NeuroAI NEUROAINEUROAI.SITE
China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories
ESC

Half the Cut Was Orders. Half Was the Price Tag.

SemiAnalysis's November 13 model update puts 2030 AI accelerator capex at 923 billion dollars. The previous version said 1.531 trillion.

2026-11-17 · 4788 words · NeuroAI
Half the Cut Was Orders. Half Was the Price Tag.

SemiAnalysis's November 13 model update puts 2030 AI accelerator capex at 923 billion dollars.

The previous version said 1.531 trillion.

A difference of 608 billion, same year, same table.

One five-year table

2026 holds at 392 billion, untouched. 2027 falls from 770 to 627 billion, down 19%. 2028 from 1.054 trillion to 758 billion, down 28%. 2029 from 1.352 trillion to 796 billion, down 41%. 2030 from 1.531 trillion to 923 billion, down 40%.

Everything cut sits after 2027.

The real revision lands in the back three years. The old version assumed infrastructure would rise another 45% in two years, pushing 2029 to 1.352 trillion and 2030 to 1.531 trillion; the new one presses those cells down 41% and 40%.

The further out, the harder the cut. 2029 is the blade.

Which 40% is being quoted

Headlines used 40%. The basis matters.

40% is single-year 2030 global AI accelerator capital spending, nominal dollars, old minus new, divided by old. 15,310 minus 9,230 is 6,080; divide by 15,310 and you get 39.7%.

It is neither a five-year cumulative nor an actual, but the gap between two forecasts. Nominal dollars ignore inflation, and no allowance is made for exchange rates or equipment deflation.

Sum 2026 through 2030 and the old version is 5.099 trillion, the new one 3.496 trillion, a cumulative 31.4% cut. Take 2027 through 2030: 4.707 trillion becomes 3.104 trillion, down 34.1%, a gap of 1.603 trillion. The 1.6 trillion in press coverage checks out.

The 8-point gap exists because 2026 did not move, diluting the decline. 40% is one year. 31% is five.

Silicon area fell only 17%

SemiAnalysis cut cumulative accelerator logic wafer area shipped from 2026 to 2030 by 17%, and 2029 alone by 24%. With die sizes constant, area proxies units: 17% less area means 17% less wafer. Nvidia accounts for 66% of the reduction.

Area beats unit count: the same wafer yields different counts depending on die size, so area combines how many and how big. Blind spots: advanced packaging, HBM stacking and yield. So 17% is a floor, not a ceiling.

Volume down 17%, money down 31%. The extra is not all price.

Divide the two

31.4 minus 17 is 14.4. But 14.4 is not a price effect; the two ratios have to be divided.

0.686 divided by 0.83 is 0.826. Capital spending per unit of logic wafer area fell 17.4%. Since capex equals area times unit price, area contributes -17% and price -17.4%, nearly half each.

It is not a clean decomposition: capex also covers HBM, packaging, power and networking. But the direction is clear: half the money went away because less was bought, half because what was bought got cheaper.

On the quote half, the report offers direct evidence. Nvidia's assumed Feynman price was cut from about $110,000 to $78,000, down 29.1%; Feynman Ultra from about $175,000 to $98,000, down 44.0%.

In the same update, 2028 global HBM shipments fall from 9.0 exabytes to 6.4, down 28.9%, and 2030 AI accelerator power demand is cut 34% to 80.7 gigawatts.

HBM and power cut in the same update suggests the whole chain is shrinking, not just chip prices.

Price gave way first.

Between two forecasts

SemiAnalysis gives three reasons. Annual recurring revenue growth at frontier AI labs slowed in the third quarter, so whether token revenue covers compute cost needs recalculating. Financing tightened: projects without a long-term cloud commitment are hard to place. US regulation and grid constraints are expected to delay 5.8 GW due online between the second half of 2027 and the first half of 2028.

Two demand reasons, one supply. Not one is "chips will not sell".

But capex is not demand. It is what the supply side plans to spend, so flat demand and lower prices still cut it. Two variables sit in one figure, which is why 40% is not enough.

On the first trading day after publication, the Philadelphia Semiconductor Index fell 5.28%, all 30 constituents down.

Lower prices are good for buyers: the same budget buys more compute and may stimulate inference. Whether the model captures that feedback, the report does not say.

Nvidia took the deepest cut

Nvidia's shipment forecasts for 2027 through 2030 are cut 22%, 39%, 56% and 66%.

Broadcom's ASIC revenue is cut only 11%, mostly through cheaper HBM rather than fewer units. AMD looks resilient. Amazon's Trainium near-term shipments are unchanged.

Why Nvidia: the largest share and the highest price assumptions. The old $110,000 for Feynman treated the most expensive part as the default; $78,000 concedes that alternatives caught up. Shipments down 66%, price down 29% — the revenue gap is not simple arithmetic.

Four companies, four outcomes, one table. The deepest cut lands on the widest moat.

A brake pedal in 2029

2028 capex rose 20.9%. 2029 is 796 billion, up only 5.0% on 758 billion; 2030 is 923 billion, up 16.0%.

A curve that has been cut flat almost flattens in 2029 and lifts in 2030. If revenue cannot cover cost, the 2030 acceleration needs another explanation — a new product cycle, perhaps. The report does not explain the 16%.

What to watch next time

The old version implies a 40.6% compound annual rate from 2026 to 2030; the new one, 23.9%. A 16.7-point drop over four years, not one year's number trimmed but the whole curve pressed flatter.

Compounding means 40.6% over four years reaches 1.531 trillion, and 23.9% reaches 923 billion, both from 392 billion.

23.9% is still growth: 923 billion in 2030 remains 2.35 times 392 billion in 2026.

So the question is not whether AI infrastructure stopped. It is whether it is worth the price.

For the next revision, watch silicon area. Area down again means demand. Area flat with money down means price.

*Disclaimer: This article is based on publicly available information and is for reference only. It does not constitute investment advice.*

Related coverage

Cite this article

NeuroAI (脑机网), "Half the Cut Was Orders. Half Was the Price Tag," 17 November 2026, https://neuroai.site/a/na-capital-capex-cut-1117

Machine-readable: BibTeX · RIS · JSON · Markdown

More in “Capital & Markets” → · Back to home · Markdown version

China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories