SemiAnalysis Cuts 2028 AI Chip Spending Forecast by 28%, Nvidia Bears Most of the Reduction
nashnova research
SemiAnalysis cut its 2028–2030 AI accelerator spending forecast by 28%–41%, removing roughly $1.6 trillion over four years, with Nvidia absorbing most of the reduction through price cuts and share loss — a signal that AI hardware's steep growth curve is flattening against reality.
How much was cut — and where did the money go?
2027 spending is now forecast at $627 billion, 19% below the prior model; 2029 at $796 billion, down 41% — the steepest single-year cut.
Growth hasn't vanished — 2027 still rises 60% from 2026's $392 billion — but it decelerates sharply after that, with 2029 growing just 5%.
This means → the once-steep curve has been "flattened": spending still climbs, but far more slowly, losing roughly $1.6 trillion versus the old model over four years.
Actual shipment volumes fell less than spending — why?
Total silicon area — the physical volume of chips shipped — drops only 17% over 2026–2030, far less than the spending cut.
In 2030, spending falls 40%, but silicon area drops just 15% and HBM (high-bandwidth memory, the ultra-fast memory paired with AI chips) bits drop only 12%.
In plain terms = the spending plunge is mostly a price story, not a volume story. For players like TSMC that bill by silicon area, the real hit is much smaller than the headline numbers suggest.
Why the downgrade — which reason matters most?
First: revenue slowdown. Frontier AI labs saw net new annualized recurring revenue decelerate in Q3 2026, prompting SemiAnalysis to reset compute-revenue expectations across the ecosystem and trim supply to keep revenue above cost. The firm still says it is more bullish on lab revenue than Street consensus.
Second: tighter financing. Debt and equity investors have grown more cautious on AI infrastructure capex; projects without hyperscaler offtake agreements face particularly difficult funding, directly shrinking the power envelope available for chip deployment.
Third: U.S. midterm elections. The election outcome pushed 5.8 GW of planned U.S. data-center supply into later years — but SemiAnalysis labels this a secondary factor, less decisive than the first two.
This means → the core logic is "AI hasn't earned enough to justify this much spending" — demand hasn't disappeared, but reality has caught up with expectations.
How do the price cuts work — and why does Nvidia get hit hardest?
Layer one: industry-wide HBM price cuts. SK Hynix HBM4 12-layer pricing for 2027 is now assumed at $2.71 per Gb, 12% below the prior model; 2028 at $2.52, down 18%. Micron's cuts are similar, mostly 15%–30%. Samsung is an exception — slightly higher in 2027, then 10%–20% lower from 2028.
Layer two: Nvidia-specific margin compression. For Feynman and later products, SemiAnalysis cut Nvidia's markup over HBM cost — reasoning that intensifying competition forces Nvidia to accept lower margins to stay competitive.
Feynman-series products see 29%–44% price cuts; Broadcom's same-generation TPU drops just 5%–6%, AMD 10%–15%. Nvidia's cut is five to seven times Broadcom's and roughly three times AMD's.
In plain terms = everyone is cutting prices, but Nvidia is taking the deepest hit because it has the most markup to give — and the most competitive pressure to give it.
What does Nvidia's revenue trajectory look like?
Under SemiAnalysis's model, Nvidia's data-center compute revenue is 22%, 39%, 56%, and 66% below the old forecast for 2027–2030 respectively.
In calendar-year terms: roughly $570 billion in 2027, $514 billion in 2028, $439 billion in 2029, $389 billion in 2030 — three consecutive years of decline starting in 2028.
Wall Street points the other way: Morgan Stanley's latest forecast still puts Nvidia's FY2029 revenue (roughly calendar 2028) at $866.7 billion, up 39.1% — about 1.7× SemiAnalysis's figure for the same period.
This reflects a fundamental divergence between SemiAnalysis and sell-side consensus on Nvidia's outlook — not a matter of degree, but of direction.
Why is Nvidia's share under pressure — and what other risks loom?
Three share-loss drivers: ① GPUs rank lower in the compute mix at frontier labs like Anthropic; ② Nvidia has the highest exposure to emerging cloud providers, whose financing and execution risks have risen; ③ the RASA Act — which would bring remote GPU access under export controls — is gaining legislative traction.
Power demand is also revised down: 2030 incremental power demand for accelerator chips falls from 122.8 GW to 80.7 GW, a 34% cut, with 60% of the reduction driven by lower per-chip power assumptions.
SemiAnalysis explicitly notes this model does not factor in Jensen Huang's departure as CEO — the downgrade would have been the same even if he stayed, but the leadership transition adds additional downside risk to the share-loss trajectory.
市场有风险,内容仅供研究参考,不构成投资建议。
