inthesingularity.ai
Wire · Forecasts · SemiAnalysis, Nvidia

W.143SemiAnalysis cut its 2030 AI accelerator spending forecast 40%, to $923B, reports say

A SemiAnalysis model update dated November 13, relayed by several outlets on November 16, lowers forecast spending on AI accelerators to $627 billion in 2027, $758 billion in 2028, $796 billion in 2029 and $923 billion in 2030, from $770 billion, $1.054 trillion, $1.352 trillion and $1.531 trillion. The 2026 figure, about $392 billion, is essentially unchanged. The four-year cut is about $1.6 trillion, 34%. Nvidia takes the largest cuts: 66% of its 2030 shipments measured by compute revenue. The update is sold to clients, and SemiAnalysis has not published or publicly confirmed the figures. The new forecast for 2028 is below the old one for 2027.

Forecast AI accelerator spending in 2030, as reported
$923B
from $1.531T; 40% lower
Cut to the 2027–2030 total, our addition
$1.6T
$4.707T to $3.104T; 34% lower
Growth the new forecast implies for 2029
5%
our division; the previous forecast implied 28%
Cut to Nvidia’s 2030 shipments, by compute revenue
66%
as reported; 59% measured by system power

What was reported, and by whom

The numbers reach the public second-hand. Deep News, republished by Tiger Brokers, and BigGo Finance carried the fullest accounts on November 16, and their tables agree. OfficeChai reported that the source is an update to a SemiAnalysis model dated November 13 that is sold to institutional subscribers, that the figures come from summaries of it shared on X, that there is no public version, and that SemiAnalysis has not publicly confirmed them. We found no SemiAnalysis post carrying them. Everything below is the forecast as those outlets relayed it.

The series is spending on AI accelerators. The outlets call it global AI accelerator capital spending. 404K Research and OfficeChai describe it as unit shipments multiplied by average selling prices, which makes it what buyers pay chip vendors. It is not total data center or AI capital spending.

The forecast still rises every year. Spending goes from about $392 billion in 2026 to $923 billion in 2030, 2.4 times as much. The previous forecast reached $1.531 trillion, 3.9 times.

The new forecast for 2028, $758B, is below the previous one for 2027; the new 2030 is below the previous 2028

Forecast spending on AI accelerators, billions of dollars, by calendar year, as relayed by Deep News and BigGo Finance from a SemiAnalysis model update dated November 13, 2026. Grey: the previous forecast. Pink: the new forecast. The percentage under each year is our division of the two. The 2026 figure is described as essentially unchanged at about $392 billion. The series is spending on accelerators, not all capital spending.

0$400B$800B$1,200B$1,600B2026, previous forecast: $392 billion3922026, new forecast: $392 billion3922026unchanged2027, previous forecast: $770 billion7702027, new forecast: $627 billion, 18.6% lower6272027−19%2028, previous forecast: $1.054 trillion1,0542028, new forecast: $758 billion, 28.1% lower7582028−28%2029, previous forecast: $1.352 trillion1,3522029, new forecast: $796 billion, 41.1% lower7962029−41%2030, previous forecast: $1.531 trillion1,5312030, new forecast: $923 billion, 39.7% lower9232030−40%
04008001,2001,6002026, previous forecast: $392 billion2026, new forecast: $392 billion3922026unchanged2027, previous forecast: $770 billion2027, new forecast: $627 billion, 18.6% lower6272027−19%2028, previous forecast: $1.054 trillion2028, new forecast: $758 billion, 28.1% lower7582028−28%2029, previous forecast: $1.352 trillion2029, new forecast: $796 billion, 41.1% lower7962029−41%2030, previous forecast: $1.531 trillion2030, new forecast: $923 billion, 39.7% lower9232030−40%
The table
YearPrevious, $BNew, $BChange, $BChangeGrowth, previousGrowth, new
202639239200
2027770627−143−18.6%+96.4%+59.9%
20281,054758−296−28.1%+36.9%+20.9%
20291,352796−556−41.1%+28.3%+5.0%
20301,531923−608−39.7%+13.2%+16.0%
2027–20304,7073,104−1,603−34.1%

The shape of the cut

The cut grows with distance through 2029. It is 19% for 2027, 28% for 2028, 41% for 2029 and 40% for 2030. Across the four years the forecast falls from $4.707 trillion to $3.104 trillion, by our addition, a cut of $1.603 trillion or 34%. BigGo gives the same total as about $1.6 trillion.

