Anthropic: $265B; Never — OpenAI: $305B · Anthropic: $340B" width="694" height="229"></a></figure>
<h2>The forecast</h2>
<p><strong>End of 2026.</strong> OpenAI expects to reach or exceed a $70B run-rate. It beat its end-2025 target by 7%, and the trend rule says $70B. We use $72B. For Anthropic, October’s spend through OpenRouter implies a run-rate of about $84B at this year’s average quarterly ratio (1.8%), and about $80B if October behaves like June, the last month with a launch-driven jump of the same size. We take $83B. Growth of about 10% a month from there reaches $100B in December, with about $22B of fourth-quarter revenue. That is at the top of what the FT reported after Anthropic’s investor day, an annualized figure below $100B, and short of the $110B analysts said was possible.</p>
<p><strong>2027.</strong> For OpenAI we assume its growth rate in December 2026, about 13% a month, halves every 18 months. That gives about $230B in December 2027. For Anthropic we take the midpoint of two methods. The same rule, applied to its growth of about 10% a month late in 2026, gives $245B. OpenRouter gives $285B if Anthropic’s share of spend there settles at 30% while the market follows the <a href="https://06.ai/packs/ai-demand">AI-demand</a> pack’s base case. The midpoint, $265B, is what you get if Anthropic enters 2027 growing about 11% a month and that rate halves every 18 months. The ranges come from varying the year-end level, the starting growth rate and the fade, from halving every 6 months to every 4½ years. On OpenAI’s accounting Anthropic would be about $245B, a little ahead of OpenAI.</p>
<p>For OpenAI, history leans to the upside: a rule that had its growth halving over 18 months ran 23–35% low eight to 18 months out. For Anthropic it cuts both ways: rules built on its trailing growth were far too low until mid-2026 and are far too high now. Scale argues for caution. Each company would add roughly $155–165B of run-rate in 2027, more than either has added in a year so far, and both depend on compute arriving on schedule. Anthropic’s own IPO projection, $190–200B of revenue in 2028, implies about $140B at the end of 2027 if its growth fades as in our base case. It has beaten every longer-range projection it has made since the start of 2025.</p>
<h2>What to watch</h2>
<p><strong>Anthropic on OpenRouter in November.</strong> If spend holds near the $4.5–5M a day it has averaged since October 6, the October jump was real demand and $100B by December looks low. If it falls back by a third or more, as it did in late July, $85–90B is more likely.</p>
<p><strong>OpenAI’s third-quarter revenue.</strong> Our fitted path says about $10.1B. Much less would mean the run-rate figures ran ahead of revenue.</p>
<p><strong>Anthropic’s prospectus.</strong> If Anthropic goes ahead with its IPO, the public filing should show quarterly revenue on one definition, next to the run-rates it has been quoting.</p>
<h2>What this is anchored on</h2>
<p>Run-rates and revenue are as reported; fitted paths, back-tests, OpenRouter spend and forecasts are our estimates. All 40 figures used in the fit are listed below with their sources.</p>
<ul><li><strong>OpenAI</strong> — About $50B run-rate at the end of September on OpenAI’s own (net) basis; the ~$70B figure came from investors applying Anthropic’s gross method; expects to reach or exceed $70B by the end of 2026 · <a href="https://www.bloomberglinea.com/negocios/openai-preve-superar-los-us70000-millones-en-ingresos-anualizados-a-finales-de-2026/">Bloomberg, via Bloomberg Línea, Oct 9</a></li><li><strong>OpenAI</strong> — About $50B at the end of September, less than the ~$70B previously indicated to investors · <a href="https://www.thestandard.com.hk/finance/article/345035/OpenAIs-September-annualized-revenue-nears-US50-billion-less-than-previously-indicated-source-says">Reuters, citing the FT, via The Standard, Oct 8</a></li><li><strong>OpenAI</strong> — Run-rate grew 70% from the end of June to the end of September (The Information, citing the CFO); enterprise sales driving most of the growth (Bloomberg) · <a href="https://thenextweb.com/news/openai-revenue-70bn-year-end-50bn-september">The Next Web, Oct 9</a></li><li><strong>OpenAI</strong> — Full-year 2026 revenue of around $35B · <a href="https://aiweekly.co/alerts/anthropic-books-gross-cloud-sales-openai-only-its-net-share">Bloomberg, via AI Weekly</a></li><li><strong>OpenAI</strong> — Annual recurring revenue “nearing $70 billion” · <a href="https://www.axios.com/2026/09/29/scoop-openais-annual-recurring-revenue-nears-70b">Axios, Sep 29</a></li><li><strong>OpenAI</strong> — More than $40B run-rate in mid-August · <a href="https://investinglive.com/stocks/openai-s-annualized-revenue-run-rate-has-surpassed-40-billion-says-a-bloomberg-report/">Bloomberg, via investingLive, Aug 13</a></li><li><strong>OpenAI</strong> — $2B of revenue a month (March 2026) · <a href="https://openai.com/index/accelerating-the-next-phase-ai/">OpenAI, Mar 31</a></li><li><strong>OpenAI</strong> — $21.4B run-rate at the end of 2025; more than $25B at the end of February · <a href="https://www.techloy.com/openai-hits-25-billion-in-annualized-revenue-while-anthropic-surges-to-19-billion/">The Information, via Techloy, Mar 5</a></li><li><strong>OpenAI</strong> — Revenue $5.7B in Q1 2026 · <a href="https://thenextweb.com/news/openai-burned-3-7-billion-q1-2026">The Information, via The Next Web, Jun 17</a></li><li><strong>OpenAI</strong> — Revenue $6.7B in Q2 2026 · <a href="https://aiweekly.co/alerts/wsj-openai-q2-revenue-grew-18-qoq-to-67b-as-operating-margin-sank-further">Wall Street Journal, via AI Weekly, Aug 18</a></li><li><strong>OpenAI</strong> — Audited revenue: $3.7B in 2024, $13.07B in 2025 · <a href="https://wheresyoured.at/exclusive-openai-financials">Where’s Your Ed At (verified by the FT), Jun 15</a></li><li><strong>OpenAI</strong> — Projected revenue of $12.7B for 2025 and $29.4B for 2026 (March 2025) · <a href="https://www.advisorperspectives.com/articles/2025/03/27/openai-revenue-triple-12-7-billion-year">Bloomberg, via Advisor Perspectives</a></li><li><strong>Anthropic</strong> — Revenue $16.8B in Q3 2026, up about 45% from Q2 · <a href="https://thenextweb.com/news/anthropic-q3-revenue-16-8bn-ipo">Financial Times, via The Next Web, Oct 14</a></li><li><strong>Anthropic</strong> — Revenue $4.7B in Q1 2026 and $11.6B in Q2 · <a href="https://www.businessday.co.za/world/international-companies/2026-10-15-anthropic-revenue-jumps-45-to-168bn/">Reuters, via BusinessDay, Oct 15</a></li><li><strong>Anthropic</strong> — Run-rate of $65B at the end of July · <a href="https://news.bgov.com/financial-accounting/anthropic-revenue-run-rate-surpasses-65-billion-ahead-of-ipo">Bloomberg, via Bloomberg Government, Aug 17</a></li><li><strong>Anthropic</strong> — Run-rate of $14B · <a href="https://www.anthropic.com/news/anthropic-raises-30-billion-series-g-funding-380-billion-post-money-valuation">Anthropic, Series G announcement, Feb 12</a></li><li><strong>Anthropic</strong> — Run-rate above $30B, up from about $9B at the end of 2025 · <a href="https://www.anthropic.com/news/google-broadcom-partnership-compute">Anthropic, Apr 6</a></li><li><strong>Anthropic</strong> — Run-rate crossed $47B in early May · <a href="https://www.anthropic.com/news/series-h">Anthropic, Series H announcement, May 28</a></li><li><strong>Anthropic</strong> — Revenue of about $386M in 2024 and $4.59B in 2025 (draft IPO prospectus) · <a href="https://www.tradingkey.com/analysis/stocks/us-stocks/262190854-anthropic-ipo-prospectus-revealed-tradingkey">Reuters, via TradingKey, Sep 29</a></li><li><strong>Anthropic</strong> — Revenue of nearly $4.6B in 2025; cloud-marketplace sales booked in full, with the partners’ cut as an expense · <a href="https://www.khaleejtimes.com/business/anthropic-ipo-prospectus-lays-bare-deep-dependence-on-big-tech-partners">Reuters, via Khaleej Times, Sep 30</a></li><li><strong>Anthropic</strong> — Internal base case of $20B of annualized revenue for 2026, best case $26B (October 2025) · <a href="https://finance.yahoo.com/news/exclusive-anthropic-aims-nearly-triple-170234463.html">Reuters, via Yahoo Finance, Oct 15, 2025</a></li><li><strong>Anthropic</strong> — Projected 2026 revenue of $18B (January 2026) · <a href="https://www.investing.com/news/stock-market-news/anthropic-ups-2026-revenue-forecast-delays-cashpositive-target--the-information-4469111">The Information, via Investing.com, Jan 27</a></li><li><strong>Anthropic</strong> — 2028 revenue of roughly $190–200B underpinning the IPO valuation · <a href="https://finance.yahoo.com/technology/ai/articles/exclusive-anthropic-ipo-valuation-hinges-001158885.html">Reuters, via Yahoo Finance, Aug 14</a></li><li><strong>Anthropic</strong> — Annualized revenue expected to stay below $100B · <a href="https://finance.yahoo.com/technology/ai/articles/anthropic-reports-45-revenue-growth-102235753.html">Financial Times, via Yahoo Finance, Oct 15</a></li><li><strong>Anthropic</strong> — Analysts who met executives: run-rate “could reach $110 billion by the end of the year” · <a href="https://www.calcalistech.com/ctechnews/article/tgb6eqw4z">Reuters, via Calcalist, Oct 15</a></li><li><strong>Anthropic</strong> — About $74B annualized in October (third-party estimate) · <a href="https://www.implicator.ai/anthropic-tells-investors-q3-revenue-hit-16-8-billion-as-growth-slowed-to-45/">SemiAnalysis, via Implicator, Oct 15</a></li><li><strong>OpenRouter</strong> — Daily tokens by model, Jan 1, 2025–Oct 18, 2026; spend is our estimate at catalog prices (see the <a href="https://06.ai/packs/ai-demand">AI-demand</a> pack) · <a href="https://openrouter.ai/rankings">OpenRouter rankings. Licensed under CC BY 4.0</a></li><li><strong>OpenRouter</strong> — List prices, including Opus 5.5 at $4 and $20 per million input and output tokens and Opus 5 at $5 and $25 · <a href="https://openrouter.ai/docs/api/api-reference/endpoints/list-all-endpoints-for-a-model.md">OpenRouter models API, Oct 19</a></li></ul>
<h2>Method</h2>
<p><strong>Figures.</strong> We collected 145 reported figures on the two companies’ revenue, run-rates, projections and users, with a link for each. Of these, 80 give a run-rate or revenue for a period; we fit 40, dropping repeats of the same figure, later restatements and figures on a different basis. Third-party estimates are shown but not fitted. OpenAI’s 2024 and 2025 full-year revenue includes items its run-rate figures leave out, so we don’t fit it. Each figure is placed on the date it describes; the back-test uses the date it was first reported.</p>
<p><strong>Fitted paths.</strong> Each company’s log run-rate is a piecewise-linear curve with knots every half month and a penalty on curvature. Run-rates are fitted as points; quarterly and annual revenue as the integral of run-rate ÷ 12 over the period, weighted more heavily because they are booked numbers. Monthly growth in the second chart is month-end to month-end on these paths.</p>
<p><strong>Back-test.</strong> At each month-end from March 2024 (January 2025 for Anthropic) to September 2026, a rule sees only figures reported by then. Its level is the latest figure (quarterly revenue counts as four times itself at mid-quarter) and its growth is the average over the previous 12 months, held constant or halving over n months. It is scored against every fitted run-rate and quarterly revenue figure dated up to 18 months later, as forecast ÷ outcome − 1.</p>
<p><strong>Plans</strong> compare projections with what was later reported. Two were made partway through the period they cover: September 2024 for 2024 revenue and May 2026 for the second quarter. 2026 outcomes are our estimates.</p>
<p><strong>OpenRouter.</strong> Spend on each lab’s models is estimated as in the <a href="https://06.ai/packs/ai-demand">AI-demand</a> pack, at catalog prices. The share is Anthropic’s spend over all spend. The scenarios scale the <a href="https://06.ai/packs/ai-demand">AI-demand</a> pack’s base path for total spend by Anthropic’s share, moving geometrically from October’s level to the end value, and convert spend to run-rate at October’s ratio.</p>
<p><strong>Forecast.</strong> End-2026: OpenAI from its own target; Anthropic from an October run-rate of about $83B, between a launch-month reading of OpenRouter and its quarterly average ratio, then about 10% a month. 2027: growth halves every 18 months, g(t) = g₀ × 2<sup>−t/18</sup>. For OpenAI g₀ is its December 2026 growth; for Anthropic it is set so December 2027 lands midway between that rule applied to its late-2026 growth and the OpenRouter lens with a 30% share. The 80% ranges simulate the year-end level (±7% for OpenAI and ±10% for Anthropic at one standard deviation), the starting growth (±25%) and a half-life between 6 and 54 months. Run-rates are each company’s own: Anthropic books sales through cloud marketplaces in full and pays the partners’ cut as an expense, about 16% of those sales, which were 47% of its 2025 revenue; we put Anthropic about 7.5% lower on OpenAI’s net method.</p>
<hr>
<p>Every chart above links to its interactive version on 06.ai, with the data behind it. The daily trackers are at <a href="https://06.ai/trackers">06.ai/trackers</a>.</p>
<p><em>Not investment advice. Estimates are labeled as estimates.</em></p>]]></content:encoded>
</item>
<item>
  <title>How fast is AI demand growing?</title>
  <link>https://inthesingularity.ai/packs/ai-demand</link>
  <guid isPermaLink="true">https://inthesingularity.ai/packs/ai-demand</guid>
  <pubDate>Sun, 18 Oct 2026 12:07:00 GMT</pubDate>
  <description>OpenRouter publishes daily token counts for the models it routes. Tokens are up 33× in a year and spend about 11× since January. What’s driving it, who captures it, and three paths for 2027.</description>
  <category>AI demand</category>
  <enclosure url="https://inthesingularity.ai/newsletter/ai-demand/cover.png" length="53899" type="image/png" />
  <media:content url="https://inthesingularity.ai/newsletter/ai-demand/cover.png" medium="image" type="image/png" />
  <content:encoded><![CDATA[<figure><a href="https://06.ai/packs/ai-demand"><img src="https://06.ai/newsletter/ai-demand/thumb.png" alt="How fast is AI demand growing?" width="421" height="415"></a><figcaption><a href="https://06.ai/packs/ai-demand">Interactive version and data on 06.ai →</a></figcaption></figure>
<p>OpenRouter routed <strong>23.8T tokens a day</strong> in October, 33× as many as a year earlier. We estimate its customers are spending about <strong>$4.2B a year</strong> on them, 11× January’s pace. Tokens are growing faster than dollars, Chinese models supply most of the tokens while US labs collect most of the dollars, and most tokens go to code and agent work.</p>
<ul><li><strong>23.8T</strong> — Tokens a day, October. 33× October 2025. 2.2 quadrillion so far this year, against OpenRouter’s May forecast of more than 1 quadrillion for all of 2026</li><li><strong>$4.2B</strong> — Annualized spend, October. Our estimate of what customers spend through OpenRouter, 11× January. About $230M a year of fees for OpenRouter at 5.5%</li><li><strong>$0.48</strong> — Average price, million tokens. October, against $0.96 in the first half, as cheaper models took share</li><li><strong>73K</strong> — Tokens per request, top 100 apps. 5.4× January. Agents re-send their working context at every step</li></ul>
<h2>Tokens: 33× in a year</h2>
<figure><a href="https://06.ai/packs/ai-demand#tokens"><img src="https://06.ai/newsletter/ai-demand/fig-01.png" alt="Tokens: 33× in a year" width="694" height="442"></a><figcaption>Tokens a week by where each model’s developer is based, January 2025 to October 18, 2026. Other is the rest of the world, stealth models and OpenRouter’s long-tail row. Each model’s tokens are counted with its own tokenizer. Shown: Stacked.</figcaption></figure>
<figure><a href="https://06.ai/packs/ai-demand#tokens"><img src="https://06.ai/newsletter/ai-demand/fig-02.png" alt="Tokens: 33× in a year — Log scale" width="694" height="473"></a><figcaption><strong>Log scale.</strong></figcaption></figure>
<h2>What OpenRouter can see</h2>
<p>OpenRouter sells access to more than 400 models through one API and publishes the daily token count of the 50 biggest, plus a single row for everything else. It sees developers, startups and agent tools. It doesn’t see ChatGPT’s subscribers or companies that call OpenAI and Anthropic directly, and its public counts leave out private and zero-data-retention traffic. So it is a slice: about $4.2B a year of spend, against roughly $120–140B of annualized revenue at OpenAI and Anthropic combined, depending on how OpenAI’s revenue is counted. But it is one of the few public daily counts of AI demand, and it has been growing faster than Google, which gives a figure from time to time. Google said in May it was processing 3.2 quadrillion tokens a month, 7× a year earlier. OpenRouter grew 12× over the same twelve months.</p>
<h2>Tokens up 23× this year, dollars 11×</h2>
