W.091Altman told Samsung executives AI demand has been double-counted and that only OpenAI and Anthropic drive its growth, Financial News reports. The two labs’ disclosed and reported contracts come to about 39% of the $2.48T backlog at four clouds
Sam Altman told about 50 presidents of Samsung companies on October 31 that “there has been double-counting of demand” and that only OpenAI and Anthropic are driving major demand growth, Korea’s Financial News reported on November 11 from materials it obtained. Microsoft, Oracle, Google and Amazon last reported $2.48T of backlog between them. The contracts disclosed or reported between the two labs and those four add up to $967B, about 39%. Among neoclouds, Anthropic is up to 43% of Nscale’s contracted total.
What Altman is reported to have said
The remarks come from one newspaper’s account of a private meeting. Financial News says it obtained question-and-answer materials from a session on October 31 at Samsung’s training centre in Yongin, attended by about 50 presidents of Samsung affiliates. The quotes here are from the paper’s English edition, which is machine-translated; the report carries no comment from OpenAI or Samsung. Asked whether demand for AI, memory and data centers could saturate over five to ten years, Altman answered that there are short-term fluctuations. Financial News sets the remark in the debate over neoclouds, the companies that rent out GPUs: he said “there has been double-counting of demand” and that only OpenAI and Anthropic are driving major demand growth, so a supply shock and an unsettled period are possible. He added that demand will grow enormously in the long term.
He did not say what had been counted twice. The paper’s reporters read the remark as referring to infrastructure announcements that overlap from one tally to the next; that reading is theirs. Altman also said the heaviest user’s consumption rose from about 100,000 tokens a month at the start of 2020 to 100 billion a month early this year, a factor of 1,000,000 on those figures, and named the chip supply chain and data center infrastructure as the main constraints of the next few years. Asked whether heavy investment was squeezing profits, he said “That’s the point,” arguing that a company with cheaper, vertically integrated infrastructure can sell tokens at scale on low margins. The paper paraphrases him as saying OpenAI’s investments in efficiency, including chips it designs, are starting to pay off. A 24/7 Wall St. headline says he revealed a plan to stop buying Nvidia’s chips; the Financial News report does not say that.
OpenAI and Anthropic contracts come to about 39% of four clouds’ $2.48T backlog
Each bar is the backlog the company last reported, $ billions: Microsoft’s commercial remaining performance obligation and Google Cloud’s backlog at September 30, 2026, Oracle’s remaining performance obligations at August 31, and the AWS backlog Amazon gave on its third-quarter call. Pink is the value of OpenAI’s contracts with that company when they were announced; strong grey is the minimum Anthropic says it is committed to spend with it; faint grey is the backlog less those amounts. The label is pink plus strong grey as a share of the bar. Contract values are not the same as what remains in backlog today: some has been delivered, and Oracle has never confirmed the $300B.
Backlog and lab contracts by cloud
| Company | Measure | Backlog ($B) | OpenAI ($B) | Anthropic, at least ($B) | Two labs ÷ backlog |
|---|---|---|---|---|---|
| Microsoft | commercial RPO, Sep 30 | 677 | 250 | 47.2 | 44% |
| Oracle | total RPO, Aug 31 | 664 | 300 | – | 45% |
| Amazon | AWS backlog, Q3 call | 578 | 138 | 110.0 | 43% |
| Google Cloud backlog, Sep 30 | 560.8 | – | 121.3 | 22% | |
| Four clouds | 2479.8 | 688 | 278.5 | 39% |
How concentrated the backlog is
On disclosed contracts, the two labs account for roughly two-fifths of the big clouds’ backlog. Microsoft reported $677B of commercial remaining performance obligation at September 30, Oracle $664B at August 31, Google Cloud $560.8B and AWS $578B: $2.48T in all. OpenAI’s contracts with three of them were worth $688B when announced: $250B of Azure services, about $300B at Oracle as the Wall Street Journal reported it, and $38B and $100B at Amazon (W.061). Anthropic’s amended prospectus commits it to at least $47.2B at Microsoft, about $121.3B at Google and $110.0B at Amazon (W.084). Together that is $967B, 39.0% of the backlog. The Information estimated in May that the two labs were roughly half of about $2 trillion at the same four companies.