Growth in 2027 stays high and 2029 nearly stalls. The new levels imply growth of 60% in 2027, 21% in 2028, 5% in 2029 and 16% in 2030. The previous levels implied 96%, 37%, 28% and 13%.

Volumes are cut less than spending, most clearly in 2030. Accelerator logic wafer area, a measure of volume, is cut 17% across 2026–2030, against a 31% cut to spending over the same years by our addition. By year the wafer cuts are 13%, 16%, 24% and 15%. In 2030 spending is cut 40% and wafer area 15%, and BigGo reports that the 2030 cut mostly reflects price assumptions. The reports say the update lowers assumed selling prices, including Nvidia’s for the Feynman generation: Feynman from about $110,000 to about $78,000, Feynman Ultra from about $175,000 to about $98,000, Feynman Next from about $191,000 to about $118,000. Those are cuts of 29%, 44%, 38%.

The previous forecast implied 28% growth in 2029; the new one implies 5%

Year-on-year growth in forecast spending on AI accelerators, percent. Each bar is our division of one year’s reported level by the year before. Grey: the previous forecast. Pink: the new forecast. Both start from the same 2026 figure of about $392 billion. BigGo Finance reports the new growth rates as 60%, 21%, 5% and 16%.

0%25%50%75%100%2027, previous forecast: $770B ÷ $392B = +96.4%96%2027, new forecast: $627B ÷ $392B = +59.9%60%20272028, previous forecast: $1.054T ÷ $770B = +36.9%37%2028, new forecast: $758B ÷ $627B = +20.9%21%20282029, previous forecast: $1.352T ÷ $1.054T = +28.3%28%2029, new forecast: $796B ÷ $758B = +5.0%5%20292030, previous forecast: $1.531T ÷ $1.352T = +13.2%13%2030, new forecast: $923B ÷ $796B = +16.0%16%2030
0%25%50%75%100%2027, previous forecast: $770B ÷ $392B = +96.4%96%2027, new forecast: $627B ÷ $392B = +59.9%60%20272028, previous forecast: $1.054T ÷ $770B = +36.9%37%2028, new forecast: $758B ÷ $627B = +20.9%21%20282029, previous forecast: $1.352T ÷ $1.054T = +28.3%28%2029, new forecast: $796B ÷ $758B = +5.0%5%20292030, previous forecast: $1.531T ÷ $1.352T = +13.2%13%2030, new forecast: $923B ÷ $796B = +16.0%16%2030

Who takes the cut

Nvidia’s cuts are the largest reported. Its shipments measured by compute revenue are lowered 22% for 2027, 39% for 2028, 56% for 2029 and 66% for 2030. Measured by total system power the cuts are 14%, 35%, 50% and 59%. Nvidia accounts for 66% of the reduction in wafer area. Deep News relays the reasoning: Nvidia has the most exposure to neocloud customers, who face financing and execution risk, and its chips sit in a marginal position on performance for the cost among the four platforms Anthropic uses.

The forecast leaves out Jensen Huang’s resignation. The reports say the model does not include any effect from it and that SemiAnalysis would have made the same cuts had he stayed. The reports add that a leadership change could raise the risk to Nvidia’s share. Nvidia announced the resignation on November 12, the day before the update is dated, and its 8-K says the board accepted it on November 11; How long can Nvidia run without Jensen Huang? covers the resignation.