<p>Since January, daily tokens are up 23× and our estimate of spend 11×. The difference is price. The average million tokens cost $0.96 in the first half of the year and $0.48 in October. We price every model at today’s list price, so the whole fall is mix: cheap models, above all from DeepSeek and Z.ai, took share. At the first half’s average price, October’s tokens would be worth $8.3B a year, not $4.2B.</p>
<p>Dollars are also where growth has started to slow. Spend grew 30% a month from April to July and 21% a month from July to October. Tokens went the other way, from 35% to 48%.</p>
<h2>Dollars are growing more slowly than tokens</h2>
<figure><a href="https://06.ai/packs/ai-demand#spend"><img src="https://06.ai/newsletter/ai-demand/fig-03.png" alt="Dollars are growing more slowly than tokens" width="694" height="422"></a><figcaption>Estimated spend through OpenRouter, annualized from each week, log scale. Dots are Sacra’s estimates of OpenRouter’s revenue divided by its 5.5% fee; July’s is the one we calibrate to. The gray line prices each week’s tokens at the first half’s average. Shading runs to the last week in which more than a tenth of tokens came from models no longer in OpenRouter’s price list.</figcaption></figure>
<h2>Chinese models supply the tokens; US labs collect the dollars</h2>
<figure><a href="https://06.ai/packs/ai-demand#labs"><img src="https://06.ai/newsletter/ai-demand/fig-04.png" alt="Chinese models supply the tokens; US labs collect the dollars" width="694" height="512"></a><figcaption>Share of tokens and of estimated spend, September 21 to October 18. For Chinese labs, “serves” is how much of the spend on their models goes to the lab’s own API; the rest goes to other hosts. Free &amp; stealth is free versions and unreleased models, counted at $0.</figcaption></figure>
<h2>China’s tokens, America’s dollars</h2>
<p>Models from Chinese labs first passed US models in weekly tokens in the week of March 23 and have led every week since April 27. Over the past four weeks they were 57% of tokens (64% last week) and 26% of estimated spend. DeepSeek is the biggest lab by tokens, with 26%, and takes 4% of the dollars. Anthropic has 5% of the tokens and 35% of the dollars; OpenAI 9% and 26%. Between them, the two take 61% of the dollars from 14% of the tokens.</p>
<p>Much of what is spent on Chinese models doesn’t reach Chinese labs either. Their models are open-weight, so anyone can host them: 28 providers serve the September version of DeepSeek V4.1-Flash, its most-used model, and DeepSeek isn’t one of them. On a snapshot of where requests go, DeepSeek serves 9% of the spend on its own models, Z.ai 6% and Moonshot 10%. The rest goes to hosts such as Together, DeepInfra and Baseten. Tencent, Xiaomi and MiniMax keep most of theirs.</p>
<p>Those hosts often charge more than the single price OpenRouter lists for a model, which is the one we use. Weighting each host’s price by its share of requests in the same snapshot would add about 17% to October’s spend, most of it on Chinese models, and lift their share of the dollars to about 37%. We leave that out of our figures because a 30-minute snapshot can’t be applied to earlier months.</p>
<h2>Agents are the demand</h2>
<p>The average request from OpenRouter’s 100 biggest apps carried 73,000 tokens in October, against 13,000 in January. Coding agents re-send their working context at every step, so one task can run to millions of tokens. Hermes Agent, Claude Code, Kilo Code and Cline are the four biggest apps, and OpenRouter classifies 63% of last week’s tokens as code or agent work. Roleplay, about a third of tokens in mid-2025, is down to about 5%.</p>
<p>Agents also make dollars harder to measure. Much of what they send is context they have sent before, which providers bill at cached rates, often a tenth of the input price or less. The cached share is the biggest unknown in our spend estimate; matching Sacra’s July revenue figure implies about 51% of input.</p>
<h2>Free tokens are a launch strategy</h2>
<p>Labs buy demand on OpenRouter with free models. Free versions and stealth models, released under code names before launch, were 15% of tokens over the past four weeks and peaked at 31% in April 2025. Z.ai ran GLM-5.3-Flash as “Ox Alpha” for a week in August; on August 24, OpenRouter said it was on track for nearly 6 trillion tokens that day. A model called Space Bunny Alpha has done 57 trillion tokens in the past 30 days without its maker being named. We count all of it at $0, which is why tokens and dollars can diverge sharply within a month.</p>
<h2>Requests got five times bigger</h2>
<figure><a href="https://06.ai/packs/ai-demand#agents"><img src="https://06.ai/newsletter/ai-demand/fig-05.png" alt="Requests got five times bigger" width="694" height="323"></a><figcaption>Left: tokens per request across the 100 biggest apps each month; labels mark when a new app became the biggest. October is the 1st–18th. Right: OpenRouter’s sampled estimates of each category’s share of tokens, four-week averages; categories can overlap, and the latest two weeks are still being classified and are left out.</figcaption></figure>
<h2>Spend on Anthropic’s models through OpenRouter has tracked its revenue</h2>
<figure><a href="https://06.ai/packs/ai-demand#anthropic"><img src="https://06.ai/newsletter/ai-demand/fig-06.png" alt="Spend on Anthropic’s models through OpenRouter has tracked its revenue" width="694" height="312"></a><figcaption>Anthropic’s reported revenue and our estimate of spend on its models through OpenRouter, each indexed to Q1 = 1. Q4 is October 1–18 at a full-quarter pace.</figcaption></figure>
<figure><a href="https://06.ai/packs/ai-demand#anthropic"><img src="https://06.ai/newsletter/ai-demand/fig-07.png" alt="Q1 2026 — Revenue: $4.7B · On OpenRouter: $83M · Share: 1.8%; Q2 2026 — Revenue: $11.5B · On OpenRouter: $238M · Share: 2.1%; Q3 2026 — Revenue: $16.8B · On OpenRouter: $285M · Share: 1.7%; Q4, Oct 1–18 pace — Revenue: $19–23B implied · On OpenRouter: $388M" width="694" height="230"></a></figure>
<h2>A read on Anthropic</h2>
<p>Spend on Anthropic’s models through OpenRouter was 1.8% of its reported revenue in the first quarter, 2.1% in the second and 1.7% in the third. In October so far it is running 36% above the third quarter’s pace. If OpenRouter’s share stays in that range, Anthropic is on course for $19–23B of revenue in the fourth quarter, against $16.8B in the third. That is the quarter in which it plans to list; bankers are reported to be pricing the IPO on $190–200B of revenue in 2028. Anthropic released Opus 5.5 and Sonnet 5.5 in late September and Haiku 5.5 in October, and launch weeks can pull spend forward.</p>
<p>OpenAI is different. Spend on its models through OpenRouter grew 4× from the first quarter to the second and 13× by the third, while its reported revenue grew 18% from the first quarter to the second. OpenRouter is not a read on OpenAI.</p>
<h2>Three paths for 2027</h2>
<figure><a href="https://06.ai/packs/ai-demand#forecast"><img src="https://06.ai/newsletter/ai-demand/fig-08.png" alt="Three paths for 2027" width="694" height="412"></a><figcaption>By month, log scale; October is the 1st–18th. Projections start from the past quarter’s monthly growth (21% for dollars, 48% for tokens) and assume that rate halves every few months. Move the slider to choose how many. Shown: Dollars.</figcaption></figure>
<p><em>At the default setting (Growth halves every 6 months), from the </em><a href="https://06.ai/packs/ai-demand#forecast"><em>interactive version</em></a><em>:</em></p>
<ul><li>Spend in December 2027, a year: <strong>$18B</strong></li><li>Multiple of October: <strong>4.2×</strong></li><li>OpenRouter fees at 5.5%: <strong>$1.0B</strong></li><li>Monthly growth by then: <strong>4.7%</strong></li></ul>
<figure><a href="https://06.ai/packs/ai-demand#forecast"><img src="https://06.ai/newsletter/ai-demand/fig-09.png" alt="3 months — Dec 2026: $5.9B · Jun 2027: $9.3B · Dec 2027: $10B · Tokens a day, Dec 2027: 171T; 6 months (base) — Dec 2026: $6.0B · Jun 2027: $12B · Dec 2027: $18B · Tokens a day, Dec 2027: 513T; 12 months — Dec 2026: $6.1B · Jun 2027: $15B · Dec 2027: $29B · Tokens a day, Dec 2027: 1,350T" width="694" height="189"></a></figure>
<figure><a href="https://06.ai/packs/ai-demand#forecast"><img src="https://06.ai/newsletter/ai-demand/fig-10.png" alt="Three paths for 2027 — Tokens" width="694" height="442"></a><figcaption><strong>Tokens.</strong></figcaption></figure>
<h2>The forecast</h2>
<p>Spend growth on OpenRouter slowed from 30% a month to 21% between the last two quarters. That is what a growth rate halving about every 6 months looks like, and it is our base case. On it, spend reaches about $18B a year by December 2027, 4.2× today’s, and is still growing about 5% a month. If growth halves every three months, spend levels off near $10B; if it halves only once a year, it reaches about $29B. Even the low case is 2.5× today’s. The high case needs another wave like 2026’s, with agents spreading from coding into other work.</p>
<p>The token paths are steeper because tokens have been growing faster than dollars. In the base case OpenRouter handles about 513T tokens a day by the end of 2027 at about $0.09 per million. That only happens if cheap models keep taking share.</p>
<p>The three paths barely differ over the next quarter: spend grows 60% to 72% from October to January on all of them. They separate during 2027, so the early quarters won’t settle which one we’re on.</p>
<h2>What it means for the earlier packs</h2>
<p><a href="https://06.ai/packs/ai-funding-scale">The </a><a href="https://06.ai/packs/ai-funding-scale">funding-scale</a><a href="https://06.ai/packs/ai-funding-scale"> pack</a> adds up to about $1.54T of debt and equity that the buildout needs through 2030. On OpenRouter, demand is not the constraint: spend is growing about 21% a month and tokens about 48%.</p>
<p>What changes is what a dollar buys. A dollar spent on OpenRouter buys 2.0× as many tokens as it did in the first half, because the mix has moved to cheaper models. This data can’t say how much of that is cheaper serving and how much is thinner margins; the models taking share are smaller, so some of it is efficiency. <a href="https://06.ai/packs/token-prices">The </a><a href="https://06.ai/packs/token-prices">token-price</a><a href="https://06.ai/packs/token-prices"> pack</a> found the same pattern in list prices: the price of a top-tier token has held while everything below it falls. It is why the labs’ share of the dollars matters more than their share of the tokens: Anthropic and OpenAI take 61% of the dollars from 14% of the tokens.</p>
<h2>What to watch</h2>
<p><strong>The dollar growth rate in the spring.</strong> Average monthly growth from December to March of about 11% would put spend on the low path, 15% on the base path and 18% on the high one. Single months are noisy: spend fell 2% in August and rose 47% in September.</p>
<p><strong>Anthropic’s share.</strong> If OpenRouter stays near 1.7–2.1% of Anthropic’s revenue, it is a weekly read on Anthropic’s growth ahead of its IPO.</p>
<p><strong>Free and stealth tokens.</strong> When they spike, token counts overstate demand.</p>
<p><strong>Who hosts the open models.</strong> DeepSeek serves 64% of requests for its new V4.1-Flash-1017 itself, against none of the September version. Chinese labs taking back hosting would move dollars back to them.</p>
<h2>What this is anchored on</h2>
<p>Source: OpenRouter (openrouter.ai/rankings), as of 2026-10-19. Licensed under CC BY 4.0. Spend, prices per token and forecasts are our estimates.</p>
<ul><li><strong>Tokens</strong> — Daily tokens for the 50 biggest models plus one row for the rest, Jan 1, 2025–Oct 18, 2026; prompt plus completion, counted with each provider’s tokenizer; public traffic only · <a href="https://openrouter.ai/rankings">OpenRouter rankings-daily dataset, as of 2026-10-19 01:29 UTC. Licensed under CC BY 4.0</a></li><li><strong>Prices</strong> — List price of every model, and of every host serving it, with each host’s requests over the 30 minutes to 01:30 UTC, Oct 19 · <a href="https://openrouter.ai/docs/api/api-reference/endpoints/list-all-endpoints-for-a-model.md">OpenRouter models and endpoints APIs</a></li><li><strong>Use and apps</strong> — Weekly token estimates by category (sampled); tokens and requests of the 100 biggest apps each month; 7-day task classification · <a href="https://openrouter.ai/docs/cookbook/administration/data-api">OpenRouter Data API. Licensed under CC BY 4.0</a></li><li><strong>OpenRouter revenue</strong> — About $19M annualized at the end of 2025 and $50M in March 2026 · <a href="https://sacra.com/research/3t-token-coinbase-of-the-inference-economy/">Sacra, Jun 1</a></li><li><strong>OpenRouter revenue</strong> — $140M annualized in July 2026 and $160M in August (estimates; OpenRouter doesn’t disclose revenue) · <a href="https://sacra.com/c/openrouter/">Sacra</a></li><li><strong>OpenRouter fee</strong> — 5.5% on pay-as-you-go credit purchases · <a href="https://openrouter.ai/blog/tutorials/how-to-get-the-lowest-cost-llm-inference-on-openrouter/">OpenRouter</a></li><li><strong>OpenRouter volume</strong> — 25T tokens a week at the May 26 Series B · <a href="https://pulse2.com/openrouter-113-million-raised-for-ai-model-exchange-platform-as-weekly-volume-reaches-25-trillion-tokens/amp/">Pulse 2.0, May 26</a></li><li><strong>OpenRouter volume</strong> — Forecast at the Series B of more than 1 quadrillion tokens for 2026 · <a href="https://www.implicator.ai/stripe-openrouter-10-billion-talks/">Implicator, Jul 24</a></li><li><strong>Google</strong> — More than 3.2 quadrillion tokens a month across its products, 7× a year earlier · <a href="https://blog.google/innovation-and-ai/sundar-pichai-io-2026/">Google, May 19</a></li><li><strong>Anthropic</strong> — Revenue $4.73B in Q1 2026 and more than $11.5B in Q2 (preliminary) · <a href="https://thenextweb.com/news/anthropic-q2-2026-revenue-11-5-billion-operating-income">Bloomberg, via The Next Web, Aug 15</a></li><li><strong>Anthropic</strong> — Revenue $16.8B in Q3 2026 · <a href="https://www.khaleejtimes.com/business/anthropic-says-revenue-jumped-to-nearly-17b-last-quarter-reports-2">AFP, via Khaleej Times, Oct 15</a></li><li><strong>Anthropic</strong> — IPO targeted for mid-to-late November; bankers reported to be pricing it on ~$190–200B of 2028 revenue · <a href="https://www.thestar.com.my/tech/tech-news/2026/08/15/exclusive-anthropic-ipo-valuation-hinges-on-190-200-billion-2028-revenue-forecast-sources-say">Bloomberg, Oct 1; Reuters, Aug 14</a></li><li><strong>OpenAI</strong> — Revenue $5.7B in Q1 2026 and $6.7B in Q2 · <a href="https://thenextweb.com/news/openai-burned-3-7-billion-q1-2026">The Information, via The Next Web; WSJ</a></li><li><strong>Ox Alpha</strong> — Stealth model revealed as Z.ai’s GLM-5.3-Flash on Aug 26; OpenRouter said it was on track for nearly 6T tokens on Aug 24 · <a href="https://www.digitalapplied.com/blog/openrouter-usage-charts-what-they-measure">Digital Applied; Pasquale Pillitteri</a></li></ul>
<h2>Method</h2>
<p><strong>Tokens</strong> come from OpenRouter’s rankings-daily dataset: the 50 biggest models each UTC day plus one row for the rest, prompt plus completion tokens counted with each provider’s tokenizer. It covers public traffic only. Weeks run Monday to Sunday. June 15 and July 15, 2025 are missing from the dataset and are filled with the average of the days either side. Region is the developer’s home country; Ox Alpha is counted as Z.ai’s, and developers we couldn’t place are in Other.</p>
<p><strong>Spend</strong> prices each model’s tokens at its standard catalog price; variants such as batch are priced as themselves, and free versions and stealth models at $0. We assume 10% of tokens are output and 51% of input is billed at cached rates, the share that reproduces Sacra’s $140M estimate of OpenRouter’s July revenue at its 5.5% fee. The level of spend depends on that assumption; growth rates and shares much less. The catalog price is often the cheapest host’s: weighting hosts’ prices by their share of requests in a 30-minute snapshot on October 19 would add about 17% to October’s spend, mostly on Chinese models. We leave it out because it can’t be applied to earlier months. Before May 2026, 10–27% of tokens came from models no longer in the price list; they are priced at the median for their lab.</p>
<p><strong>The gray line</strong> prices every week’s tokens at the first half’s average of $0.96 per million.</p>
<p><strong>Anthropic and OpenAI</strong> compare estimated spend on each lab’s models through OpenRouter, by calendar quarter, with reported revenue. Q4 scales October 1–18 to a full quarter.</p>
<p><strong>Apps</strong> divide total tokens by total requests across the 100 biggest apps in each calendar month. <strong>Categories</strong> are OpenRouter’s sampled weekly estimates, which can overlap.</p>
