SemiAnalysis Slashes AI Capex Forecast by Nearly 30% for 2028, Nvidia Hit Hardest
TREE NEWS reports: In a significant update to its AI infrastructure model, research firm SemiAnalysis has sharply cut its forecasts for AI accelerator shipments, capital expenditures, HBM demand, and power needs through 2030. The revisions, published on November 13, signal a notable cooling in the AI investment boom. The most dramatic adjustments come for 2028–2030: global AI accelerator capex for 2028 is now projected at $758 billion, down 28% from the prior $1.054 trillion; 2029 is cut 41% to $796 billion; and 2030 is reduced 40% to $923 billion, a decrease of $608 billion from the previous $1.531 trillion estimate.
While capex is still expected to grow from $392 billion in 2026 to $923 billion in 2030, the pace is far slower than previously anticipated. The downward revision is driven by two main factors: slowing growth in annual recurring revenue (ARR) at leading AI labs, which prompted a reassessment of token revenue that can sustainably cover computing costs, and a tightening financing environment as debt and equity investors grow more cautious on AI infrastructure projects, especially those without commitments from major cloud service providers (hyperscalers). Additionally, the U.S. midterm election results are expected to delay 5.8 GW of data center capacity slated for 2027–2028, further compressing available power for AI compute.
Nvidia Bears the Brunt
Nvidia is the most affected company in this revision. SemiAnalysis cut its shipment forecasts for Nvidia’s compute revenue by 22%, 39%, 56%, and 66% for 2027 through 2030, respectively. On a total power basis, the cuts are 14%, 35%, 50%, and 59%. The firm also lowered average selling price (ASP) assumptions for Nvidia’s next-generation products, citing intensifying competition. For example, the Feynman GPU’s ASP is reduced from about $110,000 to $78,000, Feynman Ultra from $175,000 to $98,000, and Feynman Next from $191,000 to $118,000. Nvidia’s share of the reduction in logic wafer shipment area is a staggering 66%, highlighting its outsized exposure.
The downgrade reflects Nvidia’s diminishing attractiveness in the procurement mix of large AI labs. At Anthropic, for instance, Nvidia GPUs are seen as marginal on a performance-per-cost basis compared to TPUs, Trainium, and AMD GPUs. OpenAI is increasingly favoring AMD GPUs and its own custom ASICs (Jalapeno and Serrano). Moreover, Nvidia has the highest exposure to neocloud customers, particularly smaller long-tail players facing greater execution and financing risks.
Broadcom and AMD Show Resilience
Broadcom emerges as relatively resilient, with its ASIC revenue forecast cut by only 11% cumulatively from 2026 to 2030, mostly due to lower HBM costs passed through to pricing rather than volume reductions. Its key projects with Google TPU, OpenAI, and Anthropic are viewed favorably on performance-per-cost and energy efficiency, and Broadcom can also provide financing support for data center infrastructure.
AMD faces cuts but less severe than Nvidia. Its total power forecast is reduced by 23%, 55%, and 57% for 2028–2030, while spending cuts are 29%, 35%, and 32%. AMD benefits from strong positions with Meta, OpenAI, and Anthropic, and its MI400 and custom MI450-MC chips are now expected to ship earlier, in Q4 2026 instead of Q4 2027. MediaTek, however, sees significant shipment reductions for its Humufish and Triggerfish projects. Amazon’s Trainium is the only platform with no near-term cuts, supported by ample data center pipeline and strategic subsidies.
Market Implications
The revised forecasts suggest that the AI hardware boom is entering a more rational phase. Investors should consider the following:
- Nvidia’s growth premium may compress: With both shipment volumes and ASPs under pressure, Nvidia’s revenue growth could disappoint relative to lofty expectations. The stock may face multiple compression as the market recalibrates AI infrastructure spending.
- Diversify within AI semiconductors: Broadcom and AMD appear better positioned relative to Nvidia, given their ASIC traction and customer diversification. However, all face a slower capex environment.
- Watch financing conditions: The ability of AI infrastructure projects to secure debt and equity financing will be a key swing factor. Any further tightening could lead to additional downgrades.
- Power constraints are real: The 34% cut to 2030 AI accelerator power demand (from 122.8 GW to 80.7 GW) underscores that electricity availability is becoming a binding constraint, which could benefit power infrastructure providers but limit chip deployment.
- HBM and memory: Lower HBM demand forecasts (2028 cut by 30%) and price assumptions suggest memory makers may see less pricing power, impacting companies like SK Hynix and Micron.
Overall, the AI investment theme remains intact but is maturing. The easy money phase of hypergrowth may be over, and investors should focus on companies with strong performance-per-cost, energy efficiency, and financing capabilities.
Key Takeaways
- SemiAnalysis cut 2028–2030 AI capex forecasts by 28–41%, with 2030 now at $923 billion vs. $1.531 trillion previously.
- Nvidia is hit hardest: shipment forecasts cut up to 66% by 2030, and ASPs lowered significantly.
- Broadcom and AMD are relatively resilient, while MediaTek and smaller neoclouds face bigger risks.
- Slowing AI lab revenue growth and tighter financing are the main drivers, not a collapse in demand.
- Investors should recalibrate expectations for AI hardware stocks and watch financing and power constraints.




