AI Infrastructure Frenzy Cooling? SemiAnalysis Sharply Cuts Five-Year Compute Forecasts, Nvidia Hit Hardest
- Key Takeaways: SemiAnalysis has sharply revised down its forecasts for AI accelerator capital expenditure and hardware demand, cutting its 2030 capex projection by $608 billion, as expectations for AI infrastructure investment growth are being recalibrated amid slowing revenue and tightening financing.
- Key Points:
- Global AI accelerator capex forecasts for 2028 through 2030 were cut by 28%, 41%, and 40% respectively, with the 2030 figure dropping from $1.531 trillion to $923 billion.
- The cumulative forecast for accelerator logic silicon wafer shipment area from 2026 to 2030 was revised down by 17%, with Nvidia accounting for 66% of the reduction.
- The 2028 global HBM shipment forecast was lowered from 9.0 EB to 6.4 EB, a 30% cut; the 2030 AI accelerator power demand forecast was reduced by 34% to 80.7 GW.
- Nvidia's shipment forecasts for 2027 through 2030 were cut by 22%, 39%, 56%, and 66% respectively, with the Feynman average selling price dropping from approximately $110,000 to about $78,000.
- Broadcom's ASIC business cumulative revenue was revised down by only 11%, primarily driven by price adjustments from lower HBM costs rather than shipment reductions.
- Frontier AI labs saw ARR growth slow in Q3, debt and equity investors showed reduced appetite for AI infrastructure investment, and the U.S. midterm elections led to the postponement of 5.8 GW of data centers.
Original author: Bu Shuqing
Original source: Wallstreetcn
The AI infrastructure investment boom is showing new signs of cooling.
In its latest model update released on November 13, semiconductor research firm SemiAnalysis significantly cut its forecasts for AI accelerator shipments, capital expenditure, HBM demand, and power demand over the coming years. The adjustments for 2028 to 2030 were particularly notable: the global AI accelerator capex forecast for 2028 was cut by 28%, while 2029 and 2030 were cut by 41% and 40%, respectively.

Under the latest model, global AI accelerator capex in 2030 will fall from the previously forecast $1.531 trillion to $923 billion, a reduction of $608 billion. Meanwhile, the cumulative forecast for accelerator logic silicon shipment area from 2026 to 2030 was cut by 17%, with Nvidia accounting for 66% of the reduction.

Nvidia took the biggest hit in this round of forecast revisions. SemiAnalysis not only sharply cut its shipment forecasts for Nvidia over the coming years, but also further lowered its average selling price assumptions for next-generation products, believing that as competition in the AI hardware market intensifies, Nvidia will need to accept lower margins to maintain competitiveness.
AI Capex Expectations Plunge, 2030 Forecast Cut by $608 Billion
The most direct signal from SemiAnalysis's adjustment is its reassessment of the scale of global AI accelerator capex.
According to its latest model, the global AI accelerator capex forecast for 2026 is essentially unchanged at around $392 billion. But starting in 2027, the previous high-growth expectations were significantly revised down.
The 2027 capex forecast was cut from $770 billion 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%; and 2030 from $1.531 trillion to $923 billion, down 40%.
This means the path of sustained rapid expansion in AI accelerator capex that SemiAnalysis previously expected is being markedly recalibrated. In particular, for 2029 and 2030, the originally projected annual investment scale of over $1 trillion has now been cut to the $800 billion to $900 billion range.
This does not mean AI infrastructure investment will stop growing. Under the new forecast, global AI accelerator capex will still grow from $392 billion in 2026 to $923 billion in 2030—just at a noticeably lower scale than previously expected.
Behind the capex downgrade are dual changes in AI industry revenue growth and the financing environment.
SemiAnalysis noted that the growth rate of new annual recurring revenue (ARR) at frontier AI labs slowed in the third quarter, prompting it to re-examine the scale of token revenue the entire ecosystem can generate. In order for AI business revenue to cover computing costs more sustainably, the firm correspondingly lowered its expectations for future compute supply expansion.
At the same time, debt and equity investors' willingness to invest in AI infrastructure capex is weakening, especially for projects lacking committed procurement support from major cloud service providers (hyperscalers). This will directly affect the development of data center land, power, and building shells, leading to reduced power supply available for deploying AI computing systems in the future.
The U.S. midterm election results have further intensified this pressure. SemiAnalysis estimates that of the U.S. data center supply originally scheduled to come online between the second half of 2027 and the first half of 2028, 5.8GW will be delayed, and regulatory factors will further compress data center capacity in 2028.
However, the firm specifically emphasized that data center construction delays caused by the midterm elections are only an additional factor in the forecast downgrade; slower AI revenue growth and tighter financing conditions are the more important drivers.
Cumulative Silicon Shipment Area Cut 17%, HBM Demand Also Sharply Revised
Beyond the capex forecast downgrade, SemiAnalysis's assessment of actual hardware demand has also become more cautious.
Under the latest model, the cumulative forecast for global AI accelerator logic silicon shipment area from 2026 to 2030 was cut by 17%. Among these, the forecasts for 2027, 2028, 2029, and 2030 were cut by 13%, 16%, 24%, and 15%, respectively.
Notably, Nvidia accounted for 66% of the silicon shipment area reduction, meaning this round of AI hardware demand expectation adjustments is not being borne evenly across suppliers—Nvidia is the primary affected party.
HBM demand forecasts were also notably revised down. The global HBM shipment forecast for 2028 was cut from 9.0EB to 6.4EB, down 30%; 2029 from 10.3EB to 7.9EB, down 23%; and 2030 from 10.9EB to 9.6EB, down 12%.