<p><strong>The forecast</strong> starts from the October 1–18 level and the monthly growth between July’s average and October’s, 2.8 months apart: 20.9% for dollars and 47.7% for tokens. That rate then halves every n months: growth in month m is g × 2<sup>−(m−1)/n</sup>. The half-life is the main assumption; the slowdown in dollar growth between the last two quarters implies 5.7 months. It inherits the spend method’s assumptions, and the token paths assume the mix keeps moving to cheaper models.</p>
<hr>
<p>Every chart above links to its interactive version on 06.ai, with the data behind it. The daily trackers are at <a href="https://06.ai/trackers">06.ai/trackers</a>.</p>
<p><em>Not investment advice. Estimates are labeled as estimates.</em></p>]]></content:encoded>
</item>
<item>
  <title>Who sets the price of a token?</title>
  <link>https://inthesingularity.ai/packs/token-prices</link>
  <guid isPermaLink="true">https://inthesingularity.ai/packs/token-prices</guid>
  <pubDate>Sat, 17 Oct 2026 12:06:00 GMT</pubDate>
  <description>DeepSeek’s newest model claims top-tier scores on agentic coding at OpenAI’s budget-tier price. What earlier DeepSeek releases did to US token prices, and what this one might do.</description>
  <category>AI infrastructure</category>
  <enclosure url="https://inthesingularity.ai/newsletter/token-prices/cover.png" length="57001" type="image/png" />
  <media:content url="https://inthesingularity.ai/newsletter/token-prices/cover.png" medium="image" type="image/png" />
  <content:encoded><![CDATA[<figure><a href="https://06.ai/packs/token-prices"><img src="https://06.ai/newsletter/token-prices/thumb.png" alt="Who sets the price of a token?" width="421" height="392"></a><figcaption><a href="https://06.ai/packs/token-prices">Interactive version and data on 06.ai →</a></figcaption></figure>
<p>DeepSeek’s V4.1-Flash-1017 costs <strong>$0.15</strong> per million input tokens and <strong>$0.60</strong> per million output tokens in US business hours, about what OpenAI charges for its cheapest model. On DeepSWE, DeepSeek’s headline coding benchmark, it scores above GPT-5.6 Sol and Claude Opus 5, which cost 33× to 42× more per output token.</p>
<ul><li><strong>$0.60</strong> — DeepSeek V4.1-Flash-1017, output. Per million tokens in US hours; input $0.15; $0.30 / $1.20 at peak</li><li><strong>33×</strong> — Cheaper than GPT-5.6 Sol. On output; 27× on input and 133× on cached input, at Sol’s discounted rate</li><li><strong>77.0</strong> — DeepSWE v1.1, DeepSeek’s figures. Claude Opus 5 74.0, GPT-5.6 Sol 73.0</li><li><strong>16B</strong> — Active parameters. Of 552B, when writing output; 8B when reading input</li></ul>
<h2>The price ladder</h2>
<figure><a href="https://06.ai/packs/token-prices#ladder"><img src="https://06.ai/newsletter/token-prices/fig-01.png" alt="The price ladder" width="694" height="408"></a><figcaption>Price per million output tokens, log scale. Filled dots are what a US customer pays today; hollow dots are DeepSeek’s peak-hour rate, Sol’s list price and Gemini Flash’s price after its launch discount ends. DeepSWE scores are DeepSeek’s figures, for the models in its table.</figcaption></figure>
<h2>The gap narrowed after R1, then widened again</h2>
<figure><a href="https://06.ai/packs/token-prices#history"><img src="https://06.ai/newsletter/token-prices/fig-02.png" alt="The gap narrowed after R1, then widened again" width="694" height="462"></a><figcaption>Launch price per million output tokens of each lab’s top model, log scale; DeepSeek at its peak rate. Dashed lines mark DeepSeek releases. Lines connect launches and skip versions at similar prices.</figcaption></figure>
<h2>What DeepSeek says you get for the price</h2>
<p>DeepSeek’s published table, 10 of the 11 benchmarks it reported. Bold marks the best score in each row. It beats GPT-5.6 Sol on 9 of 10 and Claude Opus 5 on 4 of 10; Opus 5 leads on the harder long-horizon tests. Prices are input / output per million tokens.</p>
<ul><li><strong>DeepSWE v1.1</strong> — DeepSeek V4.1-Flash-1017: 77.0 · DeepSeek V4.1-Flash (Sep): 74.2 · Kimi K3: 67.5 · GLM 5.3: 66.9 · GPT-5.6 Sol: 73.0 · Claude Opus 5: 74.0</li><li><strong>Terminal-Bench 4.0</strong> — DeepSeek V4.1-Flash-1017: 42.3 · DeepSeek V4.1-Flash (Sep): 31.2 · Kimi K3: 12.6 · GLM 5.3: 37.9 · GPT-5.6 Sol: 39.9 · Claude Opus 5: 51.8</li><li><strong>ProgramBench</strong> — DeepSeek V4.1-Flash-1017: 29.1 · DeepSeek V4.1-Flash (Sep): 20.3 · Kimi K3: 17.5 · GLM 5.3: 19.0 · GPT-5.6 Sol: 23.0 · Claude Opus 5: 37.0</li><li><strong>NL2Repo-Bench</strong> — DeepSeek V4.1-Flash-1017: 68.2 · DeepSeek V4.1-Flash (Sep): 65.4 · Kimi K3: 58.0 · GLM 5.3: 58.0 · GPT-5.6 Sol: 56.8 · Claude Opus 5: 75.3</li><li><strong>Automation-Bench v1.0.6</strong> — DeepSeek V4.1-Flash-1017: 60.5 · DeepSeek V4.1-Flash (Sep): 54.8 · Kimi K3: 46.7 · GLM 5.3: 48.8 · GPT-5.6 Sol: 45.8 · Claude Opus 5: 50.3</li><li><strong>Agents’ Last Exam</strong> — DeepSeek V4.1-Flash-1017: 35.4 · DeepSeek V4.1-Flash (Sep): 31.8 · Kimi K3: 27.6 · GLM 5.3: 28.5 · GPT-5.6 Sol: 26.7 · Claude Opus 5: 28.6</li><li><strong>Apex Bench v1.1</strong> — DeepSeek V4.1-Flash-1017: 72.1 · DeepSeek V4.1-Flash (Sep): 64.6 · Kimi K3: 47.8 · GLM 5.3: 56.6 · GPT-5.6 Sol: 54.3 · Claude Opus 5: 65.8</li><li><strong>HLE, no tools</strong> — DeepSeek V4.1-Flash-1017: 42.6 · DeepSeek V4.1-Flash (Sep): 36.8 · Kimi K3: 43.5 · GLM 5.3: 42.0 · GPT-5.6 Sol: 44.5 · Claude Opus 5: 56.3</li><li><strong>Chartography, with tools</strong> — DeepSeek V4.1-Flash-1017: 83.0 · DeepSeek V4.1-Flash (Sep): 78.9 · Kimi K3: 68.1 · GPT-5.6 Sol: 79.9 · Claude Opus 5: 84.0</li><li><strong>BabyVision</strong> — DeepSeek V4.1-Flash-1017: 92.5 · DeepSeek V4.1-Flash (Sep): 89.6 · Kimi K3: 85.7 · GPT-5.6 Sol: 88.9 · Claude Opus 5: 94.1</li></ul>
<h2>The model</h2>
<p>V4.1-Flash-1017 is a new version of the V4.1-Flash model DeepSeek released in September, with its weights posted under the MIT license, so anyone can host it. It is large, at 552B parameters plus 196B in a separate memory layer, but sparse: only 8B parameters are active when it reads and 16B when it writes, which is why it is cheap to serve. DeepSeek says it beats September’s version on all 11 benchmarks it reports and gives shorter answers. Its table shows it ahead of GPT-5.6 Sol on 9 of 10 benchmarks and of Claude Opus 5 on 4. Opus 5 still leads on the harder long-horizon tests, Terminal-Bench (51.8 against 42.3), NL2Repo (75.3 against 68.2) and ProgramBench (37.0 against 29.1), and on HLE (56.3 against 42.6). These are DeepSeek’s numbers, several from benchmarks few others report, run on its own agent harness. Independent evaluations aren’t out yet.</p>
<h2>Priced like a budget model</h2>
<p>US business hours fall in DeepSeek’s off-peak window, when it charges $0.15 per million input tokens and $0.60 per million output tokens. That is 27× less than GPT-5.6 Sol on input and 33× less on output, even at Sol’s discounted $4/$20, and 133× less on cached input. Cached input matters most for agents, which re-read the same context at every step. Against OpenAI’s cheapest model, GPT-5.6 Luna at $0.20/$1.20, it costs about the same. DeepSeek is selling a model it benchmarks against Sol and Opus at Luna’s price.</p>
<h2>What earlier DeepSeek releases did</h2>
<p>V2, in May 2024, started a price war in China: Alibaba cut a Qwen model 97%, Baidu made two Ernie models free and Zhipu cut 80%. US prices didn’t move. R1, in January 2025, roughly matched OpenAI’s o1 at $0.55/$2.19 against $15/$60, and Nvidia lost about $589B of market value on January 27. The biggest US cuts came over the following year. OpenAI released the cheaper o3-mini four days after the selloff and cut o3 by 80%, to $2/$8, in June. GPT-5 launched in August at $1.25/$10, and Anthropic took Opus from $15/$75 to $5/$25 in November. By mid-2025 the gap between DeepSeek and OpenAI’s flagship was the narrowest it has been.</p>
<p>2026 went the other way. GPT-5.5 launched the same week DeepSeek previewed V4, at $5/$30, double GPT-5.4. DeepSeek made a 75% cut to V4-Pro permanent in May, and US flagships didn’t follow. US labs cut later and selectively: OpenAI cut its cheapest model, Luna, by 80% in August and put Sol on a three-month discount, and Anthropic priced Opus 5.5 20% below Opus 5. None came near DeepSeek’s prices. DeepSeek itself raised prices in August and added peak-hour rates, a sign it couldn’t serve all the demand at its old prices.</p>
<h2>The pattern</h2>
<p>Each DeepSeek release resets the price of “good enough”. US labs answer by repricing the tier that competes and by temporary discounts, while holding flagship list prices around $4–5 for input and $20–30 for output. The price of a given level of capability falls very fast (Epoch AI measured 9× to 900× a year, depending on the benchmark), but the price of a top-tier token hasn’t. Two things limit DeepSeek’s direct pull: its own serving capacity, and the US companies and government users that won’t send data to a Chinese API. The pressure arrives through US companies hosting the open weights.</p>
<h2>What this one might do</h2>
<p><strong>Mid-tier prices go first.</strong> If independent tests confirm Sol-level agentic scores at Luna-level prices, GPT-5.6 Terra ($2/$12) and Anthropic’s Sonnet-class models are the most exposed. Expect cuts and discounts there, and more bundling into subscriptions such as Codex and Claude Code, where per-token prices are harder to compare.</p>
<p><strong>Cached input gets cheaper.</strong> It costs the labs little to serve, and most of what an agent reads is cached.</p>
<p><strong>Flagship prices are the tell.</strong> Sol’s discount runs at least to November 21. Extending it or making it permanent would be the DeepSeek effect; going back to $5/$30 would mean OpenAI still has pricing power. Google’s plan to double Gemini Flash prices on January 1 is the same test.</p>
<p><strong>DeepSeek is also going after agent developers.</strong> It now sells web search and page fetching to agents at $0.75 per 1,000 calls, free until November 1 when the calls come through Claude Code, Codex or DeepSeek’s own agent tool.</p>
<h2>What it means for the lab-funding and funding-scale packs</h2>
<p>OpenAI’s plan needs revenue to grow from $36B this year to $350B in 2030, and Anthropic’s IPO is being priced on $190–200B of revenue in 2028. Both assume usage grows much faster than prices fall. A model that caps mid-tier prices pushes more of that revenue onto the top tier, where DeepSeek still trails. Each 10% that OpenAI’s revenue falls short of plan adds about $84B to the $116B it still has to raise; try it below. For compute, cheaper tokens have so far meant more demand, not less: DeepSeek is rationing capacity with peak pricing, and V4’s launch barely moved US markets. The risk sits with the labs’ margins and the backlogs that depend on them, such as Oracle’s, Microsoft’s and CoreWeave’s, more than with the chipmakers.</p>
<h2>If token prices take OpenAI’s revenue below plan</h2>
<p>OpenAI’s July 2026 plan from <a href="https://06.ai/packs/ai-labs-funding">the </a><a href="https://06.ai/packs/ai-labs-funding">lab-funding</a><a href="https://06.ai/packs/ai-labs-funding"> pack</a>: $840B of revenue and $278B of cash burn over 2026–30, with $162B in hand. Assumes costs don’t fall with revenue, since much of its compute is contracted.</p>
<p><em>At the default setting (Revenue below plan 0%), from the </em><a href="https://06.ai/packs/token-prices#sens"><em>interactive version</em></a><em>:</em></p>
<ul><li>Cash burn, 2026–30: <strong>$278B</strong></li><li>Still to raise: <strong>$116B</strong></li><li>Total need in the <a href="https://06.ai/packs/ai-funding-scale">funding-scale</a> pack: <strong>$1.54T</strong></li></ul>
<h2>What this is anchored on</h2>
<p>List prices per million tokens as of October 17, 2026, standard rates for short prompts.</p>
<ul><li><strong>V4.1-Flash-1017</strong> — Released Oct 17 as a new version of September’s V4.1-Flash; 552B parameters plus 196B ‘Engram’ parameters, 8B active on input and 16B on output; 1M-token context; $0.30/$1.20 at peak and half off-peak; beats the September version on all 11 benchmarks reported · <a href="https://www.marktechpost.com/2026/10/17/deepseek-just-released-v4-1-flash-1017-scores-77-0-on-deepswe-opens-search-api/">MarkTechPost, Oct 17</a></li><li><strong>V4.1-Flash-1017</strong> — Weights on Hugging Face under the MIT license; Search and Fetch APIs at $0.75 per 1,000 requests, free via DeepSeek Harness, Claude Code or Codex until Nov 1 · <a href="https://aiweekly.co/alerts/deepseek-opens-v41-flash-1017-weights-and-075-search-api">AI Weekly, Oct 17</a></li><li><strong>V4.1-Flash-1017</strong> — $0.15 input, $0.60 output, $0.003 cached input per million tokens · <a href="https://openrouter.ai/deepseek/deepseek-v4.1-flash-1017">OpenRouter</a></li><li><strong>DeepSeek rates</strong> — Peak hours 01:00–04:00 and 06:00–10:00 UTC on weekdays; V4.1-Flash $0.30/$1.20 peak and $0.15/$0.60 off-peak; V4-Pro $1.32/$3.96 at peak · <a href="https://www.yottalabs.ai/post/deepseek-v4-1-flash-pricing-specs-v4-pro-routing-2026">Yotta Labs</a></li><li><strong>GPT-5.6 Sol</strong> — $4 input, $0.40 cached input, $20 output; discounted pricing available at least through Nov 21, 2026 · <a href="https://developers.openai.com/api/docs/models/gpt-5.6-sol.md">OpenAI</a></li><li><strong>GPT-5.6 Sol</strong> — Cut from $5/$30 to $4/$20 for three months on Aug 21; Claude Opus 5 at $5/$25 and Claude Fable 5 at $10/$50 · <a href="https://letsdatascience.com/news/openai-cuts-gpt-56-sol-api-prices-fbfd85f3">Let’s Data Science, Aug 22</a></li><li><strong>GPT-5.6 family</strong> — Terra $2/$12; Luna $0.20/$1.20; general availability Jul 9, 2026 · <a href="https://costbench.com/software/llm-api-providers/gpt-5-6/">CostBench</a></li><li><strong>Claude Opus 5.5</strong> — Released Sep 22, 2026 at $4/$20, against $5/$25 for Opus 5 · <a href="https://metricnexus.ai/blog/claude-opus-5-5-benchmarks-pricing">Metric Nexus</a></li><li><strong>August 2026</strong> — Luna cut 80% to $0.20/$1.20; Gemini 3.7 Flash launched at $0.75/$3.75, set to double on Jan 1; DeepSeek V4-Flash moved from $0.14/$0.28 to $0.22/$0.66 off-peak and $0.44/$1.32 at peak · <a href="https://devtk.ai/en/blog/ai-api-price-war-august-2026/">DevTk.AI</a></li><li><strong>DeepSeek V4-Pro</strong> — Output price raised from $0.87 to $3.96 at peak and $1.98 off-peak on Aug 16; Nvidia lost about $589B of market value on Jan 27, 2025 · <a href="https://www.regardsofwallstreet.com/news/deepseek-price-increase-openai-pause-compute-demand">Regards of Wall Street</a></li><li><strong>DeepSeek V4-Pro</strong> — Launched about Apr 24, 2026 (1.6T parameters, 49B active, open weights) at $3.48 output; 75% discount · <a href="https://thenextweb.com/news/deepseek-v4-pro-price-cut-75-percent">The Next Web, Apr 27</a></li><li><strong>DeepSeek V4-Pro</strong> — 75% cut made permanent May 23; GPT-5.5 at $5/$30 and Claude Opus 4.7 at $5/$25; no US price response · <a href="https://www.thestreet.com/markets/chinese-ai-rival-hits-openai-anthropic-where-it-really-hurts">TheStreet, May 25</a></li><li><strong>DeepSeek V4</strong> — Muted initial market reaction to the Apr 24 preview · <a href="https://letsdatascience.com/news/deepseek-launches-v4-pressures-wall-street-stocks-9d72545c">Let’s Data Science</a></li><li><strong>GPT-5.5</strong> — Launched at $5/$30, double GPT-5.4’s $2.50/$15 · <a href="https://www.aipricing.guru/news/openai-gpt-5-5-launches-pricing-april-2026/">AI Pricing Guru, Apr 24</a></li><li><strong>Opus 4.5</strong> — Released Nov 24, 2025 at $5/$25, against $15/$75 for Opus 4 and 4.1 · <a href="https://claudefa.st/blog/models/claude-opus-4-5">ClaudeFast</a></li><li><strong>GPT-5</strong> — Launched at $1.25/$10, against $15/$75 for Claude Opus 4.1 · <a href="https://techcrunch.com/2025/08/08/openai-priced-gpt-5-so-low-it-may-spark-a-price-war">TechCrunch, Aug 8, 2025</a></li><li><strong>DeepSeek V3.2-Exp</strong> — API prices cut by more than half on Sep 29, 2025, to $0.28 input and $0.42 output · <a href="https://gigazine.net/gsc_news/en/20250930-deepseek-v3-2-exp">GIGAZINE, Sep 30, 2025</a></li><li><strong>o3</strong> — Cut 80% on Jun 10, 2025, from $10/$40 to $2/$8 · <a href="https://venturebeat.com/ai/openai-announces-80-price-drop-for-o3-its-most-powerful-reasoning-model">VentureBeat</a></li><li><strong>R1 vs o1</strong> — DeepSeek-R1 at $0.55/$2.19 (Jan 20, 2025) against o1 at $15/$60 · <a href="https://anotherwrapper.com/tools/llm-pricing/deepseek-r1/o1-2024-12-17">AnotherWrapper, citing both companies’ price pages</a></li><li><strong>o3-mini</strong> — OpenAI released o3-mini on Jan 31, 2025, in the wake of DeepSeek · <a href="https://bnnbloomberg.ca/business/technology/2025/01/31/openai-releases-cheaper-model-o3-mini-in-wake-of-deepseek">BNN Bloomberg</a></li><li><strong>China, 2024</strong> — DeepSeek-V2 at ¥1–2 per million tokens (May 6, 2024); Alibaba cut Qwen-Long 97%, Baidu made Ernie Speed and Lite free, Zhipu cut GLM-3-Turbo 80% · <a href="https://kr-asia.com/llm-prices-hit-rock-bottom-in-china-as-alibaba-cloud-enters-the-fray">KrASIA, May 30, 2024</a></li><li><strong>Price trends</strong> — The price of reaching a fixed level of performance fell 9× to 900× a year, depending on the benchmark · <a href="https://epoch.ai/blog/llm-inference-price-trends">Epoch AI, Mar 2025</a></li></ul>