AMD is cut less than Nvidia on spending, and less on power in 2027 and 2028 but about as much in 2029 and 2030. On a spending basis AMD is lowered 11%, 29%, 35% and 32% for 2027 to 2030. On a power basis it is roughly flat for 2027 and lowered 23%, 55% and 57% after. About 25,000 MI400 chips and about 30,000 of the MI450-MC, a version made for Meta, move up from the fourth quarter of 2027 to the fourth quarter of 2026.

Custom chips are cut mostly through price. Broadcom’s cumulative 2026–2030 ASIC revenue is lowered 11%, mainly because lower HBM costs reduce selling prices. MediaTek’s Humufish loses 470,000 units of lifetime shipments over 2028–2029, and its Triggerfish chip for Anthropic loses 42% of its lifetime shipments. Amazon’s Trainium shipments for 2027 and 2028 are not cut.

Nvidia’s 2030 shipments by compute revenue are cut 66%, against 40% for all accelerator spending

Size of the cut to each year’s forecast, percent; a longer bar is a deeper cut. Black: all accelerator spending, our division of the reported levels. Pink: Nvidia’s shipments on a compute-revenue basis, as reported. Grey: AMD on a spending basis, as reported. The three rows are measured on bases the reports name differently, so they are close to comparable but not identical.

0−20%−40%−60%−80%2027All accelerators, spending2027, all accelerator spending: $770B to $627B, 18.6% lower−19%Nvidia, shipments by compute revenue2027, Nvidia shipments on a compute-revenue basis: 22% lower, as reported−22%AMD, spending basis2027, AMD on a spending basis: 11% lower, as reported−11%2028All accelerators, spending2028, all accelerator spending: $1.054T to $758B, 28.1% lower−28%Nvidia, shipments by compute revenue2028, Nvidia shipments on a compute-revenue basis: 39% lower, as reported−39%AMD, spending basis2028, AMD on a spending basis: 29% lower, as reported−29%2029All accelerators, spending2029, all accelerator spending: $1.352T to $796B, 41.1% lower−41%Nvidia, shipments by compute revenue2029, Nvidia shipments on a compute-revenue basis: 56% lower, as reported−56%AMD, spending basis2029, AMD on a spending basis: 35% lower, as reported−35%2030All accelerators, spending2030, all accelerator spending: $1.531T to $923B, 39.7% lower−40%Nvidia, shipments by compute revenue2030, Nvidia shipments on a compute-revenue basis: 66% lower, as reported−66%AMD, spending basis2030, AMD on a spending basis: 32% lower, as reported−32%
0−20%−40%−60%−80%2027All accelerators2027, all accelerator spending: $770B to $627B, 18.6% lower−19%Nvidia2027, Nvidia shipments on a compute-revenue basis: 22% lower, as reported−22%AMD2027, AMD on a spending basis: 11% lower, as reported−11%2028All accelerators2028, all accelerator spending: $1.054T to $758B, 28.1% lower−28%Nvidia2028, Nvidia shipments on a compute-revenue basis: 39% lower, as reported−39%AMD2028, AMD on a spending basis: 29% lower, as reported−29%2029All accelerators2029, all accelerator spending: $1.352T to $796B, 41.1% lower−41%Nvidia2029, Nvidia shipments on a compute-revenue basis: 56% lower, as reported−56%AMD2029, AMD on a spending basis: 35% lower, as reported−35%2030All accelerators2030, all accelerator spending: $1.531T to $923B, 39.7% lower−40%Nvidia2030, Nvidia shipments on a compute-revenue basis: 66% lower, as reported−66%AMD2030, AMD on a spending basis: 32% lower, as reported−32%
Cuts by vendor, as reported
Series2027202820292030
All accelerator spending (our division)−18.6%−28.1%−41.1%−39.7%
Nvidia, compute-revenue basis−22%−39%−56%−66%
Nvidia, total-system-power basis−14%−35%−50%−59%
AMD, spending basis−11%−29%−35%−32%
AMD, total-power basisabout flat−23%−55%−57%
Accelerator logic wafer area−13%−16%−24%−15%

Memory and power

HBM shipments are cut most for 2028. The forecast for 2028 falls from 9.0 to 6.4 EB, for 2029 from 10.3 to 7.9 and for 2030 from 10.9 to 9.6. The reports print the unit as EB without spelling it out. They say assumed HBM prices were lowered as well.

Accelerator power demand is cut by a third. It falls from 108.7 GW to 72.5 GW for 2029 and from 122.8 GW to 80.7 GW for 2030, cuts of 33% and 34%. Deep News says part of the change is lower assumed power per chip. The reports do not say whether the series is power shipped in the year or installed in total.

5.8 GW of US data center supply is pushed back. The update estimates that this much capacity due between the second half of 2027 and the first half of 2028 will be delayed, and ties the delay to the US midterm elections and regulation. The reports describe it as a secondary cause.

The stated causes are revenue and financing. The reports say growth in frontier AI labs’ annualized recurring revenue slowed in the third quarter, and that debt and equity investors have become less willing to fund AI infrastructure, especially projects without a hyperscaler’s purchase commitment.

The HBM forecast for 2028 falls from 9.0 to 6.4 EB

Forecast HBM shipments by calendar year, in EB as the reports print it. Grey: the previous forecast. Pink: the new forecast. The percentage under each year is our division; the reports give the 2028 cut as 30%, which the rounded levels put at 28.9%.

0369122028, previous forecast: 9.0 EB9.02028, new forecast: 6.4 EB, 28.9% lower by our division6.42028−29%2029, previous forecast: 10.3 EB10.32029, new forecast: 7.9 EB, 23.3% lower by our division7.92029−23%2030, previous forecast: 10.9 EB10.92030, new forecast: 9.6 EB, 11.9% lower by our division9.62030−12%
0369122028, previous forecast: 9.0 EB9.02028, new forecast: 6.4 EB, 28.9% lower by our division6.42028−29%2029, previous forecast: 10.3 EB10.32029, new forecast: 7.9 EB, 23.3% lower by our division7.92029−23%2030, previous forecast: 10.9 EB10.92030, new forecast: 9.6 EB, 11.9% lower by our division9.62030−12%

Markets on November 16

Chip stocks fell, and most accounts gave other reasons. The Philadelphia Semiconductor index closed 5.28% lower at 9,610.65. Micron fell 7.62%, Sandisk 8.58%, AMD 5.58%, Nvidia 4.12% and Broadcom 3.92%. Invezz listed the forecast cut among several causes of the falls in Micron and Sandisk. TradingKey’s market wrap reported the cut as separate industry news without linking it to the fall. Seoul Economic Daily attributed the day to Jensen Huang’s departure from Nvidia. Motley Fool cited new US restrictions on AI chips, covered in W.141, the leadership transition and analysts’ price-target cuts. Neither mentioned SemiAnalysis.

Nvidia reports on November 17. In its last reported quarter, which ended July 26, its Data Center revenue was $89.0 billion, and it guided to $108.0 billion of total revenue for the quarter it reports next. Four quarters at $89.0 billion would be $356 billion a year, which gives a sense of scale for a forecast of $392 billion for all accelerator vendors in 2026. Nvidia’s fiscal quarters and segment do not match the forecast’s calendar years or scope.

