BlockBeats news, November 16 — AI research firm SemiAnalysis significantly lowered its forecasts for AI chip and server shipments and related capital expenditures after 2028 in its latest report, arguing that slowing ARR growth at frontier AI labs, tightening financing conditions, and delays in U.S. data center construction will jointly suppress computing power expansion over the coming years.
Under the firm's new model, global AI accelerator capital expenditure is projected at $758 billion in 2028, down 28% from its previous forecast; 2029 and 2030 are reduced to $796 billion and $923 billion respectively, representing cuts of 41% and 40% from prior estimates. The report attributes the downgrades mainly to falling chip prices, lower shipment volumes, and reduced assumptions for per-chip power consumption of future products.
Among major suppliers, Nvidia is seen as facing the biggest impact. SemiAnalysis expects Nvidia's shipment forecasts for 2027 through 2030, measured by compute revenue, to be cut by 22%, 39%, 56%, and 66%, respectively. The firm also lowered its average selling price expectations for Nvidia's future Feynman-series products, with Feynman dropping from about $110,000 to about $78,000 and Feynman Ultra from about $175,000 to about $98,000.
The report argues that pressure on Nvidia comes not only from slower overall demand but also from customers reprioritizing cost, energy efficiency, and financing capacity. Anthropic favors solutions such as TPU, Trainium, and AMD GPUs, while OpenAI is expected to prioritize deploying its in-house ASIC projects. At the same time, Nvidia has high exposure to neocloud customers, and such customers are facing greater financing and execution risks.
By contrast, Broadcom is seen as the least affected, with its ASIC revenue forecasts for 2026 through 2030 cut by only 11%, mainly due to price adjustments from lower HBM costs rather than significant shipment reductions. AMD's expectations were also lowered, but its Meta, OpenAI, and Anthropic customer base, along with related warrant arrangements, is believed to provide some support for demand. Amazon Trainium forecasts for 2027 and 2028 were not cut.
The report's overall judgment is that AI computing infrastructure buildout will not stop, but future growth will be more constrained by financing capacity, data center supply, chip prices, and customer progress on in-house chips; in this round of repricing, Nvidia may become the most visibly affected supplier.