<h2>Method</h2>
<p><strong>Prices</strong> are list prices per million tokens on October 17, 2026, at standard rates for prompts under each provider’s long-context threshold. DeepSeek charges its peak rate, double the off-peak rate, from 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays; US business hours fall outside both windows. GPT-5.6 Sol is shown at its discounted rate, available at least through November 21, with its list price as the hollow dot. Gemini 3.7 Flash is shown at its launch price, which is set to double on January 1.</p>
<p><strong>Benchmarks</strong> are DeepSeek’s own table. Kimi K3 and GLM 5.3 have no prices here because their pay-as-you-go API prices weren’t published in the sources used.</p>
<p><strong>The price history</strong> uses launch list prices of each lab’s top model: OpenAI’s o1, o3, GPT-5, GPT-5.5 and GPT-5.6 Sol; Anthropic’s Opus line; DeepSeek’s R1, V3.2-Exp, V4-Pro and V4.1-Flash-1017, at peak rates where DeepSeek has them. It is a picture of list prices, not of what any customer paid after discounts.</p>
<p><strong>The funding sensitivity</strong> scales the <a href="https://06.ai/packs/ai-labs-funding">lab-funding</a> pack’s plan: a revenue shortfall adds to cash burn one for one, with no cost savings. The <a href="https://06.ai/packs/ai-funding-scale">funding-scale</a> pack’s total adds the same amount to its five-year figure.</p>
<hr>
<p>Every chart above links to its interactive version on 06.ai, with the data behind it. The daily trackers are at <a href="https://06.ai/trackers">06.ai/trackers</a>.</p>
<p><em>Not investment advice. Estimates are labeled as estimates.</em></p>]]></content:encoded>
</item>
<item>
  <title>Can markets fund the AI buildout?</title>
  <link>https://inthesingularity.ai/packs/ai-funding-scale</link>
  <guid isPermaLink="true">https://inthesingularity.ai/packs/ai-funding-scale</guid>
  <pubDate>Sat, 17 Oct 2026 12:05:00 GMT</pubDate>
  <description>The hyperscaler, neocloud, leveraged-middle and lab-funding packs added up: about $1.54T of debt or equity through 2030, set against government borrowing, mortgages and the markets that would have to supply it.</description>
  <category>AI infrastructure</category>
  <enclosure url="https://inthesingularity.ai/newsletter/ai-funding-scale/cover.png" length="52135" type="image/png" />
  <media:content url="https://inthesingularity.ai/newsletter/ai-funding-scale/cover.png" medium="image" type="image/png" />
  <content:encoded><![CDATA[<figure><a href="https://06.ai/packs/ai-funding-scale"><img src="https://06.ai/newsletter/ai-funding-scale/thumb.png" alt="Can markets fund the AI buildout?" width="421" height="398"></a><figcaption><a href="https://06.ai/packs/ai-funding-scale">Interactive version and data on 06.ai →</a></figcaption></figure>
<p>Thirteen companies need about <strong>$1.54T</strong> of debt or equity over 2026–30, <strong>$435B</strong> of it in 2028. Next to Treasury borrowing and mortgages, that’s small. Next to the new money US businesses borrow each year, and the markets the weakest of these borrowers use, it isn’t.</p>
<ul><li><strong>$1.54T</strong> — Total need, 2026–30. Debt or equity · $1.26T for infrastructure, $278B for the labs</li><li><strong>$435B</strong> — Peak year. 2028 · 1.3% of GDP</li><li><strong>$438B</strong> — Hyperscalers. 28% of the total · peak $155B in 2027</li><li><strong>$476B</strong> — Builders. 31% · Oracle $65B, SpaceX $411B</li><li><strong>$347B</strong> — Neoclouds. 23% · plus $75B of GPU loans to roll</li><li><strong>$278B</strong> — Labs. 18% · all OpenAI; Anthropic is cash-flow positive</li></ul>
<h2>Debt or equity needed, by year</h2>
<figure><a href="https://06.ai/packs/ai-funding-scale#by-year"><img src="https://06.ai/newsletter/ai-funding-scale/fig-01.png" alt="Debt or equity needed, by year" width="694" height="508"></a><figcaption>What capex leaves after operating cash flow and customer prepays, plus the labs’ own cash burn, in calendar years. Each group’s pack has the company detail. Hover or tap a bar for the split. Shown: Street consensus.</figcaption></figure>
<p>Oracle is converted from fiscal years ending May. Neocloud figures are new money only: another $75B of GPU-backed loans comes due over the period and has to be refinanced. The labs’ figure is OpenAI’s own plan; on Anthropic’s own figures it needs no outside money.</p>
<h2>Against government, mortgages and credit markets</h2>
<figure><a href="https://06.ai/packs/ai-funding-scale#vs-markets"><img src="https://06.ai/newsletter/ai-funding-scale/fig-02.png" alt="Against government, mortgages and credit markets" width="694" height="532"></a><figcaption>Shown: Per year.</figcaption></figure>
<figure><a href="https://06.ai/packs/ai-funding-scale#vs-markets"><img src="https://06.ai/newsletter/ai-funding-scale/fig-03.png" alt="Against government, mortgages and credit markets — Outstanding" width="694" height="448"></a><figcaption><strong>Outstanding.</strong></figcaption></figure>
<h2>Each group against the market it borrows in</h2>
<p>The total hides where the money has to come from. Each group’s peak-year need against a year of issuance in the market it mostly draws on. For the neoclouds, the peak includes GPU-backed loans that have to be refinanced that year.</p>
<p>Hyperscalers · High grade</p>
<p><strong>6.8%</strong> · Peak-year need as a share of a year of US investment-grade bonds</p>
<ul><li><strong>Need, 2026–30</strong>: $438B</li><li><strong>Peak year</strong>: $155B in 2027</li><li><strong>That market</strong>: $2.27T a year</li><li><strong>Raises through</strong>: Investment-grade bonds, equity</li></ul>
<p>Builders · BBB / BBB−</p>
<p><strong>4.9%</strong> · Peak-year need as a share of a year of US investment-grade bonds</p>
<ul><li><strong>Need, 2026–30</strong>: $476B</li><li><strong>Peak year</strong>: $110B in 2028</li><li><strong>That market</strong>: $2.27T a year</li><li><strong>Raises through</strong>: Investment-grade bonds, chip financing, equity</li></ul>
<p>Neoclouds · High yield or unrated</p>
<p><strong>12%</strong> · Peak-year need as a share of a year of high-yield bonds, convertibles and direct-lending funds</p>
<ul><li><strong>Need, 2026–30</strong>: $347B + $75B to roll</li><li><strong>Peak year</strong>: $98B in 2027 incl. $8B to roll</li><li><strong>That market</strong>: $799B a year</li><li><strong>Raises through</strong>: GPU-backed loans, high-yield bonds, converts, equity</li></ul>
<p>Labs · Private, unrated</p>
<p><strong>22%</strong> · Peak-year need as a share of a year of US IPOs and share sales</p>
<ul><li><strong>Need, 2026–30</strong>: $278B</li><li><strong>Peak year</strong>: $111B in 2028</li><li><strong>That market</strong>: $502B a year</li><li><strong>Raises through</strong>: Private funding rounds, strategic investors, an IPO</li></ul>
<h2>Small next to Treasuries and mortgages</h2>
<p>At the 2028 peak the need is $435B, about 1.3% of GDP. That is 22% of the $1.99T deficit the federal government ran in FY2026, and 21% of the $2.1T of mortgages lenders will write this year. Debt across every US borrower outside finance (households, businesses and governments) is growing about $4.4T a year, and the AI peak would be 10% of that. Even on stated targets, where SpaceX builds to 10 GW and 2027 reaches $755B, the peak is 38% of one year’s deficit.</p>
<h2>Large next to what’s actually new</h2>
<p>Those comparisons flatter it. Most mortgage lending and bond issuance replaces debt that already exists; the AI need is new money. With mortgage rates near 7%, US mortgage debt grew only $182B over the past year, so the 2028 peak is 2.4× the net growth in all US mortgage debt. If it were all borrowed, it would also be 39% of the $1.1T a year that all US nonfinancial businesses are adding in debt, and 68% on stated targets.</p>
<h2>Hyperscalers are the easy part</h2>
<p>The four need $438B over five years, peaking at $155B in 2027: 6.8% of a year of US investment-grade issuance at the 2026 pace. Hyperscalers as a group have gone from about 2% of that market in 2023–24 to 11% this year, by Vanguard’s count, including Amazon’s $54B sale in March, the largest corporate bond deal on record, and investment-grade spreads have barely moved: 84bp on October 15, 4bp wider than at the end of August. The need is counted before dividends and buybacks. If the four keep paying out, they raise more; if markets tighten, payouts are the first thing they can cut.</p>
<h2>The strain is in three places</h2>
<p><strong>SpaceX on its stated targets.</strong> Building to about 10 GW by end-2027 needs about $370B in 2027 from one BBB-rated issuer. That is roughly what the whole US high-yield market raised in 2025, 6.9× the largest corporate bond deal ever, and 4.3× what SpaceX raised in its IPO, itself a record. On Street numbers it needs $80B that year. Either the build is slower than Musk has described, or chip vendors and partners finance most of it.</p>
<p><strong>The neoclouds.</strong> They need $347B of new money plus $75B of GPU loans to roll, mostly through high-yield bonds, convertibles and private credit. At the 2027 peak that is about 12% of what those markets raise in a year, and convertibles alone carry up to $32B a year, 17% of US convertible issuance. Those markets are already straining. High-yield spreads have widened 60bp since the end of August, to 323bp, after the heaviest month of junk supply this year, while investment grade moved 4bp. CoreWeave’s bonds yield over 15%.</p>
<p><strong>OpenAI.</strong> Its own plan burns $278B over 2026–30, $116B more than the cash it started the year with plus the $122B it raised in March. That has to come from private rounds and, eventually, an IPO that Altman says won’t happen in 2026. At the 2028 peak the need is 22% of what US IPOs and share sales raised at this year’s record pace, and two deals, SpaceX’s IPO and Alphabet’s raise, were about two-thirds of that. Anthropic is the opposite case: on its own figures it already funds itself, and its risk is $518B of mostly non-cancellable compute commitments that need its revenue to keep growing.</p>
<h2>Issuance is already running ahead of the gap</h2>
<p>Goldman counts nearly $600B of AI-linked debt priced this year, against a $259B gap for 2026 here. Part of the difference is issuers raising ahead of need: by Vanguard’s count, hyperscalers had sold about $132B of bonds by the end of July, against a 2026 gap here of $83B for the four, and Alphabet alone raised $85B in June. The rest is borrowers beyond these companies: data-center developers and the lease-backed vehicles that build sites and buy chips for the labs.</p>
<h2>Some of the money goes in a circle</h2>
<p>Since the start of 2025, at least $110B of the labs’ new equity has come from companies they buy compute from. Amazon alone put $39B into the two labs in the first half of 2026, close to its $45B capex gap for the whole year. That doesn’t change the total here. It moves part of the labs’ need onto their suppliers, whose gaps in the <a href="https://06.ai/packs/hyperscaler-capex-funding">hyperscaler</a> pack don’t include these investments.</p>
<h2>Read the total as a floor</h2>
<p>It excludes about $830B of data-center leases the four hyperscalers have signed but not started, which developers finance with their own debt; the debt partners and lessors raise to build and equip the labs’ capacity (almost $100B against OpenAI’s sites by late 2025, and more than $100B of packages this year for chips and sites serving Anthropic, among others); power and grid; and payouts. Broader estimates run higher: Morgan Stanley sees a $1.5T data-center financing gap for 2025–28, and JPMorgan expects more than $2.1T of data-center financing from investment-grade bonds alone over five years.</p>
<h2>The telecom comparison</h2>
<p>The last debt-funded network buildout is the obvious reference. Telecom companies raised about $1.3T of debt in the five years to 2001, by Cambridge Associates’ count, in an economy a third the size of today’s, and by 2001 telecom bonds were nearly a quarter of the US high-yield market. This one needs $1.54T over five years: more in dollars, far less relative to the economy. Where it rhymes is in who is stretching: the weakest borrowers, funded in high yield, selling to a few customers.</p>
<p>So the total is manageable. The risk is in its shape: a few issuers, the weakest credit markets, and revenue that traces back to two AI labs, one of which still has to raise another $116B.</p>
<h2>What the comparisons are anchored on</h2>
<p>Latest data as of October 17, 2026. Company figures come from the <a href="https://06.ai/packs/hyperscaler-capex-funding">hyperscaler</a>, <a href="https://06.ai/packs/neocloud-capex-funding">neocloud</a>, <a href="https://06.ai/packs/leveraged-middle">leveraged-middle</a> and <a href="https://06.ai/packs/ai-labs-funding">lab-funding</a> packs, which cite their own sources.</p>
<ul><li><strong>Government</strong> — Figure: FY2026 deficit $1,993B (preliminary), against $1,775B in FY2025 · <a href="https://www.cbo.gov/publication/61985">CBO, Monthly Budget Review: September 2026</a></li><li><strong>All borrowers</strong> — Figure: Q2 2026: domestic nonfinancial debt $84.1T, growing at a 5.2% annual rate; nonfinancial business $23.97T (4.6%); federal government $34.92T; home mortgages $14.01T · <a href="https://www.federalreserve.gov/releases/z1/20260911/z1.pdf">Federal Reserve, Financial Accounts of the US (Z.1), Q2 2026</a></li><li><strong>Mortgages</strong> — Figure: Mortgage balances $13.117T at Q2 2026, up $182B on the year; Q2 originations $505B · <a href="https://www.newyorkfed.org/newsevents/news/research/2026/20260811">New York Fed, Household Debt and Credit Report, Q2 2026</a></li><li><strong>Mortgages</strong> — Figure: Originations about $2.1T in 2026 and $1.99T in 2027, with rates near 7% · <a href="https://www.housingwire.com/articles/mba-originations-forecast-2027-rates/">MBA forecast, via HousingWire, Oct 11</a></li><li><strong>Investment grade</strong> — Figure: $1.7T issued Jan–Sep 2026, up 31%; $195B in September · <a href="https://www.incomeresearch.com/monthly-fixed-income-market-update-september-2026/">IR+M, Monthly Fixed Income Market Update, September 2026</a></li><li><strong>Investment grade</strong> — Figure: Spread 80bp at Aug 31, 84bp at Oct 15 (ICE BofA US Corporate OAS) · <a href="https://fred.stlouisfed.org/series/BAMLC0A0CM">FRED, BAMLC0A0CM</a></li><li><strong>Hyperscalers</strong> — Figure: Bond sales $93B in 2025 and about $132B in 2026 to July 31; about 11% of IG issuance in 2026, against about 2% in 2023–24 · <a href="https://corporate.vanguard.com/content/corporatesite/us/en/corp/vemo/ai-buildout-comes-to-bond-market.html">Vanguard, The AI buildout comes to the bond market (Aug 19); share as charted by CIO</a></li><li><strong>Investment grade</strong> — Figure: Amazon’s ~$53.8B two-currency sale in March 2026, the largest corporate bond sale on record (Verizon’s $49B in 2013 was the previous one) · <a href="https://www.ifre.com/bonds/2397565/amazon-makes-market-history-with-largest-corporate-bond-sale-ever">IFR</a></li><li><strong>Corporate bonds</strong> — Figure: $12.1T outstanding at Q2 2026 · <a href="https://www.sifma.org/resources/research/statistics/us-corporate-bonds-statistics/">SIFMA, US Corporate Bonds Statistics</a></li><li><strong>High yield</strong> — Figure: Record $219B issued in H1 2026; $364B in 2025. Other providers count less for 2025 ($298–328B) · <a href="https://www.aranca.com/assets/docs/US-High-Yield-Bond-Newsletter-1H-26.pdf">Aranca, US High-Yield Newsletter 1H’26 (Bloomberg data)</a></li><li><strong>High yield</strong> — Figure: $38.5B issued in September, the heaviest month of 2026; Goldman counts nearly $600B of AI-linked debt priced this year · <a href="https://www.briefs.co/?p=22964">Bloomberg, via Briefs, Sep 28</a></li><li><strong>High yield</strong> — Figure: Spread 263bp at Aug 31, 312bp at Sep 30, 323bp at Oct 15 (ICE BofA US High Yield OAS) · <a href="https://fred.stlouisfed.org/series/BAMLH0A0HYM2">FRED, BAMLH0A0HYM2</a></li><li><strong>High yield</strong> — Figure: $1.53T par outstanding at Sep 30, 2026 (ICE index) · <a href="https://www.morganstanley.com/content/dam/im/assets/publication/thought-leadership/article/article_highyieldmarketmonitor_q32026.pdf">Morgan Stanley IM, High Yield Market Monitor Q3 2026</a></li><li><strong>Convertibles</strong> — Figure: $131B raised by US-listed companies this year to early September, a record; 44% AI-related · <a href="https://www.advisorperspectives.com/articles/2026/09/11/convertible-market-record-ai-spending-surges">Bloomberg, via Advisor Perspectives, Sep 11</a></li><li><strong>Private credit</strong> — Figure: Direct-lending funds raised $129.15B in Jan–Sep 2026, up 35% · <a href="https://www.ksl.com/article/51636914/private-credit-faces-deployment-squeeze-as-fundraising-hits-record">Reuters, via KSL, Oct 15</a></li><li><strong>Private credit</strong> — Figure: US direct lending about $1.7T (March 2026) · <a href="https://www.blackrock.com/gls-download/literature/market-commentary/1h2026-credit-currents-chartbook-stamped.pdf">BlackRock, 1H2026 Credit Currents chartbook</a></li><li><strong>Leveraged loans</strong> — Figure: Morningstar LSTA index outstandings a record $1.55T at end-2025 · <a href="https://www.lsta.org/content/morningstar-lsta-leveraged-loan-index-analysis-december-2025/">LSTA</a></li><li><strong>Equity</strong> — Figure: Record $251B of US IPOs and share sales in H1 2026, including SpaceX’s $86.2B IPO (the largest ever) and Alphabet’s $85B raise · <a href="https://www.briefs.co/?p=10024">Bloomberg, via Briefs, Jun 29</a></li><li><strong>Labs</strong> — Figure: Amazon paid $15.0B and $13.7B into OpenAI in Q1 and Q2 2026, and $5B into each of Anthropic’s Series G and H · <a href="https://www.sec.gov/Archives/edgar/data/0001018724/000101872426000026/amzn-20260630.htm">Amazon 10-Q, quarter to Jun 30, 2026</a></li><li><strong>Economy</strong> — Figure: Nominal GDP $32.56T at an annual rate, Q2 2026 · <a href="https://fred.stlouisfed.org/graph/?g=wo6u">BEA, via FRED</a></li><li><strong>Street estimates</strong> — Figure: ~$2.9T of data-center capex 2025–28 and a ~$1.5T financing gap: ~$800B private credit, ~$200B corporate bonds, ~$150B ABS/CMBS, ~$350B other · <a href="https://prod-mssip.morganstanley.com/content/dam/msdotcom/en/assets/pdfs/Research_Bridging-Data-Center-Gap.pdf">Morgan Stanley Research, July 2025</a></li><li><strong>Street estimates</strong> — Figure: More than $2.1T of data-center financing from high-grade bonds over five years · <a href="https://fortune.com/2026/06/25/what-bubble-jpmorgan-5-5-trillion-ai-capex-explosion-profitable-for-now/">JPMorgan, via Fortune, Jun 25</a></li><li><strong>Telecom, 1996–2001</strong> — Figure: About $1.3T of debt raised in the five years to 2001 (unsourced in the report) · <a href="https://publishedresearch.cambridgeassociates.com/wp-content/uploads/2015/01/Telecom-Investing-2002.pdf">Cambridge Associates, Telecom Investing 2002</a></li><li><strong>Telecom, 1996–2001</strong> — Figure: Telecom bonds nearly a quarter of US high-yield bonds outstanding in 2001 · <a href="https://www.bis.org/publ/cgfs19board2.pdf">Federal Reserve Board staff paper, for the BIS CGFS, April 2001</a></li></ul>