SemiAnalysis believes that the HBM supply-demand balance is improving, so it lowered future HBM price assumptions and passed the price changes through to the bill of materials (BOM) costs of various AI chips.
Previously, a considerable portion of the growth in AI infrastructure spending in 2027 came from price increases, especially memory prices. In the firm's view, the downward revision of price forecasts also means the supply chain is returning to a more normal state.
In addition, AI accelerator power demand forecasts were also sharply cut. Under the latest forecast, AI accelerator power demand in 2030 will fall from the previous 122.8GW to 80.7GW, down 34%; the 2029 forecast was cut from 108.7GW to 72.5GW, down 33%.
This change stems both from lower chip shipments and from adjustments to future per-chip power assumptions. SemiAnalysis noted that power continues to be a limiting factor, and the highest chip power levels in future accelerator product roadmaps may also become more moderate.
Nvidia Under the Most Pressure, Future Chip Shipment Forecasts Sharply Cut
Among major AI chip suppliers, Nvidia is the company most affected by this round of forecast adjustments.
On a compute revenue basis, SemiAnalysis cut Nvidia's shipment forecasts for 2027 through 2030 by 22%, 39%, 56%, and 66%, respectively; on a total power consumption basis including entire systems, the cuts were 14%, 35%, 50%, and 59%, respectively.
On pricing, SemiAnalysis not only considered the impact of lower HBM costs, but also believes Nvidia will need to reduce its markup on Feynman and subsequent products to cope with increasingly fierce competition.
Specifically, the average selling price forecast for Feynman was cut from about $110,000 to about $78,000, Feynman Ultra from about $175,000 to about $98,000, and Feynman Next from about $191,000 to about $118,000.
The dual downgrades to shipments and prices will put greater pressure on Nvidia's future revenue forecasts. This adjustment covers not only GPUs but also Groq LPUs and standalone CPU products.
SemiAnalysis explained that Nvidia was hit harder mainly because its chips have become less attractive relative to other platforms in the compute procurement mix of some large AI labs.
Take Anthropic as an example: the four major computing platforms it uses include TPU, Trainium, AMD GPUs, and Nvidia GPUs. SemiAnalysis believes that in terms of performance-per-cost, Nvidia GPUs are at the margin among these four platform types, and Anthropic prefers the other platforms.
On the OpenAI side, although GPUs still occupy an important position in its infrastructure, AMD GPUs have already become an attractive option. In addition, OpenAI will prioritize deploying in-house ASIC projects such as Jalapeno and the next-generation Serrano.
New cloud service providers (Neoclouds) constitute another risk for Nvidia. SemiAnalysis noted that Nvidia has the highest exposure to Neocloud customers, especially smaller long-tail customers. These companies currently face greater project execution and financing risks.
Notably, SemiAnalysis emphasized that this model adjustment does not incorporate the potential impact of Jensen Huang stepping down as Nvidia CEO; even if Jensen Huang continues as CEO, the firm would make the same forecast downgrades. However, the firm believes that leadership changes could further increase the downside risk of Nvidia losing market share.
Broadcom Relatively Resilient; AMD and MediaTek Face Varying Degrees of Adjustment
In sharp contrast to Nvidia, Broadcom showed greater resilience in this round of forecast adjustments.
SemiAnalysis cut its cumulative ASIC business revenue forecast for Broadcom from 2026 to 2030 by 11%, mainly due to average selling price adjustments from lower HBM costs rather than large-scale shipment cuts.
The firm believes that Broadcom's major ASIC projects—including Google TPU, OpenAI, and Anthropic-related projects—have advantages in performance-per-cost and energy efficiency, and Broadcom can also work with creditors to provide financing support for data center land, power, building shells, and system deployment.
AMD also showed some resilience but still faces notable downgrades. On a total power consumption basis, SemiAnalysis cut AMD's forecasts for 2028 through 2030 by 23%, 55%, and 57%, respectively; on a spending basis, the cuts were 29%, 35%, and 32%, respectively. The 2027 forecast was relatively stable, essentially flat on a power basis and down 11% on a spending basis.
AMD was less affected than Nvidia, mainly thanks to its positioning with large customers such as Meta, OpenAI, and Anthropic. SemiAnalysis believes that the warrant arrangements between Meta and OpenAI and AMD make purchasing AMD systems highly attractive; Anthropic also views AMD as the choice with the second-best performance-per-cost in inference computing after TPU.
In addition, AMD's recent product deployment pace has actually moved forward. SemiAnalysis moved the shipment timing of about 25,000 MI400 chips and about 30,000 Meta-customized MI450-MC chips from the fourth quarter of 2027 to the fourth quarter of 2026 to reflect stronger production ramp-up.
MediaTek, by contrast, faces more pronounced shipment adjustments. SemiAnalysis cut the cumulative Humufish lifecycle shipments for 2028 to 2029 by 470,000 units and cut the lifecycle shipments of the Triggerfish project for Anthropic by 42%.
Amazon Trainium was one of the few platforms whose near-term forecasts were not cut in this adjustment. SemiAnalysis did not reduce Trainium shipment forecasts for 2027 and 2028 because Amazon has a relatively ample pipeline of data center projects, Trainium has lower manufacturing costs, and Amazon is willing to support its key customers through strategic subsidies.
Overall, SemiAnalysis's adjustment does not mean AI compute demand will disappear entirely, but rather that growth expectations previously built on continuous financing, rapid data center expansion, and high-priced hardware are being recalibrated.
As AI lab revenue growth slows and capital market financing tightens, competition among chip suppliers will shift from pure performance contests to a comprehensive contest of performance-per-cost, energy efficiency, and financing capability.
Nvidia still holds a strong position in the AI chip market, but in SemiAnalysis's view, its future growth will face greater pressure than some ASIC suppliers.