The share is an approximation, and the companies do not publish it. The sum sets contract values at signing and contractual minimums against backlog at different dates. Microsoft’s chief financial officer said about 45% of its $625B of commercial backlog at the end of 2025 came from OpenAI, about $281B; using that in place of the $250B contract gives 40.2%. Amazon’s chief executive described AWS’s backlog as “well diversified” on the third-quarter call. On these figures OpenAI and Anthropic contracts are 43% of it.
At the neoclouds and their landlords, single lab contracts are a large part of the total
Contracted totals as each company describes them, $ billions, with the part tied to an AI lab: pink where the lab is named, dashed where the company calls it only a frontier AI lab or AI laboratory, faint grey for the rest. CoreWeave’s OpenAI figure is the cap on contracts announced in 2025, not what remains in backlog; CoreWeave said it expanded its relationship with OpenAI in the third quarter without giving a value. Nscale’s total is active and contracted agreements in its IPO filing, of which about $2.6B was active. The three leases are single contracts shown whole.
Where the same demand appears twice
One contract shows up in more than one company’s disclosures. That is how contracts are reported, not an error, but it inflates any total that adds the lists together. Anthropic’s lease at TeraWulf’s Justified campus, $19B over 20 years, is contracted revenue for TeraWulf and a lease commitment for Anthropic, and is backstopped by a Google credit facility, according to 24/7 Wall St. Alphabet reports $48B of such backstops at September 30 and does not list them. Anthropic’s commitment of up to about $44.6B to Nscale is 43% of the $103.4B Nscale lists as active and contracted. CoreWeave is the largest tenant of Applied Digital, which reported more than $23B of contracted revenue in May; CoreWeave in turn reports $132.5B of backlog from its own customers, so the same capacity can sit behind both figures. Google Cloud’s backlog includes a $16.8B deal with Crux AI, a venture related to Alphabet.
The lab’s side of the ledger lists the same counterparties. Anthropic’s filing puts its committed spend at $900.9B. The seven largest named amounts, $646.1B together, are with Broadcom, Google, SpaceXAI, Amazon, Microsoft, Nscale and Lambda. Each is also backlog, contracted revenue or lease income on the other side. OpenAI has not published a comparable schedule; Bernstein counted its contracts at $614B of spending through 2030.
Anthropic’s $900.9B of commitments, by counterparty
The seven largest named amounts in Anthropic’s amended prospectus of November 10, $ billions, with the filing’s qualifier under each name: “at least” is a minimum, “up to” a ceiling. Each amount is the customer’s side of a contract that appears as backlog or contracted revenue at the supplier. The amounts are as set out in W.084.
What this changes on inthesingularity.ai
- The compute demand pack and W.076 use backlog as a measure of demand. Lead: a tracker row for each cloud’s backlog with the lab-attributed part beside it, updated each quarter, so that totals are not added across suppliers and customers.
- W.061 counted OpenAI’s contracts from the supplier announcements. Anthropic now publishes its side by counterparty. Lead: if OpenAI files a prospectus, set its schedule against the $688B the three clouds’ announcements imply.