<h2>Method</h2>
<p><strong>The AI need</strong> is each pack’s debt-or-equity line summed by calendar year: capex minus customer prepays minus operating cash flow, with no surplus carried between years. Hyperscalers come from the <a href="https://06.ai/packs/hyperscaler-capex-funding">hyperscaler</a> pack and the builders from the <a href="https://06.ai/packs/leveraged-middle">leveraged-middle</a> pack. Oracle’s fiscal years end in May, so calendar year t is 5/12 of FY t plus 7/12 of FY t+1 (FY31 is taken as equal to FY30). That is why Oracle shows $65B here against $79B in the <a href="https://06.ai/packs/leveraged-middle">leveraged-middle</a> pack: seven months of FY26 fell in 2025. Neoclouds are the <a href="https://06.ai/packs/neocloud-capex-funding">neocloud</a> pack’s base case, new money only. The labs are the <a href="https://06.ai/packs/ai-labs-funding">lab-funding</a> pack: OpenAI’s July 2026 plan ($278B over 2026–30, split by year as described there), and zero for Anthropic, which reports adjusted free cash flow above zero. Adding the labs doesn’t double count: what a lab pays a cloud is a cost in its burn and revenue in the cloud’s operating cash flow, so the two cancel out. Supplier investments in the labs move need between groups without changing the total.</p>
<p><strong>Scenarios.</strong> “Street consensus” uses Street capex for the hyperscalers and SpaceX. “Stated targets” uses Meta’s $600B-through-2028 pledge, Amazon’s doubled power by end-2027 and SpaceX building toward ~10 GW by end-2027. Oracle has no separate target, and the neocloud and lab figures are already each company’s own plan, so none of those change.</p>
<p><strong>Benchmarks.</strong> Flows are annual, and partial-year 2026 figures are annualized as labeled. Net flows (the deficit, the change in mortgage debt, net business borrowing) are the like-for-like comparison for new money; gross issuance includes refinancing. The AI need includes equity, so comparing it with borrowing assumes all of it is borrowed. High-yield totals vary by provider: Bloomberg-based counts put 2025 at $364B, others at $298–328B. Mortgage debt growth uses the New York Fed’s household data; debt outstanding uses the Fed’s Financial Accounts, which measure mortgages slightly differently. Several neoclouds are based outside the US but raise mostly in US markets.</p>
<hr>
<p>Every chart above links to its interactive version on 06.ai, with the data behind it. The daily trackers are at <a href="https://06.ai/trackers">06.ai/trackers</a>.</p>
<p><em>Not investment advice. Estimates are labeled as estimates.</em></p>]]></content:encoded>
</item>
<item>
  <title>Who funds the labs?</title>
  <link>https://inthesingularity.ai/packs/ai-labs-funding</link>
  <guid isPermaLink="true">https://inthesingularity.ai/packs/ai-labs-funding</guid>
  <pubDate>Sat, 17 Oct 2026 12:04:00 GMT</pubDate>
  <description>OpenAI and Anthropic are the customers behind the hyperscaler, neocloud and leveraged-middle packs. What each still has to raise to pay for its compute, and how much of that money comes from the companies selling it to them.</description>
  <category>AI infrastructure</category>
  <enclosure url="https://inthesingularity.ai/newsletter/ai-labs-funding/cover.png" length="45321" type="image/png" />
  <media:content url="https://inthesingularity.ai/newsletter/ai-labs-funding/cover.png" medium="image" type="image/png" />
  <content:encoded><![CDATA[<figure><a href="https://06.ai/packs/ai-labs-funding"><img src="https://06.ai/newsletter/ai-labs-funding/thumb.png" alt="Who funds the labs?" width="421" height="373"></a><figcaption><a href="https://06.ai/packs/ai-labs-funding">Interactive version and data on 06.ai →</a></figcaption></figure>
<p>OpenAI and Anthropic, the customers behind the <a href="https://06.ai/packs/hyperscaler-capex-funding">hyperscaler</a>, <a href="https://06.ai/packs/neocloud-capex-funding">neocloud</a> and <a href="https://06.ai/packs/leveraged-middle">leveraged-middle</a> packs, have opposite problems. OpenAI’s own plan burns <strong>$278B</strong> of cash through 2030, <strong>$116B</strong> more than it has in hand. Anthropic reports it is already cash-flow positive, but it has signed <strong>$518B</strong> of compute it mostly can’t cancel.</p>
<h3>OpenAI</h3>
<p><strong>$278B</strong> · Cash burn over 2026–30 in its own July 2026 plan</p>
<ul><li><strong>In hand</strong>: $162B: about $40B of cash at the end of 2025 plus the $122B round that closed in March</li><li><strong>Still to raise</strong>: <strong>$116B</strong>, with the money in hand running out in 2028</li><li><strong>Revenue</strong>: About $50B a year at the end of September (about $70B counted Anthropic’s way); the plan has $350B in 2030</li><li><strong>Compute</strong>: $856B planned on compute and infrastructure through 2030, about $171B a year, 3.4× current revenue</li></ul>
<h3>Anthropic</h3>
<p><strong>$518B</strong> · Compute commitments over about a decade, roughly 80% non-cancellable</p>
<ul><li><strong>Cash flow</strong>: Positive since Q3 2026 on an adjusted basis, by its own figures</li><li><strong>Raised</strong>: $121.5B since 2025, $105B of it this year, plus an IPO of up to $100B planned for November</li><li><strong>Revenue</strong>: $16.8B in Q3 2026, about $67B annualized, against $4.6B in all of 2025</li><li><strong>Compute</strong>: About $52B a year of commitments on average, 0.8× current revenue</li></ul>
<h2>OpenAI keeps raising its own burn forecast</h2>
<figure><a href="https://06.ai/packs/ai-labs-funding#vintages"><img src="https://06.ai/newsletter/ai-labs-funding/fig-01.png" alt="OpenAI keeps raising its own burn forecast" width="694" height="362"></a><figcaption>Cash burn over 2026–30 in each of OpenAI’s reported plans, by when the plan was made. Under each bar is the 2030 revenue the same plan assumed: the burn has grown even as the revenue target has.</figcaption></figure>
<h2>On its own plan, OpenAI runs through what it has raised in 2028</h2>
<figure><a href="https://06.ai/packs/ai-labs-funding#runway"><img src="https://06.ai/newsletter/ai-labs-funding/fig-02.png" alt="On its own plan, OpenAI runs through what it has raised in 2028" width="694" height="422"></a><figcaption>Bars are yearly cash burn: the July plan’s $278B total, split by year in the proportions of February’s plan. The line is cumulative burn. The dashed line is the money in hand at the start of 2026.</figcaption></figure>
<p>Every 10% that revenue falls short of plan over the period adds about $84B to the burn if costs don’t fall with it. Much of the compute is contracted.</p>
<h2>Who writes the checks</h2>
<figure><a href="https://06.ai/packs/ai-labs-funding#checks"><img src="https://06.ai/newsletter/ai-labs-funding/fig-03.png" alt="Who writes the checks" width="694" height="224"></a><figcaption>New equity raised since January 2025, by investor. Suppliers are companies the lab buys compute from. Anthropic’s supplier figure is a floor: Microsoft’s and Nvidia’s shares of its Series G weren’t disclosed. Dashed: what’s coming next.</figcaption></figure>
<h2>Where the debt sits</h2>
<p>Little of it is in the labs’ own names. Most is raised by the partners and lessors who build sites and buy chips for them, and repaid from the labs’ rent.</p>
<ul><li><strong>Credit lines in its own name</strong> — OpenAI: About $4.7B, undrawn; a $520M Bank of America line · Anthropic: $2.5B, being expanded to $15B</li><li><strong>Loans from suppliers</strong> — OpenAI: None reported · Anthropic: Broadcom, up to $42B, convertible into shares, toward $125B of TPU leases</li><li><strong>Debt raised by partners and lessors</strong> — OpenAI: Almost $100B lined up against Stargate sites by late 2025 (Abilene ~$10B, New Mexico $18B, Vantage $38B); $16B for Oracle’s Michigan campus · Anthropic: $35B from Apollo and Blackstone; a $60B Broadcom chip package that also serves others; ~$15B for a campus in Hubbard, Texas</li><li><strong>Supplier money not yet paid in</strong> — OpenAI: $21.3B of Amazon’s $50B, due after June 30 · Anthropic: Google up to $30B; Amazon up to $15B; AMD up to $5B; Nvidia up to $10B in the IPO, in talks</li></ul>
<h2>Where the labs’ spending lands</h2>
<p>The backlogs behind the operating cash flow in the <a href="https://06.ai/packs/hyperscaler-capex-funding">hyperscaler</a>, <a href="https://06.ai/packs/neocloud-capex-funding">neocloud</a> and <a href="https://06.ai/packs/leveraged-middle">leveraged-middle</a> packs, and how much of each is these two labs.</p>
<ul><li><strong>Oracle</strong> — Backlog: $664B (Aug 2026) · From the labs: About half OpenAI · Pack: <a href="https://06.ai/packs/leveraged-middle">leveraged-middle</a></li><li><strong>Microsoft</strong> — Backlog: $625B (Dec 2025) · From the labs: About 45% OpenAI · Pack: <a href="https://06.ai/packs/hyperscaler-capex-funding">hyperscaler</a></li><li><strong>Google Cloud</strong> — Backlog: $514B (Jun 2026) · From the labs: Anthropic’s $200B commitment was over 40% of the Q1 backlog · Pack: <a href="https://06.ai/packs/hyperscaler-capex-funding">hyperscaler</a></li><li><strong>Amazon (AWS)</strong> — Backlog: $496B (Jun 2026) · From the labs: About half OpenAI and Anthropic · Pack: <a href="https://06.ai/packs/hyperscaler-capex-funding">hyperscaler</a></li><li><strong>CoreWeave</strong> — Backlog: $99B (Mar 2026) · From the labs: Under 30% from non-investment-grade AI labs · Pack: <a href="https://06.ai/packs/neocloud-capex-funding">neocloud</a></li><li><strong>SpaceX</strong> · From the labs: Anthropic pays $1.25B a month, cancellable on 90 days’ notice · Pack: <a href="https://06.ai/packs/leveraged-middle">leveraged-middle</a></li></ul>
<h2>OpenAI: the plan needs another $116B, and the plan keeps growing</h2>
<p>OpenAI’s July plan has it burning $278B of cash over 2026–30 while revenue grows from $36B to $350B. It started the year with about $40B and closed a $122B round in March, so on its own plan it runs through that money in 2028 and needs about $116B more. The plan has moved the wrong way: the same five years were $69B of burn in its September 2025 plan and $179B in February’s. And it assumes the revenue arrives. With costs unchanged, every 10% revenue shortfall over the period adds about $84B.</p>
<h2>Half of its new money comes from its suppliers</h2>
<p>Of the $164B OpenAI has raised since the start of 2025, at least $80B came from companies it buys compute from: $50B from Amazon and $30B from Nvidia. Another $60B came from SoftBank, its partner in Stargate. The March round also arrives in installments: Amazon had paid $28.7B of its $50B by the end of June. Supplier money doesn’t reduce what has to be raised. It moves part of it onto the suppliers, and the capex gaps in the <a href="https://06.ai/packs/hyperscaler-capex-funding">hyperscaler</a> pack don’t include these investments.</p>
<h2>Anthropic: revenue caught up, commitments are fixed</h2>
<p>Anthropic’s revenue went from $4.6B in all of 2025 to $16.8B in the third quarter of 2026 alone, and it reported positive adjusted free cash flow for the quarter. On those figures it no longer needs outside money to run the business, but it is raising it anyway: $105B of equity this year and an IPO of up to $100B planned for November. The reason is on the other side of the ledger. It has signed at least $518B of compute over about a decade, about 80% of it non-cancellable, and bankers are pricing the IPO on $190–200B of revenue in 2028. Two customers are about a quarter of its revenue. Revenue comparisons between the two also flatter it: counted Anthropic’s way, OpenAI’s ~$50B run-rate would be ~$70B.</p>
<h2>The debt sits one step away</h2>
<p>Neither lab has borrowed much in its own name. OpenAI’s ~$4.7B credit line is undrawn, and Anthropic’s is growing to $15B. The borrowing happens at the companies that build and equip their capacity. Partners had lined up almost $100B of debt against OpenAI’s Stargate sites by late 2025. Lessors buying chips and building sites for Anthropic are raising $35B from Apollo and Blackstone and about $15B in Texas; Broadcom is arranging a $60B chip package that serves Anthropic among others, and will lend Anthropic up to $42B toward leasing its TPUs. That debt is repaid from the labs’ rent. It carries their credit risk without sitting on their balance sheets, or in the <a href="https://06.ai/packs/hyperscaler-capex-funding">hyperscaler</a>, <a href="https://06.ai/packs/neocloud-capex-funding">neocloud</a> and <a href="https://06.ai/packs/leveraged-middle">leveraged-middle</a> packs.</p>
<h2>Why the hyperscaler, neocloud and leveraged-middle packs depend on this</h2>
<p>About half of Oracle’s $664B backlog is OpenAI. OpenAI was about 45% of Microsoft’s commercial backlog at the end of 2025. Anthropic’s $200B Google commitment was more than 40% of Google Cloud’s backlog, and the two labs are about half of Amazon’s. That backlog is what turns into the operating cash flow in the <a href="https://06.ai/packs/hyperscaler-capex-funding">hyperscaler</a>, <a href="https://06.ai/packs/neocloud-capex-funding">neocloud</a> and <a href="https://06.ai/packs/leveraged-middle">leveraged-middle</a> packs. For OpenAI it depends on raising another $116B or more; for Anthropic, on revenue that keeps growing into commitments it can’t cancel. Added to the <a href="https://06.ai/packs/hyperscaler-capex-funding">hyperscaler</a>, <a href="https://06.ai/packs/neocloud-capex-funding">neocloud</a> and <a href="https://06.ai/packs/leveraged-middle">leveraged-middle</a> packs, the labs’ own need lifts the five-year total to $1.54T: see <a href="https://06.ai/packs/ai-funding-scale">the </a><a href="https://06.ai/packs/ai-funding-scale">funding-scale</a><a href="https://06.ai/packs/ai-funding-scale"> pack</a>.</p>
<h2>What this is anchored on</h2>
<p>Public reporting as of October 17, 2026. Anthropic’s figures come from press accounts of its draft IPO prospectus and investor updates, and are unaudited. OpenAI’s plans are internal projections reported by the press.</p>
<ul><li><strong>OpenAI</strong> — July 2026 plan: negative free cash flow of $278B over 2026–30; revenue $36B in 2026 rising to $350B in 2030 ($840B in total); ~$856B on compute and infrastructure; the March 2026 raise used up by 2028 · <a href="https://www.thestar.com.my/tech/tech-news/2026/09/19/openai-forecasts-cash-burn-near-280-billion-by-2030-ft-reports">FT, via Reuters, Sep 18</a></li><li><strong>OpenAI</strong> — May 2026 projection: ~$305B of burn over 2026–30 · <a href="https://thenextweb.com/news/openai-856bn-compute-vs-600bn-reset-burn-improved">The Next Web, citing the FT</a></li><li><strong>OpenAI</strong> — February 2026 plan: cash flow −$25B (2026), −$57B, −$85B, −$51B, +$39B (2030); 2025 revenue $13.1B, burn ~$8B, ~$40B of cash at year-end · <a href="https://the-decoder.com/openai-adds-111-billion-to-its-cash-burn-forecast-as-ai-costs-spiral-beyond-projections/">The Information, via The Decoder, Feb 21</a></li><li><strong>OpenAI</strong> — September 2025 plan: cash flow −$17B (2026), −$35B, −$47B, −$8B, +$38B (2030); 2030 revenue ~$200B · <a href="https://the-decoder.com/openai-has-reportedly-misjudged-its-cash-burn-by-80-billion/">The Information, via The Decoder, Sep 6, 2025</a></li><li><strong>OpenAI</strong> — 2025 audited: revenue $13.07B, costs $34B, operating loss $20.92B · <a href="https://wheresyoured.at/exclusive-openai-financials">Where’s Your Ed At (verified by the FT), Jun 15</a></li><li><strong>OpenAI</strong> — Q1 2026 revenue $5.7B, burn ~$3.7B, &gt;$73B of cash and securities at quarter-end; Q2 revenue $6.7B · <a href="https://thenextweb.com/news/openai-burned-3-7-billion-q1-2026">The Information, via The Next Web; WSJ, via Outlook Business</a></li><li><strong>OpenAI</strong> — Annualized revenue ~$50B at end-September (~$70B counted the way Anthropic counts cloud-partner sales); expects at least $70B by year-end · <a href="https://www.bloomberglinea.com/negocios/openai-preve-superar-los-us70000-millones-en-ingresos-anualizados-a-finales-de-2026/">Bloomberg, via Bloomberg Línea, Oct 9</a></li><li><strong>OpenAI</strong> — About $1.4T of data-center commitments over the next eight years · <a href="https://techcrunch.com/2025/11/06/sam-altman-says-openai-has-20b-arr-and-about-1-4-trillion-in-data-center-commitments/">TechCrunch, Nov 6, 2025</a></li><li><strong>OpenAI</strong> — $41.0B round fully funded by Dec 2025, $30B from SoftBank and $11B from co-investors · <a href="https://group.softbank/en/news/press/20251231">SoftBank Group, Dec 31, 2025</a></li><li><strong>OpenAI</strong> — $110B committed in February by Amazon ($50B), Nvidia ($30B) and SoftBank ($30B); round closed Mar 31 at $122B and an $852B valuation, ~$3B from individuals · <a href="https://techcrunch.com/2026/03/31/openai-not-yet-public-raises-3b-from-retail-investors-in-monster-122b-fund-raise/">OpenAI; TechCrunch, Mar 31</a></li><li><strong>OpenAI</strong> — Amazon paid $15.0B in Q1 2026 and $13.7B in Q2 toward its $50B, with $21.3B to come after June 30 · <a href="https://www.sec.gov/Archives/edgar/data/0001018724/000101872426000026/amzn-20260630.htm">Amazon 10-Q, quarter to Jun 30</a></li><li><strong>OpenAI</strong> — Talks to raise at least $30B at ~$1.4T; IPO ‘not 2026’ · <a href="https://propakistani.pk/2026/09/30/openai-reportedly-seeks-30-billion-at-1-4-trillion-valuation/amp/">Bloomberg, via ProPakistani, Sep 30; Reuters, Sep 13</a></li><li><strong>OpenAI</strong> — Partners’ borrowing tied to OpenAI sites ‘almost $100 billion’ (Abilene ~$10B, New Mexico $18B, $38B for Vantage in talks); OpenAI’s own credit line undrawn · <a href="https://www.irishtimes.com/business/2025/11/28/openai-partners-amass-100bn-debt-pile-to-fund-its-ambitions/">FT, via Irish Times, Nov 28, 2025</a></li><li><strong>OpenAI</strong> — Credit line expanded to ~$4.7B, undrawn; Bank of America $520M line in July · <a href="https://www.kitco.com/news/off-the-wire/2026-07-08/bofa-extends-first-520-million-loan-openai-ahead-ipo-source-says">TechCrunch, Mar 31; Reuters, via Kitco, Jul 8</a></li><li><strong>Anthropic</strong> — 2025 revenue $4.59B, operating loss $8.06B, $7.33B of compute and infrastructure costs; $20.28B of cash and short-term investments at Dec 31; two customers about a quarter of revenue · <a href="https://www.thestar.com.my/tech/tech-news/2026/09/29/exclusive-anthropic039s-ipo-prospectus-shows-sweeping-ai-vision-surging-costs">Reuters, on the draft IPO prospectus, Sep 28</a></li><li><strong>Anthropic</strong> — At least $518B of compute commitments over about a decade with six partners, ~80% non-cancellable: Broadcom ~$161.2B of equipment leases, Google ≥$111.1B, Amazon $110B, xAI/SpaceX up to $84.5B (cancellable on 90 days), Microsoft $31.4B, AMD &gt;$20B · <a href="https://www.zawya.com/en/business/anthropics-518bln-ai-buildout-hinges-largely-on-deals-that-cannot-be-canceled-filing-shows-1392642">Reuters, on the draft IPO prospectus, Sep 29</a></li><li><strong>Anthropic</strong> — Q3 2026 revenue $16.8B, adjusted operating margin 11%, adjusted free cash flow positive · <a href="https://www.khaleejtimes.com/business/anthropic-says-revenue-jumped-to-nearly-17b-last-quarter-reports-2">AFP, via Khaleej Times, citing US media reports and the FT and Bloomberg, Oct 15</a></li><li><strong>Anthropic</strong> — Q1 2026 revenue $4.73B; Q2 more than $11.5B (preliminary) · <a href="https://thenextweb.com/news/anthropic-q2-2026-revenue-11-5-billion-operating-income">Bloomberg, via The Next Web, Aug 15</a></li><li><strong>Anthropic</strong> — Bankers price the IPO on ~$190–200B of 2028 revenue · <a href="https://www.thestar.com.my/tech/tech-news/2026/08/15/exclusive-anthropic-ipo-valuation-hinges-on-190-200-billion-2028-revenue-forecast-sources-say">Reuters, Aug 14</a></li><li><strong>Anthropic</strong> — Series E $3.5B (Mar 2025), Series F $13B (Sep 2025), Series G $30B at $380B (Feb 2026) · <a href="https://www.datacenterdynamics.com/en/news/anthropic-raises-30bn-from-gic-coatue-d-e-shaw-ventures-mgx-and-others/">Orrick; Anthropic; DatacenterDynamics</a></li><li><strong>Anthropic</strong> — Series H $65B at $965B (May 2026), including $15B previously committed by hyperscalers ($5B from Amazon) · <a href="https://anthropic.com/news/series-h">Anthropic, May 28</a></li><li><strong>Anthropic</strong> — Google: $10B in cash in April, up to $30B more tied to performance targets · <a href="https://news.bgov.com/capital-markets/google-plans-to-invest-up-to-40-billion-in-anthropic">Bloomberg, Apr 24</a></li><li><strong>Anthropic</strong> — Amazon: $5B in Series G and $5B in Series H; milestone facility for further investment reduced to $15B · <a href="https://www.sec.gov/Archives/edgar/data/0001018724/000101872426000026/amzn-20260630.htm">Amazon 10-Q, quarter to Jun 30</a></li><li><strong>Anthropic</strong> — Broadcom to lend up to $42B, convertible into shares, toward a $125.2B five-year TPU lease commitment · <a href="https://www.thestar.com.my/tech/tech-news/2026/10/01/exclusive-broadcom-to-lend-anthropic-up-to-42-billion-to-lease-its-chips-filing-says">Reuters, on the draft prospectus, Oct 1</a></li><li><strong>Anthropic</strong> — Lessor financing for its capacity: $35B initial tranche from Apollo and Blackstone; a $60B Broadcom chip package ($42B senior, $18B junior); ~$15B for a Hubbard, Texas campus · <a href="https://www.advisorperspectives.com/articles/2026/10/02/banks-amass-60-billion-broadcoms-ai-chip">Apollo; Bloomberg, via Advisor Perspectives, Oct 2; Reuters, Jul 30</a></li><li><strong>Anthropic</strong> — Credit line expanding from $2.5B to $15B · <a href="https://www.advisorperspectives.com/articles/2026/09/05/anthropic-finalizing-15-billion-credit-facility">Bloomberg, via Advisor Perspectives, Sep 5</a></li><li><strong>Anthropic</strong> — IPO targeted for mid-to-late November on Nasdaq, raising up to $100B at about $2T; Nvidia in talks to anchor with up to $10B · <a href="https://news.bgov.com/banking-law/anthropic-said-to-target-mega-ipo-before-thanksgiving-holiday">Bloomberg, Oct 1; WSJ, via PYMNTS; Reuters, Sep 11</a></li><li><strong>OpenAI</strong> — Related Digital finalizing $16B of financing for Oracle’s Saline Township, Michigan campus serving OpenAI · <a href="https://news.bloomberglaw.com/private-equity/oracle-data-center-nears-16-billion-financing-after-twisty-path">Bloomberg, via Bloomberg Law, Apr 1</a></li><li><strong>Oracle</strong> — RPO $664B in the quarter to August, roughly half tied to OpenAI · <a href="https://erp.today/oracle-q1-fy27-results-664b-backlog-ai-contracts/">ERP Today, Sep 11</a></li><li><strong>Microsoft</strong> — About 45% of the $625B commercial RPO at Dec 31, 2025 came from OpenAI commitments · <a href="https://cloudwars.com/cloud/microsoft-q2-rpo-soars-110-led-by-openai-commitment-of-281-billion/">Cloud Wars, on Microsoft’s FY26 Q2 slides</a></li><li><strong>Google</strong> — Google Cloud backlog $513.9B at Jun 30, 2026; Anthropic’s $200B commitment was more than 40% of the backlog disclosed for Q1 · <a href="https://www.sec.gov/Archives/edgar/data/0001652044/000165204426000071/goog-20260630.htm">Alphabet 10-Q; Reuters, citing The Information, May 5</a></li><li><strong>Amazon</strong> — AWS backlog ~$496B at Jun 30, 2026; OpenAI and Anthropic about half of the announced backlog · <a href="https://michaelparekh.substack.com/p/ai-anthropic-and-opemai-half-of-2">Amazon 10-Q; Michael Parekh, citing The Information</a></li><li><strong>CoreWeave</strong> — Backlog $99.4B at Mar 31, 2026, of which non-investment-grade AI labs were under 30% · <a href="https://investors.coreweave.com/news/news-details/2026/CoreWeave-Reports-Strong-Second-Quarter-2026-Results/default.aspx">CoreWeave Q2 2026 results</a></li><li><strong>SpaceX</strong> — Anthropic pays $1.25B a month through May 2029; either side can end it on 90 days’ notice · <a href="https://simonwillison.net/2026/May/20/spacex-s1/">SpaceX S-1, via Simon Willison, May 20</a></li></ul>
<h2>Method</h2>
<p><strong>OpenAI’s need</strong> is the $278B of negative free cash flow over 2026–30 in its July 2026 plan, as reported by the FT. That plan’s year-by-year split hasn’t been reported. We keep the February 2026 plan’s 2026 figure ($25B) and spread the rest over 2027–29 in February’s proportions, with 2030 at zero. If 2030 turns cash-positive, as it did in February’s plan, the money needed before then is larger.</p>
<p><strong>Money in hand</strong> is about $40B of cash at the end of 2025 plus the $122B round that closed in March. Part of that round is paid in installments: Amazon had paid $28.7B of its $50B by June 30.</p>
<p><strong>Burn forecasts by vintage.</strong> The September 2025 and February 2026 figures sum each plan’s reported cash flows for 2026–30, including a positive 2030. May and July 2026 are reported totals.</p>
<p><strong>Anthropic</strong> is shown with no outside need on its own figures: it reports positive adjusted free cash flow since Q3 2026. Its cash flow for the first half of 2026 hasn’t been disclosed.</p>
<p><strong>Raised</strong> counts new equity only, not tender offers or buybacks. Supplier money is investment from companies the lab buys compute from. <strong>Compute a year</strong> is OpenAI’s planned compute and infrastructure spending over 2026–30 divided by five, and Anthropic’s signed commitments divided by their roughly ten-year span. One is a spending plan and the other signed contracts, so they are shown for scale only.</p>
<hr>
<p>Every chart above links to its interactive version on 06.ai, with the data behind it. The daily trackers are at <a href="https://06.ai/trackers">06.ai/trackers</a>.</p>
<p><em>Not investment advice. Estimates are labeled as estimates.</em></p>]]></content:encoded>
</item>
<item>
  <title>The leveraged middle</title>
  <link>https://inthesingularity.ai/packs/leveraged-middle</link>
  <guid isPermaLink="true">https://inthesingularity.ai/packs/leveraged-middle</guid>
  <pubDate>Sat, 17 Oct 2026 12:03:00 GMT</pubDate>
  <description>Oracle and SpaceX build like hyperscalers but fund like neoclouds. What each needs from outside, and who’s lending.</description>
  <category>AI infrastructure</category>
  <enclosure url="https://inthesingularity.ai/newsletter/leveraged-middle/cover.png" length="47853" type="image/png" />
  <media:content url="https://inthesingularity.ai/newsletter/leveraged-middle/cover.png" medium="image" type="image/png" />
  <content:encoded><![CDATA[<figure><a href="https://06.ai/packs/leveraged-middle"><img src="https://06.ai/newsletter/leveraged-middle/thumb.png" alt="The leveraged middle" width="421" height="415"></a><figcaption><a href="https://06.ai/packs/leveraged-middle">Interactive version and data on 06.ai →</a></figcaption></figure>
<p>Over five years Oracle needs about <strong>$78.7B</strong> of debt or equity even after <strong>$67.1B</strong> of customer prepays — and has already raised about $67.9B. SpaceX needs <strong>$411B</strong> on Street numbers, against $93.5B of cash.</p>
<ul><li><strong>$1.08T</strong> — Capex, next 5 years. Both · Street consensus</li><li><strong>$526B</strong> — Covered by operating cash flow. Capped at each year's capex</li><li><strong>$67.1B</strong> — Covered by customer prepays. Cash received ahead of service</li><li><strong>$490B</strong> — Needs debt or equity. 45% of capex</li></ul>
<h2>Capex by year, and how it gets paid for</h2>
<figure><a href="https://06.ai/packs/leveraged-middle#chart"><img src="https://06.ai/newsletter/leveraged-middle/fig-01.png" alt="Capex by year, and how it gets paid for" width="694" height="825"></a><figcaption>Oracle on fiscal years ending May (FY27 = Jun 2026–May 2027); SpaceX on calendar years. Scales differ by panel. Strong grey is customer prepays; pink is what prepays and operating cash flow can’t cover; the black tick marks operating cash flow. Shown: Street consensus · Both.</figcaption></figure>
<h2>Cumulative next 5 years by company</h2>
<p>$B. Dividends and buybacks aren't deducted, so the true need is larger where payouts continue.</p>
<figure><a href="https://06.ai/packs/leveraged-middle#breakdown"><img src="https://06.ai/newsletter/leveraged-middle/fig-02.png" alt="Oracle — Capex: $408B · Cash-flow funded: $262B · Customer prepays: $67.1B · Debt / equity: $78.7B · Share: 19% · Peak gap: FY27; SpaceX — Capex: $675B · Cash-flow funded: $264B · Customer prepays: $0B · Debt / equity: $411B · Share: 61% · Peak gap: 2028; Both — Capex: $1.08T · Cash-flow funded: $526B · Customer prepays: $67.1B · Debt / equity: $490B · Share: 45% · Peak gap: 2028" width="694" height="207"></a></figure>
<h2>How credit markets price them</h2>
<p>Ratings, default-swap levels and lender exposure. CoreWeave included as the neocloud reference point.</p>
<ul><li><strong>ORCL</strong> — Rating: BBB- (S&amp;P, cut Jul 2026) · 5-year CDS: 288bp 5y — record, Oct 15 2026 · Debt: $125.3B (Aug 31 2026) · Customer concentration: OpenAI ≈ half of $664B RPO</li><li><strong>SPCX</strong> — Rating: BBB · 5-year CDS: No liquid CDS quoted · Debt: $38.4B principal (Jun 30 2026); $40B chip financing in talks · Customer concentration: Anthropic $1.25B/mo + Google $920M/mo, cancelable on 90 days</li><li><strong>CRWV</strong> — Rating: High yield · 5-year CDS: &gt;1,000bp 5y, Oct 2026 · Debt: 2032 bonds ≈77¢, &gt;15% yield · Customer concentration: Reference point: neocloud</li></ul>
<h2>Oracle: the gap is funded; the counterparty isn’t diversified</h2>
<p>Oracle’s FY27 guide is $90–95B of capex but “not more than $70B” of net cash capex. Customers prepay the difference: $11.4B came in during the first quarter alone, and most new contracts are prepay or bring-your-own-hardware. Add $67.9B of notes, convertible preferred and at-the-market equity raised since FY26, and most of the five-year gap is already financed.</p>
<p>The risk sits elsewhere. Oracle books these prepays with a “significant financing component” — economically, a loan from the customer that it repays in compute. About half of its $664B backlog is one customer, OpenAI. That’s why its CDS hit a record 288bp in October with the funding gap largely closed: the market is pricing who pays, not how much is needed. If you treat the prepays as debt, Oracle’s external need rises to $146B.</p>
<h2>SpaceX: cash now, a cliff if it builds to plan</h2>
<p>On Street numbers SpaceX needs about $411B over five years. The $93.5B of IPO cash covers 2026 and roughly half of 2027. Building to the ~10 GW Musk has described for end-2027 raises the five-year need to $651B, and $426B of that lands in 2026–27 alone, while the build is happening. Against that, the $40B chip financing now in talks is a down payment. There are no prepays: Anthropic and Google pay monthly, and either can walk on 90 days’ notice. The revenue backing this buildout can be cancelled faster than the debt can be repaid.</p>
<h2>What this is anchored on</h2>
<p>Company filings, calls and press reports as of October 17, 2026.</p>
<ul><li><strong>ORCL</strong> — FY26: capex $55.7B, OCF $32.0B (incl. $4.6B customer prepays), FCF −$23.7B · Oracle FY26 results; Investing.com, Aug 2026</li><li><strong>ORCL</strong> — FQ1'27: capex $28.5B, OCF $23.1B incl. $11.4B prepays booked as deferred revenue with a financing component; FCF −$5B · Oracle Q1 FY27 release &amp; 10-Q, Sep 2026</li><li><strong>ORCL</strong> — FY27 capex guide $90–95B with net cash capex 'not more than $70B'; most new contracts via prepay or bring-your-own-hardware · Oracle Q1 FY27 call (CFO Hilary Maxson)</li><li><strong>ORCL</strong> — Raised $43B senior notes + $5B mandatory convertible (FY26) and $19.9B ATM equity (FQ1'27) · Oracle 10-Q, Aug 31 2026</li><li><strong>ORCL</strong> — New RPO won't hit capex or revenue until FY28+; ~half of $664B RPO converts within 36 months · Oracle Q1 FY27 call</li><li><strong>SPCX</strong> — H1'26 capex $28.5B (Q2 $18.4B, of which AI $15.8B); H1 operating cash flow $3.5B · SpaceX Q2 2026 10-Q</li><li><strong>SPCX</strong> — $14.3B deferred revenue relates to Space and Connectivity contracts — no AI-compute prepayment disclosed · SpaceX Q2 2026 10-Q</li><li><strong>SPCX</strong> — Cash $93.5B after ~$86B IPO; $25B senior notes (5.86% avg coupon); seeking $40B ($30B IG + $10B loans) for Nvidia chips · SpaceX 10-Q; FT/Reuters, Oct 6 2026</li><li><strong>SPCX</strong> — &gt;2 GW compute end-2026, 'closer to 10 GW' end-2027 · SpaceX Q2 2026 call (Musk)</li></ul>