What this changes on inthesingularity.ai

  • Will AI compute run short or long through 2030? Lead: a supply-side forecaster now puts 2030 accelerator power demand at 80.7 GW, down from 122.8 GW, and 2029 at 72.5 GW. If the pack’s supply path leans on the earlier SemiAnalysis numbers, it should be rerun on the new ones once they can be confirmed.
  • How much memory can each AI chip get? Lead: HBM for 2028 cut from 9.0 to 6.4 EB, with lower assumed HBM prices. The unit needs confirming before it is used.
  • How long can Nvidia run without Jensen Huang? Lead: a forecast that cuts Nvidia’s 2030 shipments by 66% on a revenue basis and says the cut does not depend on the leadership change.
  • Lead: the primary document. Every figure here is relayed. A public SemiAnalysis post, or its confirmation or correction of the relayed table, would settle the unit and series definitions.

Sources and method

ItemFigureSource
Spending forecastGlobal AI accelerator capital spending, previous to new: 2026 about $392 billion, essentially unchanged; 2027 $770B to $627B; 2028 $1.054T to $758B; 2029 $1.352T to $796B; 2030 $1.531T to $923B; 2030 lower by $608 billion; model update dated November 13Deep News via Tiger Brokers, Nov 16
Spending forecastThe same table; a cumulative cut of about $1.6 trillion over 2027–2030; revised growth of 60%, 21%, 5% and 16%; the update came 11 days after the previous version; the 2030 cut mostly reflects price assumptionsBigGo Finance, Nov 16
Spending forecast2028 lowered by 28%; accelerator logic wafer area lowered 17% over five yearsPANews, Nov 16
ProvenanceThe figures come from summaries, shared on X, of a model update sold to institutional subscribers, with no public version; SemiAnalysis has not publicly confirmed them; the series is unit shipments multiplied by selling prices and is not total capital spendingOfficeChai, Nov 16
ProvenanceThe series defined as shipments multiplied by average selling prices; wafer area defined as logic area excluding HBM and interposers404K Research, Nov 16
Wafer areaAccelerator logic wafer area: 2027 −13%, 2028 −16%, 2029 −24%, 2030 −15%; cumulative 2026–2030 −17%; Nvidia accounts for 66% of the reductionDeep News via Tiger Brokers
HBMHBM shipments, in EB: 2028 9.0 to 6.4 (−30%); 2029 10.3 to 7.9 (−23%); 2030 10.9 to 9.6 (−12%); HBM price assumptions also loweredDeep News via Tiger Brokers
PowerAI accelerator power demand: 2029 108.7 GW to 72.5 GW; 2030 122.8 GW to 80.7 GW; partly from lower power per chipDeep News via Tiger Brokers
NvidiaShipments on a compute-revenue basis: 2027 −22%, 2028 −39%, 2029 −56%, 2030 −66%; on a total-system-power basis: −14%, −35%, −50%, −59%; selling prices: Feynman about $110,000 to about $78,000; Feynman Ultra about $175,000 to about $98,000; Feynman Next about $191,000 to about $118,000Deep News via Tiger Brokers
AMDSpending basis: 2027 −11%, 2028 −29%, 2029 −35%, 2030 −32%; power basis: 2027 roughly flat, 2028 −23%, 2029 −55%, 2030 −57%; about 25,000 MI400 and 30,000 MI450-MC chips moved up from the fourth quarter of 2027 to the fourth quarter of 2026Deep News via Tiger Brokers
OthersBroadcom: cumulative 2026–2030 ASIC revenue −11%, mainly from selling prices; MediaTek: Humufish lifetime shipments cut by 470,000 units over 2028–2029, Triggerfish −42%; Amazon Trainium 2027 and 2028 shipments not cut; 5.8 GW of US data center supply due between the second half of 2027 and the first half of 2028 delayedDeep News via Tiger Brokers
ReasonsSlower growth in frontier labs’ annualized recurring revenue in the third quarter; tighter debt and equity financing; competition; lower HBM prices; the forecast excludes any effect of Jensen Huang’s resignation, though a leadership change could add to the risk to Nvidia’s shareDeep News via Tiger Brokers
ReasonsThree reasons listed: slower lab revenue growth, tighter financing and delays tied to the US midterm elections; 2030 power demand −34%Digital Today, Nov 17
MarketsMicron and Sandisk fell on November 16; the report cites the forecast cut among several causesInvezz via TradingView, Nov 16
MarketsPhiladelphia Semiconductor index -5.28%; Micron -7.62%; Nasdaq -2.49%; S&P 500 -1.92%; attributed to Huang’s departure from Nvidia; SemiAnalysis not mentionedSeoul Economic Daily, Nov 17
MarketsPhiladelphia Semiconductor index closed at 9,610.65, -5.28%; the forecast cut reported as separate industry newsTradingKey, Nov 16
MarketsSandisk -8.58%TradingKey, Nov 16
MarketsNvidia -4.12%; Broadcom -3.92%; AMD -5.58%; attributed to new US restrictions on AI chips, Nvidia’s leadership transition and analysts’ price-target cutsMotley Fool, Nov 16
NvidiaQuarter ended July 26, 2026: revenue $96.2 billion, Data Center $89.0 billion; outlook for the next quarter $108.0 billion, plus or minus 2%Nvidia 8-K exhibit, Aug 26
NvidiaThe board accepted Jensen Huang’s resignation on November 11, 2026, effective immediately; announced November 12Nvidia 8-K, Nov 12