Sources and method
| Item | Figure | Source |
|---|---|---|
| Altman | Remarks to about 50 presidents of Samsung affiliates at Samsung’s training centre in Yongin on Oct 31, from question-and-answer materials the paper obtained: demand has been double-counted; only OpenAI and Anthropic are driving major demand growth; a supply shock and an unsettled period are possible; demand will grow enormously in the long term; the heaviest user’s tokens rose from about 100,000 a month at the start of 2020 to 100 billion a month early this year; in-house chips; “That’s the point” | Financial News (English edition, machine-translated), Nov 11 |
| Altman | Secondary account; its headline says Altman revealed a plan to stop buying Nvidia’s chips, which the Financial News report does not say | 24/7 Wall St., Nov 12, 00:19 UTC |
| Microsoft | Commercial remaining performance obligation $677B at Sep 30, 2026, up 73% | Microsoft, first-quarter fiscal 2027 release, Oct 28 |
| Microsoft | Commercial RPO $625B at Dec 31, 2025, about 45% of it from OpenAI (Amy Hood, on the call) | Fierce Network, January 2026 |
| Oracle | Remaining performance obligations $664B at Aug 31, 2026 | Oracle, first-quarter fiscal 2027 results, Sep 10 |
| Google Cloud backlog $560.8B at Sep 30, 2026, from $513.9B at Jun 30; a $16.8B deal with Crux AI, a related-party venture, is in the backlog | Let’s Data Science, on Alphabet’s 10-Q and Bloomberg | |
| Backstops for partners’ data center lease and energy payments, notional maximum $48B | The Next Web, on the 10-Q | |
| Amazon | AWS backlog $578B; Andy Jassy: “growing triple digits year over year, and well diversified” | Amazon, third-quarter call remarks |
| OpenAI | Contracts: Microsoft $250B of incremental Azure services; Oracle about $300B, reported and not confirmed by Oracle; Amazon $38B and $100B; CoreWeave up to about $22.4B | W.061, from company releases and reports |
| Anthropic | Committed spend $900.9B: Broadcom equipment leases, about $161.2B; Google cloud, at least about $121.3B; SpaceXAI compute, up to $120.5B; Amazon cloud, at least $110.0B; Microsoft cloud, at least $47.2B; Nscale compute, up to about $44.6B; Lambda compute, about $41.3B | W.084, from the S-1/A of Nov 10 |
| Estimate | OpenAI and Anthropic were roughly half of about $2 trillion of combined backlog at Amazon, Microsoft, Google and Oracle; Anthropic committed about $200B to Google Cloud over five years | The Decoder, on The Information, May 6 |
| CoreWeave | Revenue backlog about $132.5B at Sep 30, 2026 | CoreWeave, third-quarter release, Nov 10 |
| Nscale | $103.4B of active and contracted agreements at Aug 31, 2026, about $2.6B of it active; Anthropic up to approximately $44.6B | Pulse 2.0, on Nscale’s IPO filing |
| Riot | About $28B over 15 years for 910 MW at Corsicana with an unnamed “leading Frontier AI Lab” | Riot Platforms, third-quarter release, Nov 4 |
| TeraWulf | $19B, 20-year lease of 401 MW at its Justified campus, for Anthropic with Google credit support | 24/7 Wall St., on TeraWulf’s third quarter |
| TeraWulf | 20-year lease with Anthropic, approximately $19 billion in revenue, 401 MW of critical IT load at the Justified site | The Block, Jul 6 |
| Cipher | Contracted revenue of about $9.1B at Barber Lake with an unnamed “leading artificial intelligence laboratory” | Finviz, on Cipher Digital’s third quarter |
| Applied Digital | More than $23B of total contracted revenue; CoreWeave its largest tenant | TIKR, May 2026 |
Altman’s remarks are as Financial News reported them in its machine-translated English edition; we have not seen the Korean original or the materials. Backlog: 677 + 664 + 578 + 560.8 = 2479.8. OpenAI: 250 + 300 + (38 + 100) = 688. Anthropic: 47.2 + 110.0 + 121.3 = 278.5. Share: (688 + 278.5) ÷ 2479.8 = 39.0%. With Microsoft’s own statement for the end of 2025, 625 × 45% = 281.25 in place of 250: 40.2%. The numerator uses contract values when announced and contractual minimums; the denominator is backlog at each company’s latest date, on each company’s own definition. Part of each contract has been delivered and has left backlog, parts may be cancellable, and contracts signed after the reporting date, including part of Anthropic’s Microsoft commitment, are not in the denominator, so the share is indicative. Oracle’s $300B is press reporting. Neoclouds: 22.4 ÷ 132.5 = 16.9%; 44.6 ÷ 103.4 = 43.1%. Anthropic: 161.2 + 121.3 + 120.5 + 110.0 + 47.2 + 44.6 + 41.3 = 646.1. Tokens: 100,000,000,000 ÷ 100,000 = 1,000,000.
Wire reports are written within hours of the news, on a six-hour schedule. Each lists its sources and shows the arithmetic behind its figures. Not investment advice.