<h2>Method</h2>
<p><strong>Oracle</strong> uses fiscal years ending May. FY26 is actual (capex $55.7B; OCF $32.0B split into $27.4B ex-prepay and $4.6B prepays, which reproduces the reported −$23.7B FCF). FY27 uses the $92.5B capex midpoint, with prepays set so net cash capex lands at the $70B ceiling. FY28–30 are our estimates: capex peaks in FY28 as the RPO Oracle says “won’t impact capex until fiscal 28” is built, then falls; prepays fade as fewer new contracts are signed; operating cash flow ramps as about half the backlog converts within three years.</p>
<p><strong>SpaceX</strong> is calendar-year. Capex: Street blends ~$65B for 2026 (H1 $28.5B, H2 at the Q2 run rate) with analyst paths above $100B for 2027; the stated-target path builds toward ~10 GW by end-2027 at about $50B per GW ($400B of capex in 2027), then spends about $150B a year in 2028–30. Musk hasn’t given a target beyond end-2027, so the path doesn’t keep building at the 2027 pace. Operating cash flow is our estimate from H1 2026’s $3.5B and the ~$26B annual run rate of the Anthropic and Google compute contracts; it is the softest number here.</p>
<p><strong>Prepays</strong> appear only where disclosed. SpaceX’s $14.3B of deferred revenue relates to space and connectivity contracts, not AI compute, so it isn’t counted.</p>
<hr>
<p>Every chart above links to its interactive version on 06.ai, with the data behind it. The daily trackers are at <a href="https://06.ai/trackers">06.ai/trackers</a>.</p>
<p><em>Not investment advice. Estimates are labeled as estimates.</em></p>]]></content:encoded>
</item>
<item>
  <title>Who funds the neoclouds?</title>
  <link>https://inthesingularity.ai/packs/neocloud-capex-funding</link>
  <guid isPermaLink="true">https://inthesingularity.ai/packs/neocloud-capex-funding</guid>
  <pubDate>Fri, 16 Oct 2026 12:02:00 GMT</pubDate>
  <description>CoreWeave, Nebius, IREN, Lambda and Nscale need about $623B of capex through 2030. What prepayments and cash flow cover, what has to be borrowed, and what that money costs now.</description>
  <category>AI infrastructure</category>
  <enclosure url="https://inthesingularity.ai/newsletter/neocloud-capex-funding/cover.png" length="48249" type="image/png" />
  <media:content url="https://inthesingularity.ai/newsletter/neocloud-capex-funding/cover.png" medium="image" type="image/png" />
  <content:encoded><![CDATA[<figure><a href="https://06.ai/packs/neocloud-capex-funding"><img src="https://06.ai/newsletter/neocloud-capex-funding/thumb.png" alt="Who funds the neoclouds?" width="421" height="397"></a><figcaption><a href="https://06.ai/packs/neocloud-capex-funding">Interactive version and data on 06.ai →</a></figcaption></figure>
<p>CoreWeave, Nebius, IREN, Lambda and Nscale need about <strong>$623B</strong> of capex through 2030 to hit their own targets. Prepayments and operating cash flow cover <strong>$278B</strong>. The other <strong>$347B</strong> has to come from debt or equity, plus <strong>$75B</strong> of GPU loans to roll, at a time when CoreWeave’s bonds yield 15%.</p>
<ul><li><strong>$623B</strong> — Capex, 2026–30. All five · base case</li><li><strong>$136B</strong> — Covered by operating cash flow. 22% of capex; $54B in 2030</li><li><strong>$142B</strong> — Covered by customer prepays. 23% of capex</li><li><strong>$347B</strong> — Needs debt or equity. 56% of capex · peak 2027</li><li><strong>$422B</strong> — Gross issuance. Incl. $75B of GPU loans to roll</li></ul>
<h2>Capex by year, and how it gets paid for</h2>
<figure><a href="https://06.ai/packs/neocloud-capex-funding#chart"><img src="https://06.ai/newsletter/neocloud-capex-funding/fig-01.png" alt="Capex by year, and how it gets paid for" width="824" height="472"></a><figcaption>Shown: All five · Funding sources.</figcaption></figure>
<p><em>Sliders at their defaults: Capex per GW 100% · Prepay coverage 100% · EBITDA vs base 100%.</em></p>
<figure><a href="https://06.ai/packs/neocloud-capex-funding#chart"><img src="https://06.ai/newsletter/neocloud-capex-funding/fig-02.png" alt="Capex by year, and how it gets paid for — By funding type" width="824" height="502"></a><figcaption><strong>By funding type.</strong></figcaption></figure>
<h3>By company</h3>
<figure><a href="https://06.ai/packs/neocloud-capex-funding#chart"><img src="https://06.ai/newsletter/neocloud-capex-funding/fig-03.png" alt="Capex by year, and how it gets paid for — CoreWeave" width="824" height="502"></a><figcaption><strong>CoreWeave</strong>: Capex, 2026–30 $234B · Covered by operating cash flow $62B · Covered by customer prepays $24B · Needs debt or equity $148B · Gross issuance $188B</figcaption></figure>
<figure><a href="https://06.ai/packs/neocloud-capex-funding#chart"><img src="https://06.ai/newsletter/neocloud-capex-funding/fig-04.png" alt="Capex by year, and how it gets paid for — Nebius" width="824" height="502"></a><figcaption><strong>Nebius</strong>: Capex, 2026–30 $146B · Covered by operating cash flow $21B · Covered by customer prepays $56B · Needs debt or equity $69B · Gross issuance $71B</figcaption></figure>
<figure><a href="https://06.ai/packs/neocloud-capex-funding#chart"><img src="https://06.ai/newsletter/neocloud-capex-funding/fig-05.png" alt="Capex by year, and how it gets paid for — IREN" width="824" height="502"></a><figcaption><strong>IREN</strong>: Capex, 2026–30 $106B · Covered by operating cash flow $25B · Covered by customer prepays $33B · Needs debt or equity $48B · Gross issuance $61B</figcaption></figure>
<figure><a href="https://06.ai/packs/neocloud-capex-funding#chart"><img src="https://06.ai/newsletter/neocloud-capex-funding/fig-06.png" alt="Capex by year, and how it gets paid for — Lambda" width="824" height="502"></a><figcaption><strong>Lambda</strong>: Capex, 2026–30 $74B · Covered by operating cash flow $20B · Covered by customer prepays $7B · Needs debt or equity $46B · Gross issuance $54B</figcaption></figure>
<figure><a href="https://06.ai/packs/neocloud-capex-funding#chart"><img src="https://06.ai/newsletter/neocloud-capex-funding/fig-07.png" alt="Capex by year, and how it gets paid for — Nscale" width="824" height="502"></a><figcaption><strong>Nscale</strong>: Capex, 2026–30 $64B · Covered by operating cash flow $8B · Covered by customer prepays $22B · Needs debt or equity $36B · Gross issuance $48B</figcaption></figure>
<h2>Company by company</h2>
<figure><a href="https://06.ai/packs/neocloud-capex-funding#by-company"><img src="https://06.ai/newsletter/neocloud-capex-funding/fig-08.png" alt="Company by company" width="824" height="354"></a><figcaption>Same scale for all five; the gridline marks $30B. Each header shows the 2026–30 debt or equity need and its share of capex. Click a company to focus the page on it.</figcaption></figure>
<h2>The gap peaks in 2027, then cash flow takes over</h2>
<p>In the base case, net new money peaks at $90B in 2027. Operating cash flow is only $3B in 2026, with Nscale still losing money and IREN about break-even, but it reaches $54B by 2030 and covers 22% of capex over the five years. CoreWeave is $148B of the $347B the group needs.</p>
<h2>Most of that cash is already spoken for</h2>
<p>The operating cash flow here is after $61B of interest and excludes $69B of revenue customers prepaid in earlier years. Much of it comes from capacity financed with amortizing GPU loans, which get paid down first: $75B of scheduled amortization has to be raised again, which takes gross issuance to $422B. The 2026 need is roughly covered already: the five raised or committed about $76B this year against a $68B net need. The open question is 2027–30.</p>
<h2>What the money costs</h2>
<p>Rate at issue on each 2026 raise, with earlier deals for reference. Rings show where CoreWeave’s bonds trade now and what Nebius’ accreting converts yield if held to maturity. Floating loans are shown at SOFR of 3.88% plus their spread.</p>
<p><strong>5.9–6.9% </strong>Secured GPU debt rated or priced investment grade · <strong>8.4–9.4% </strong>Secured GPU debt below investment grade · <strong>15.1% </strong>CoreWeave 9.75% 2031s on Oct 16, issued at par in April · <strong>0.5–4.5% </strong>Convert cash coupons; 3.2–7.1% for Nebius with accretion</p>
<figure><a href="https://06.ai/packs/neocloud-capex-funding#costs"><img src="https://06.ai/newsletter/neocloud-capex-funding/fig-09.png" alt="What the money costs" width="824" height="814"></a></figure>
<p>Convertible coupons understate the cost: the rest is paid in conversion optionality and dilution. Nscale’s $3.36B pre-IPO converts (September, including $1B from Nvidia) are left off because no coupon has been disclosed; they convert at a double-digit discount to the IPO price, capped at a $30B valuation.</p>
<h2>What they raised in 2026</h2>
<figure><a href="https://06.ai/packs/neocloud-capex-funding#raised"><img src="https://06.ai/newsletter/neocloud-capex-funding/fig-10.png" alt="What they raised in 2026" width="824" height="302"></a><figcaption>Facilities and offerings announced Jan 1 to Oct 16, 2026, at committed size, excluding customer prepayments. Each company’s mix here drives the “By funding type” view above.</figcaption></figure>
<h2>Secured debt got cheaper. Unsecured debt didn’t.</h2>
<p>GPU loans backed by investment-grade customers now price at 5.9–6.9%, against about 15% on CoreWeave’s first GPU loan and 13–15% on Nscale’s earlier private credit. Unsecured paper went the other way. CoreWeave’s 9.75% 2031s, issued at par in April, traded at 81.8 on October 16 for a 15.1% yield, and its five-year CDS hit a record 1,194.5 bp the day before. Its September convert needed a 2.875% coupon and a 22.5% premium, against 1.75% and 30% in April.</p>
<h2>A third of the need leans on converts</h2>
<p>If each company keeps its 2026 financing mix, $139B of the $347B comes as secured GPU debt and $123B as convertibles. Nebius drives that: 93% of what it raised this year was convertible. The cash coupons look cheap at 0.5–4.5%, but the notes accrete to 110–125% at maturity, a 3.2–7.1% yield if they never convert, plus the dilution if they do. That channel depends on the share prices, and CoreWeave and Nebius each lost about a fifth of their value this week.</p>
<h2>The numbers</h2>
<p>$B, calendar years, for the selection above. Reflects the sliders.</p>
<figure><a href="https://06.ai/packs/neocloud-capex-funding#numbers"><img src="https://06.ai/newsletter/neocloud-capex-funding/fig-11.png" alt="Capex — 2026: 93.5 · 2027: 134.7 · 2028: 143.5 · 2029: 133.7 · 2030: 117.4 · 2026–30: 622.8; Customer prepays — 2026: 23.1 · 2027: 32.6 · 2028: 34.1 · 2029: 29.4 · 2030: 22.5 · 2026–30: 141.6; Operating cash flow applied — 2026: 3.4 · 2027: 12.9 · 2028: 26.1 · 2029: 39.7 · 2030: 54.0 · 2026–30: 136.2; Debt / equity needed — 2026: 68.0 · 2027: 89.9 · 2028: 83.4 · 2029: 64.6 · 2030: 40.9 · 2026–30: 346.7; of which secured GPU debt — 2026: 27.9 · 2027: 37.2 · 2028: 31.7 · 2029: 24.5 · 2030: 17.9 · 2026–30: 139.1; of which unsecured notes — 2026: 6.1 · 2027: 6.5 · 2028: 6.3 · 2029: 5.8 · 2030: 4.7 · 2026–30: 29.4; of which convertibles — 2026: 24.5 · 2027: 32.2 · 2028: 32.4 · 2029: 23.3 · 2030: 10.1 · 2026–30: 122.6; of which equity — 2026: 9.4 · 2027: 14.0 · 2028: 13.0 · 2029: 11.1 · 2030: 8.2 · 2026–30: 55.6; GPU loans to roll — 2026: 2.6 · 2027: 8.2 · 2028: 15.7 · 2029: 22.0 · 2030: 26.9 · 2026–30: 75.4; Gross issuance — 2026: 70.6 · 2027: 98.1 · 2028: 99.0 · 2029: 86.6 · 2030: 67.7 · 2026–30: 422.0; Memo: revenue — 2026: 19.8 · 2027: 53.1 · 2028: 94.6 · 2029: 137.3 · 2030: 174.6 · 2026–30: 479.3; Memo: EBITDA — 2026: 9.2 · 2027: 27.8 · 2028: 52.5 · 2029: 76.8 · 2030: 97.8 · 2026–30: 264.1; Memo: cash interest — 2026: 4.1 · 2027: 8.2 · 2028: 12.7 · 2029: 16.4 · 2030: 19.3 · 2026–30: 60.7; Memo: prepaid revenue excluded — 2026: 2.7 · 2027: 7.3 · 2028: 13.8 · 2029: 20.6 · 2030: 24.5 · 2026–30: 68.9; Memo: debt outstanding — 2026: 90.3 · 2027: 166.2 · 2028: 236.6 · 2029: 290.2…" width="824" height="670"></a></figure>
<h3>Cumulative 2026–30 by company</h3>
<figure><a href="https://06.ai/packs/neocloud-capex-funding#numbers"><img src="https://06.ai/newsletter/neocloud-capex-funding/fig-12.png" alt="CoreWeave — Capex: $234B · Prepays: $24B · Cash-flow funded: $62B · Debt / equity: $148B · Share: 63% · Peak year: 2027 · GPU loans to roll: $40B; Nebius — Capex: $146B · Prepays: $56B · Cash-flow funded: $21B · Debt / equity: $69B · Share: 47% · Peak year: 2028 · GPU loans to roll: $2B; IREN — Capex: $106B · Prepays: $33B · Cash-flow funded: $25B · Debt / equity: $48B · Share: 45% · Peak year: 2027 · GPU loans to roll: $13B; Lambda — Capex: $74B · Prepays: $7B · Cash-flow funded: $20B · Debt / equity: $46B · Share: 63% · Peak year: 2028 · GPU loans to roll: $8B; Nscale — Capex: $64B · Prepays: $22B · Cash-flow funded: $8B · Debt / equity: $36B · Share: 56% · Peak year: 2027 · GPU loans to roll: $12B; All five — Capex: $623B · Prepays: $142B · Cash-flow funded: $136B · Debt / equity: $347B · Share: 56% · Peak year: 2027 · GPU loans to roll: $75B" width="824" height="310"></a></figure>
<h2>What this is anchored on</h2>
<p>Guidance, filings and market data as of October 16, 2026.</p>
<ul><li><strong>CRWV</strong> — 2026 capex $35–39B; revenue $12.4–13.2B; ≥1.85 GW active power by year-end; ≥8 GW by 2030; 4.2 GW contracted · CoreWeave Q2 2026 call, Aug 11 2026</li><li><strong>CRWV</strong> — Q2 adj. EBITDA $1.51B (59%); net interest $640M; deferred revenue $9.7B at Jun 30; H1 operating cash flow $3.66B · CoreWeave Q2 2026 results</li><li><strong>CRWV</strong> — DDTL 4.0 $8.5B (A3, ~5.9% fixed / SOFR+225); DDTL 5.0 $3.1B (Ba2, SOFR+450); DDTL 5.5 $2.6B (SOFR+550) · Company releases Mar 31 and May 18 2026; Stocktitan debt summary</li><li><strong>CRWV</strong> — 9.75% 2031 notes $2.75B (Apr); 9.625% 2032 $1.25B and 8.50% € 2032 €2B (Jun); converts 1.75% 2032 $4.0B (Apr), 2.875% 2033 $4.2B (Sep) · Company releases, Apr–Sep 2026</li><li><strong>CRWV</strong> — 9.75% 2031s at 81.8 / 15.08% yield (Oct 16); 5-year CDS record 1,194.5 bp (Oct 15); Nvidia $2B equity (Jan) · ICE via Bloomberg/Benzinga, Oct 16 2026; Bloomberg, Jan 26 2026</li><li><strong>NBIS</strong> — 2026 capex $20–25B; revenue $3.0–3.4B; ~40% adj. EBITDA margin; 0.8–1 GW connected and 5 GW contracted by year-end; &gt;$9B prepayments in 2026 covering 50–60% of associated capex · Nebius Q1/Q2 2026 results and calls</li><li><strong>NBIS</strong> — Converts $4.34B 1.25% 2031 / 2.625% 2033 accreting to 120% (Mar); $5.75B 0.50% 2030 (110%) / 4.50% 2034 (125%) (Aug); first secured loan ~$775M at SOFR+250 (Jul) · Nebius 6-K filings, Mar and Aug 2026</li><li><strong>IREN</strong> — FY27 capex $25–30B; 0.3 GW IT in 2026, 0.8 GW cumulative in 2027; $4B contracted run rate for 2026 capacity; prepays 45–55% of GPU capex · IREN FY26 results and call, Aug 27 2026</li><li><strong>IREN</strong> — $3.65B investment-grade GPU financing at 6.0% (Jun); $2.4B Blue Owl/PIMCO at 9.0% (Aug); $3.0B 1.00% 2033 converts (May) · IREN releases, May–Aug 2026</li><li><strong>Lambda</strong> — 3 GW by 2030 target; backlog $15B (Jun) to $50B (Sep); 2026 revenue above $1.5B; raising up to $4B pre-IPO · WSJ via TechCrunch, Oct 6 2026; Bloomberg</li><li><strong>Lambda</strong> — $926M term loan B (Baa2, SOFR+300, Aug); $1.008B DDTL at 6.78% fixed (Baa1, Oct); $1B secured facility (May) · Lambda releases, May–Oct 2026</li><li><strong>Nscale</strong> — $103.4B contracted (Microsoft ≤$43.8B, Anthropic ≤$44.6B for 460 MW); $24B equipment + $3.5B construction committed, mainly 2026–27; H1 capex $3.2B; deferred revenue $2.0B to $6.5B · Nscale S-1, Sep 18 2026</li><li><strong>Nscale</strong> — $3.36B pre-IPO converts incl. $1B Nvidia (Sep); newest facilities SOFR+237.5 (IG); legacy Sandton loans 13–15% · Nscale release Sep 25 2026; S-1</li><li><strong>Rates</strong> — SOFR 3.88% and fed funds 3.75–4.00% after a September hike; 10-year UST 5.12%; US high-yield index effective yield ~8.1% · Fed/market data, Oct 13–16 2026</li></ul>
<h2>Method</h2>
<p><strong>Capex.</strong> 2026 is company guidance at midpoints, with IREN converted from its June fiscal year and Nscale built from the commitments in its S-1. Later years turn each company’s stated power target into capacity added per year, priced at a company-specific capex per GW. Lambda’s whole path is our estimate.</p>
<p><strong>Prepayments</strong> are gross cash received, from guidance (Nebius, IREN), filings (CoreWeave, Nscale) or an assumed share of capex.</p>
<p><strong>Operating cash flow</strong> is revenue × adjusted EBITDA margin, less cash interest on the debt the model raises (solved year by year), less revenue a customer prepaid in an earlier year. No cash taxes (bonus depreciation, losses) and no working-capital swings. It is the biggest swing factor in 2028–30; the EBITDA slider flexes it.</p>
