SemiAnalysis slashes AI spending forecasts: 2028 capex cut 28%, 2030 cut 40%
The research firm's largest revision in years cuts a combined $1.6 trillion from AI accelerator spending in 2027–2030 — and Nvidia takes the hardest hit.
📊 Scorecard: our Bearish call on $NVDA is scored from $156.14, the last close before it ran. Its first result comes after the next close.

The News
Research firm SemiAnalysis made its largest revision in years in a model update dated November 13, first widely reported on November 16–17. It cut its global AI accelerator capex forecast by 19% for 2027 (to $627 billion from $770 billion), 28% for 2028 (to $758 billion from $1.054 trillion), 41% for 2029 (to $796 billion from $1.352 trillion) and 40% for 2030 (to $923 billion from $1.531 trillion) — a combined $1.6 trillion lower across 2027–2030.
The 2028 HBM shipment forecast was cut 30%, cumulative 2026–2030 silicon wafer shipment area was revised down 17%, and 2030 AI accelerator power demand fell 34% (to 80.7 GW from 122.8 GW). Nvidia took the hardest hit, accounting for 66% of the silicon-area reduction, with its 2030 compute-revenue forecast cut 66% and its assumed price for the future Feynman Next chip cut 38%, from $191,000 to $118,000 (the base Feynman chip was cut to about $78,000 from about $110,000).
The firm attributed the revision to weaker data center funding, construction delays and a shift to price-performance competition. Spending still grows — 60% in 2027, 21% in 2028, 5% in 2029 and 16% in 2030 — but the growth curve flattens sharply. Counterpoint: UBS went the other way, raising its broader aggregate AI capex forecast — a wider measure than SemiAnalysis's accelerator-only figures — to $1.0 trillion for 2026, $1.4 trillion for 2027 and $1.5 trillion for 2028 in a November 2 note. Caveat: the underlying SemiAnalysis note is subscription-only and the firm has not publicly confirmed the figures — they come from summaries of it.
The AI boom is not over — but the forecasters just took $1.6 trillion off the top of it.
Why It Matters
Forecasts matter because the AI trade is built on them. Data center operators, chipmakers and memory suppliers all plan years ahead on the assumption that AI spending keeps compounding. When the most respected independent research firm in the sector makes its largest revision in years — cutting 2030 spending by 40% — every link in the chain has to re-plan, and the stocks re-price first.
The HBM cut is the memory angle, and it is the sharpest one. A 30% cut to 2028 high-bandwidth memory shipments hits SK hynix, Samsung and Micron directly, because HBM is the fastest-growing and most profitable part of their business. It also explains why Monday's memory selloff was worse than the broader chip rout: memory investors got two warnings at once, the spending cut and tighter U.S. controls on AI chip exports.
The Nvidia concentration is the other warning. One company accounts for two-thirds of the silicon-area reduction, and its 2030 compute-revenue forecast was cut by two-thirds. That is the risk when one supplier dominates a market: the upside is enormous, but so is the downside when the forecast turns. The Feynman price cut — $191,000 to $118,000 — says even the next generation of Nvidia chips will have to compete on price, not just performance.
Analyst's Take
Our call is Bearish. Capex forecasts are the demand model for the whole AI supply chain, so a $1.6 trillion cut reprices everything downstream at once. That is why the same forecast was cited as a driver of Monday's memory selloff, and why Nvidia took the hardest hit in the forecast itself, accounting for 66% of the silicon-area reduction.
The honest counterpoint is that spending still grows: a 60% jump in 2027 and 21% in 2028 is not a bust, and UBS raised its own forecasts. The debate is now about the slope of the curve, not its direction — and the slope just got much flatter. The 30% cut to 2028 HBM shipments is the number to watch.
For readers, the practical point is to treat the figures as reported summaries, not confirmed numbers: the SemiAnalysis note is subscription-only and the firm has not publicly confirmed them. Even so, several summaries of the same note agree on the figures, and the market has already voted with its feet. Watch Nvidia's guidance on Tuesday for the first official word on whether the slowdown is real. Reader poll: Bullish or bearish on AI spending now? Vote: Bullish, Bearish, Mixed.
Are you bullish or bearish on this story?
By The Numbers
FAQ
What is AI accelerator capex?
Capex is short for capital expenditure — money spent on long-lived assets. AI accelerator capex is what data center operators spend on the specialised chips, like Nvidia's GPUs, that train and run AI models.
What is HBM?
HBM stands for high-bandwidth memory, a fast type of memory stacked in layers that sits next to AI chips. A 30% cut to the 2028 HBM shipment forecast is a direct hit to memory makers like SK hynix, Samsung and Micron.
If the forecasts were cut, why is spending still growing?
The cuts lower the growth curve, they do not turn it negative. SemiAnalysis still expects AI accelerator spending to grow 60% in 2027 and 21% in 2028 — just from a lower starting point than it forecast before.
👁
Source: BigGo Finance