What we read. The reports listed, which relay a SemiAnalysis client model update dated November 13, 2026. We did not see the update. We used a figure only where the accounts we read agree or where a single account gives it without contradiction, and we name the outlet in the table. Arithmetic. Changes: 627 ÷ 770 − 1 = −18.6%; 758 ÷ 1,054 − 1 = −28.1%; 796 ÷ 1,352 − 1 = −41.1%; 923 ÷ 1,531 − 1 = −39.7%. The reports give −19%, −28%, −41% and −40%. Totals for 2027–2030: 770 + 1,054 + 1,352 + 1,531 = 4,707; 627 + 758 + 796 + 923 = 3,104; 4,707 − 3,104 = 1,603; 1,603 ÷ 4,707 = 34.1%. With 2026: 392 + 4,707 = 5,099; 392 + 3,104 = 3,496; 1 − 3,496 ÷ 5,099 = 31.4%. Growth, previous: 770 ÷ 392 = +96.4%; 1,054 ÷ 770 = +36.9%; 1,352 ÷ 1,054 = +28.3%; 1,531 ÷ 1,352 = +13.2%. Growth, new: 627 ÷ 392 = +59.9%; 758 ÷ 627 = +20.9%; 796 ÷ 758 = +5.0%; 923 ÷ 796 = +16.0%. 2030 over 2026: 923 ÷ 392 = 2.35; 1,531 ÷ 392 = 3.91. HBM: 6.4 ÷ 9.0 − 1 = −28.9%; 7.9 ÷ 10.3 − 1 = −23.3%; 9.6 ÷ 10.9 − 1 = −11.9%. Power: 72.5 ÷ 108.7 − 1 = −33.3%; 80.7 ÷ 122.8 − 1 = −34.3%. Selling prices: 78 ÷ 110 − 1 = −29.1%; 98 ÷ 175 − 1 = −44.0%; 118 ÷ 191 − 1 = −38.2%. Nvidia: 89.0 × 4 = 356. Conflicts between sources. Headlines say the 2028 forecast fell nearly 30%; the bodies and the levels give 28%. The reported 30% cut to 2028 HBM is 28.9% on the rounded levels. Digital Today lists three reasons for the cut and Deep News lists more. Not established. Whether EB means exabytes or exabits; whether power demand is annual or cumulative. The previous version is said to be 11 days older, which would be about November 2; no outlet gives its date. Not used. The names of the analysts and of the model, which one outlet gives; a claim in one account that wafer area still grows 80% in 2028; closing figures for TSMC’s US shares, which we could not confirm; and comparisons with hyperscalers’ capital spending guidance, which covers a different scope.

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.