<p><strong>Debt or equity needed</strong> is what’s left, split by each company’s 2026 financing mix. Scheduled amortization of secured GPU loans (about five years) is shown as refinancing on top. Bullet maturities inside the window, such as CoreWeave’s 2030 notes and Nebius’ February 2030 converts if they don’t convert, are not included.</p>
<p><strong>Not included:</strong> GPU refresh as 2024–25 fleets come off their first contracts; leased powered shells, whose rent is already inside EBITDA; and Crusoe, whose Abilene campus sits in a separately financed joint venture. Power units differ: CoreWeave counts active power, Nebius connected power, IREN critical IT load.</p>
<h3>CoreWeave CRWV</h3>
<p>At least 8 GW active by 2030 (1.85 GW guided for year-end 2026; 4.2 GW contracted in August).</p>
<ul><li>Capex: 2026 at the $35–39B guidance midpoint. Then +6.15 GW to reach 8 GW (1.35 GW in 2027, then 1.5 / 1.6 / 1.7 GW) at about $32B per GW; most data centers are leased.</li><li>Prepayments counted gross: about $2B arrived in H1 2026, so 2026 is set at $4.0B, then 10% of capex.</li><li>Cash flow: $12.8B of revenue in 2026, then about $9.5B per GW-year of average active power (consistent with the ~$30B exit run rate pointed to for 2027). 60% adjusted EBITDA margin (Q2: 59%). 8% cash interest on all debt. The $8.2B deferred-revenue balance comes out over four years.</li><li>2026 raised: $14.2B secured, $6.3B unsecured notes, $8.2B converts, $3B equity.</li></ul>
<h3>Nebius NBIS</h3>
<p>5 GW contracted by year-end 2026, nearly all online within 2 to 3.5 years; 0.8–1 GW connected at year-end 2026; more than 1 GW of new capacity in 2027.</p>
<ul><li>Capex: 2026 at the $20–25B guidance midpoint. Then about 4.1 GW to reach roughly 5 GW, phased 1.1 / 1.4 / 1.1 / 0.5 GW so it lands by early 2030, at about $30B per GW (more than 75% owned sites).</li><li>Prepayments: more than $9B guided for 2026, then 38% of capex (about 70% of Q2 deals carried prepays covering 50–60% of capex).</li><li>Cash flow: $3.2B of revenue in 2026, then about $8B per GW-year of average connected power. EBITDA margin 40% (guided), 45% in 2027, 50% after. 3% cash interest; convert accretion is paid at maturity, not counted here.</li><li>2026 raised: $10.1B of converts and a first $0.78B secured loan.</li></ul>
<h3>IREN IREN</h3>
<p>0.3 GW IT delivered in 2026 and 0.8 GW IT cumulative in 2027; more than 5 GW of secured power; no stated pace after 2027.</p>
<ul><li>Capex: FY27 (Jul 2026–Jun 2027) guided at $25–30B. Calendar 2026 ≈ $3.3B H1 actual plus about half of FY27 ≈ $16B; calendar 2027 ≈ $25.5B. 2028–30 assumes 0.5 GW IT a year at about $43B per GW IT.</li><li>Prepayments at 33% of capex: recent prepays fund 45–55% of GPU capex, and GPUs are about two-thirds of the total.</li><li>Cash flow: about $1.4B of revenue in calendar 2026, then about $13B per GW-year of IT load ($4B of contracted run rate on 0.3 GW). EBITDA margin 35% in 2026, 60–65% after (owned sites, cheap power). 5.5% cash interest.</li><li>2026 raised: $6.45B secured GPU financing, $3.0B converts, about $2.5B equity.</li></ul>
<h3>Lambda private</h3>
<p>3 GW of compute under management by 2030 (private; no capex disclosure).</p>
<ul><li>Capex: backlog went from $15B in June to $50B in September, largely one AI lab, so the path is front-loaded: $6B / $13B / $17B / $19B / $18.5B at about $25B per GW. Prepayments assumed at 10% of capex.</li><li>Cash flow: $1.6B of revenue in 2026 (reported above $1.5B), then about $9B per GW-year. EBITDA margin 30% rising to 45%, lower than owners of their own sites because colocation rent is an operating cost. 7% cash interest.</li><li>2026 raised: $2.9B secured (August and October facilities rated investment grade), plus up to $4B of equity in progress.</li></ul>
<h3>Nscale private</h3>
<p>$103.4B contracted (Microsoft up to $43.8B; Anthropic up to $44.6B for 460 MW at Monarch, WV); 1.37 GW active and contracted; Monarch to 2 GW by H1 2028.</p>
<ul><li>Capex, from the S-1: $3.2B in H1 2026 plus $24B of equipment and $3.5B of construction committed, payable mainly in 2026–27, so about $12B in 2026 and $20B in 2027. 2028 covers the Anthropic and Monarch build (~$15B); 2029–30 assumed at $8–9B a year.</li><li>Prepayments average about 23% of contract value at signing; deferred revenue went from $2.0B to $6.5B in H1, so 2026 is set at $6B and later years at 30% of capex.</li><li>Cash flow: revenue from the $103.4B of contracts as they deliver, about $0.75B / $5B / $11B / $16B / $18B. EBITDA margin −50% in 2026 (H1: −142%), 35% in 2027, 50–55% after. 6.5% cash interest. Early losses add to the external need.</li><li>The S-1 says no binding financing is in place yet for the Anthropic contracts.</li></ul>
<p><strong>Source links</strong></p>
<ul><li><a href="https://finance.yahoo.com/quote/CRWV/earnings/CRWV-Q2-2026-earnings_call-662772.html">CoreWeave Q2 2026 call</a></li><li><a href="https://investors.coreweave.com/news/news-details/2026/CoreWeave-Reports-Strong-Second-Quarter-2026-Results/default.aspx">CoreWeave Q2 2026 results</a></li><li><a href="https://www.theglobeandmail.com/investing/markets/stocks/CRWV-Q/pressreleases/3917877/coreweave-crwv-q2-2026-earnings-call-transcript/">CoreWeave 8 GW target</a></li><li><a href="https://investors.coreweave.com/news/news-details/2026/CoreWeave-Closes-Landmark-8-5-Billion-Financing-Facility-Achieving-First-Investment-Grade-Rated-GPU-backed-Financing/default.aspx">CoreWeave DDTL 4.0</a></li><li><a href="https://investors.coreweave.com/news/news-details/2026/CoreWeave-Closes-3-1-Billion-Loan-Facility-Expanding-Access-to-Public-Markets-for-GPU-Backed-Financing/default.aspx">CoreWeave DDTL 5.0</a></li><li><a href="https://investors.coreweave.com/news/news-details/2026/CoreWeave-Announces-Upsize-and-Pricing-of-1750-Million-of-Senior-Notes/default.aspx">CoreWeave 9.75% 2031 notes</a></li><li><a href="https://www.businesswire.com/news/home/20260611950890/en/CoreWeave-Announces-Pricing-of-$1.25-Billion-of-Senior-Notes-and-%E2%82%AC2-Billion-of-Senior-Notes">CoreWeave June notes</a></li><li><a href="https://investors.coreweave.com/news/news-details/2026/CoreWeave-Prices-Upsized-3-5-Billion-Convertible-Senior-Notes-Offering/default.aspx">CoreWeave April converts</a></li><li><a href="https://blockspace.media/insight/coreweave-closes-4-2-billion-convertible-notes-offering/">CoreWeave September converts</a></li><li><a href="https://www.stocktitan.net/articles/coreweave-nebius-ai-cloud-debt-when-it-comes-due">CoreWeave and Nebius debt stack</a></li><li><a href="https://www.benzinga.com/markets/tech/26/10/62412918/coreweave-bonds-distress-ai-darwin-moment-hartnett">CoreWeave CDS and bond prices, Oct 16</a></li><li><a href="https://finance.yahoo.com/markets/options/articles/coreweave-ai-bonds-tumble-raising-111700097.html">CoreWeave bonds and the HY index</a></li><li><a href="https://www.fool.com/earnings/call-transcripts/2026/08/19/nebius-nbis-q2-2026-earnings-call-transcript/">Nebius Q2 2026 call</a></li><li><a href="https://www.sec.gov/Archives/edgar/data/1513845/000110465926029863/tm268409d3_ex99-1.htm">Nebius March 2026 converts</a></li><li><a href="https://www.sec.gov/Archives/edgar/data/0001513845/000110465926098924/tm2623617d1_ex99-1.htm">Nebius August 2026 converts</a></li><li><a href="https://www.globenewswire.com/news-release/2026/08/27/3352401/0/en/iren-reports-fy26-results.html">IREN FY26 results</a></li><li><a href="https://www.marketbeat.com/instant-alerts/transcript-iren-q4-earnings-call-highlights-2026-08-27/">IREN FY27 capex guidance</a></li><li><a href="https://www.globenewswire.com/news-release/2026/06/01/3304211/0/en/iren-closes-3-65bn-investment-grade-gpu-financing.html">IREN investment-grade GPU financing</a></li><li><a href="https://www.globenewswire.com/news-release/2026/05/14/3295320/0/en/iren-closes-3-0-billion-convertible-notes-offering.html">IREN May 2026 converts</a></li><li><a href="https://lambda.ai/blog/lambda-prices-926-million-senior-secured-term-loan-b-facility">Lambda term loan B</a></li><li><a href="https://lambda.ai/blog/lambda-closes-1-billion-senior-secured-fixed-rate-financing">Lambda fixed-rate financing</a></li><li><a href="https://techcrunch.com/2026/10/06/ai-computing-startup-lambda-to-raise-4b-ahead-of-planned-ipo/">Lambda pre-IPO raise</a></li><li><a href="https://www.theenergymag.com/news/2026-09-19/nscale-ipo-ai-contract-funding">Nscale S-1 contracts and commitments</a></li><li><a href="https://www.mostlymetrics.com/p/nscale-ipo-s1-breakdown">Nscale S-1 breakdown</a></li><li><a href="https://www.nscale.com/press-releases/pre-ipo-convertible-financing">Nscale pre-IPO converts</a></li><li><a href="https://ryanoconnellfinance.com/fed-rate-dashboard/">Fed funds and SOFR, Oct 15</a></li></ul>
<hr>
<p>Every chart above links to its interactive version on 06.ai, with the data behind it. The daily trackers are at <a href="https://06.ai/trackers">06.ai/trackers</a>.</p>
<p><em>Not investment advice. Estimates are labeled as estimates.</em></p>]]></content:encoded>
</item>
<item>
  <title>Who pays for the AI buildout?</title>
  <link>https://inthesingularity.ai/packs/hyperscaler-capex-funding</link>
  <guid isPermaLink="true">https://inthesingularity.ai/packs/hyperscaler-capex-funding</guid>
  <pubDate>Fri, 16 Oct 2026 12:01:00 GMT</pubDate>
  <description>Capex the four hyperscalers need through 2030, and how much of it operating cash flow can’t cover.</description>
  <category>AI infrastructure</category>
  <enclosure url="https://inthesingularity.ai/newsletter/hyperscaler-capex-funding/cover.png" length="47598" type="image/png" />
  <media:content url="https://inthesingularity.ai/newsletter/hyperscaler-capex-funding/cover.png" medium="image" type="image/png" />
  <content:encoded><![CDATA[<figure><a href="https://06.ai/packs/hyperscaler-capex-funding"><img src="https://06.ai/newsletter/hyperscaler-capex-funding/thumb.png" alt="Who pays for the AI buildout?" width="421" height="397"></a><figcaption><a href="https://06.ai/packs/hyperscaler-capex-funding">Interactive version and data on 06.ai →</a></figcaption></figure>
<p>On Street consensus numbers, Microsoft, Alphabet, Amazon and Meta need about <strong>$438B</strong> of debt or equity over 2026–30 — <strong>8%</strong> of capex. The rest comes out of operating cash flow.</p>
<ul><li><strong>$5.29T</strong> — Capex, 2026–30. All four · Street consensus</li><li><strong>$4.85T</strong> — Covered by operating cash flow. Capped at each year's capex</li><li><strong>$438B</strong> — Needs debt or equity. 8% of capex</li></ul>
<h2>Capex by year, and how it gets paid for</h2>
<figure><a href="https://06.ai/packs/hyperscaler-capex-funding#chart"><img src="https://06.ai/newsletter/hyperscaler-capex-funding/fig-01.png" alt="Capex by year, and how it gets paid for" width="694" height="462"></a><figcaption>Full bar = capex. Pink is what operating cash flow can’t cover. The black tick marks operating cash flow — above the bar is a cushion. Shown: Street consensus · All four.</figcaption></figure>
<h3>By company</h3>
<figure><a href="https://06.ai/packs/hyperscaler-capex-funding#chart"><img src="https://06.ai/newsletter/hyperscaler-capex-funding/fig-02.png" alt="Capex by year, and how it gets paid for — MSFT" width="694" height="493"></a><figcaption><strong>MSFT</strong>: Capex, 2026–30 $1.15T · Covered by operating cash flow $1.15T · Needs debt or equity $0B</figcaption></figure>
<figure><a href="https://06.ai/packs/hyperscaler-capex-funding#chart"><img src="https://06.ai/newsletter/hyperscaler-capex-funding/fig-03.png" alt="Capex by year, and how it gets paid for — GOOGL" width="694" height="493"></a><figcaption><strong>GOOGL</strong>: Capex, 2026–30 $1.53T · Covered by operating cash flow $1.36T · Needs debt or equity $165B</figcaption></figure>
<figure><a href="https://06.ai/packs/hyperscaler-capex-funding#chart"><img src="https://06.ai/newsletter/hyperscaler-capex-funding/fig-04.png" alt="Capex by year, and how it gets paid for — AMZN" width="694" height="493"></a><figcaption><strong>AMZN</strong>: Capex, 2026–30 $1.5T · Covered by operating cash flow $1.39T · Needs debt or equity $110B</figcaption></figure>
<figure><a href="https://06.ai/packs/hyperscaler-capex-funding#chart"><img src="https://06.ai/newsletter/hyperscaler-capex-funding/fig-05.png" alt="Capex by year, and how it gets paid for — META" width="694" height="493"></a><figcaption><strong>META</strong>: Capex, 2026–30 $1.11T · Covered by operating cash flow $945B · Needs debt or equity $163B</figcaption></figure>
<h2>Cumulative 2026–30 by company</h2>
<p>$B. Dividends and buybacks aren't deducted, so the true need is larger where payouts continue.</p>
<figure><a href="https://06.ai/packs/hyperscaler-capex-funding#breakdown"><img src="https://06.ai/newsletter/hyperscaler-capex-funding/fig-06.png" alt="Microsoft — Capex: $1.15T · Cash-flow funded: $1.15T · Debt / equity: $0B · Share: 0%; Alphabet — Capex: $1.53T · Cash-flow funded: $1.36T · Debt / equity: $165B · Share: 11% · Peak gap: 2027; Amazon — Capex: $1.5T · Cash-flow funded: $1.39T · Debt / equity: $110B · Share: 7% · Peak gap: 2026; Meta — Capex: $1.11T · Cash-flow funded: $945B · Debt / equity: $163B · Share: 15% · Peak gap: 2028; All four — Capex: $5.29T · Cash-flow funded: $4.85T · Debt / equity: $438B · Share: 8% · Peak gap: 2027" width="694" height="270"></a></figure>
<h2>The gap is real, but small against the balance sheets</h2>
<p>Microsoft funds everything from cash flow on every path. Alphabet ($165B), Meta ($163B) and Amazon ($110B) run gaps concentrated in 2027–28, and some of that is already pre-funded — Alphabet raised $84.75B of equity in June, and all three have issued bonds. Moving to management’s stated targets adds only about $36B: their own targets already roughly match Street numbers.</p>
<h2>Why there’s no prepay line</h2>
<p>These four are the funding source. They invest in the labs, and the labs commit that money back as cloud spend. None of them reports capex-linked customer prepayments. That changes if a cash-rich lab — after an IPO, say — starts paying hyperscalers up front; that would be worth watching.</p>
<h2>Where Oracle and SpaceX went</h2>
<p>Both build at hyperscaler scale but fund like something in between: capex far above cash flow, customer prepays or chip financing, concentrated lab customers, BBB-area credit. They have their own pack: <a href="https://06.ai/packs/leveraged-middle">The leveraged middle</a>.</p>
<h2>What this is anchored on</h2>
<p>Guidance, stated targets and analyst estimates as of October 16, 2026.</p>
<ul><li><strong>MSFT</strong> — CY26 capex incl. finance leases ~$175B (was ~$190B before lease-accounting change); FY27 to grow YoY; FY26 OCF $183B · Microsoft FY26 Q4 call, Jul 29 2026</li><li><strong>MSFT</strong> — $329B of data-center leases signed but not yet commenced (operating leases, outside capex) · Microsoft FY26 10-K</li><li><strong>GOOGL</strong> — 2026 capex $195–205B; 2027 to rise 'significantly'; Q2 FCF −$5.9B (first negative since IPO) · Alphabet Q2 2026 results, Jul 2026</li><li><strong>GOOGL</strong> — $84.75B equity raise (Jun 2026) incl. $40B ATM from Q3; LT debt $98.2B at Jun 30 · Alphabet release Jun 3 2026; Q2 balance sheet</li><li><strong>AMZN</strong> — 2026 cash capex ~$220B; goal to double power capacity by end-2027 vs 2025; TTM FCF −$7.6B · Amazon Q2 2026 call, Jul 30 2026</li><li><strong>META</strong> — 2026 capex $130–145B; 'at least $600B' US infra spend through 2028; ~14 GW compute target · Meta Q2 2026; Zuckerberg (Sep 2025); Citi</li><li><strong>ALL</strong> — 2027 Street capex: Citi GOOGL $308B / META $205B / AMZN $288B; Morgan Stanley META $225B→$250B (2028), AMZN $308B→$318B · Citi Jul 2026; Morgan Stanley Sep 2026</li></ul>
<h2>Method</h2>
<p><strong>Capex</strong> is calendar-year. 2026 is latest company guidance at midpoints (Microsoft on its post-lease-change basis). 2027–28 blend Citi, Morgan Stanley and LSEG consensus; 2029–30 assume growth flattening toward 2030, as consensus implies. “Stated targets” uses paths that hit what management has said: Meta’s $600B-through-2028 pledge and Amazon’s doubled power by end-2027. Microsoft and Alphabet have no quantified target beyond “grow”, so they match Street.</p>
<p><strong>Operating cash flow</strong> is our estimate, triangulated from FY26 actuals and published 2027 FCF estimates (S&amp;P, MarketScreener, BofA, TIKR); 2029–30 are extrapolated. Each year, cash flow covers capex up to its level; a surplus isn’t carried into the next year.</p>
<p><strong>Not in capex:</strong> operating leases signed but not commenced — $329B Microsoft, ~$279B Meta, $137B Amazon, $85B Alphabet — a second, off-balance-sheet funding channel.</p>
<hr>
<p>Every chart above links to its interactive version on 06.ai, with the data behind it. The daily trackers are at <a href="https://06.ai/trackers">06.ai/trackers</a>.</p>
<p><em>Not investment advice. Estimates are labeled as estimates.</em></p>]]></content:encoded>
</item>
</channel>
</rss>
