SemiAnalysis Significantly Lowers AI Computing Power Forecast: 2028 Capital Expenditure Projection Cut by Nearly 30%, Five-Year Wafer Shipment Area Reduced by 17%, Nvidia Hit the Hardest
SemiAnalysis has lowered its forecast for global AI accelerator capital expenditure in 2028 by 28% and cut the 2030 forecast by 40%. The cumulative shipment area of accelerator chips from 2026 to 2030 has been revised down by 17%, demand for HBM in 2028 has been slashed by 30%, and power consumption demand in 2030 has been revised down by 34%. Nvidia is the most impacted, accounting for 66% of the global chip shipment reduction, with its 2030 shipment forecast—based on revenue from computing—also reduced by as much as 66%.
The AI infrastructure investment boom is showing new signs of cooling down.
On November 13, semiconductor research firm SemiAnalysis released its latest model update, significantly lowering future projections for AI accelerator shipments, capital expenditures, HBM demand, and power demand over the next several years. Notably, the adjustments for 2028-2030 are particularly marked: the global AI accelerator capital expenditure forecast for 2028 has been revised down by 28%, while the forecasts for 2029 and 2030 are cut by 41% and 40%, respectively.

According to the latest model, global AI accelerator capital expenditure is projected to drop from the previous estimate of $1.531 trillion to $923 billion by 2030, a reduction of $608 billion. At the same time, the cumulative forecast for AI accelerator logic wafer shipment area from 2026 to 2030 has been revised down by 17%, with Nvidia accounting for 66% of the reduction.

In this round of forecast adjustments, Nvidia is the most affected. SemiAnalysis not only significantly reduced its shipment forecasts for Nvidia for the coming years, but also further lowered the average selling price assumptions for new generation products. As competition in the AI hardware market intensifies, SemiAnalysis believes Nvidia will have to accept lower profit margins to remain competitive.
AI Capital Expenditure Expectations Plummet, 2030 Forecast Cut by $608 Billion
The most direct signal in this adjustment by SemiAnalysis is the reevaluation of the scale of global AI accelerator capital expenditures.
Based on the latest model, the forecast for global AI accelerator capital expenditure in 2026 remains largely unchanged at around $392 billion. However, starting from 2027, the previously projected high-growth scenario is notably revised downward.
The 2027 capital expenditure forecast is cut from $770 billion to $627 billion, a 19% decrease; 2028 is reduced from $1.054 trillion to $758 billion, down 28%; 2029 is cut from $1.352 trillion to $796 billion, down 41%; and 2030 drops from $1.531 trillion to $923 billion, down 40%.
This means that SemiAnalysis’ previous expectation for sustained rapid expansion of AI accelerator capital expenditures is being clearly corrected. Especially for 2029 and 2030, the annual investment scale, which previously exceeded $1 trillion, is now lowered to the $800–900 billion range.
This does not mean AI infrastructure investment will stop growing. According to the new forecast, global AI accelerator capital expenditures will still rise from $392 billion in 2026 to $923 billion in 2030, just at a slower pace than previously expected.
The downward revisions in capex are driven by a dual shift in both AI industry revenue growth and the financing environment.
SemiAnalysis points out that the quarter-on-quarter growth rate of annual recurring revenue (ARR) from leading AI labs has slowed, prompting a reassessment of the total token revenue the ecosystem can generate. To ensure AI business revenue can more continuously cover computing costs, the firm has 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 large Hyperscaler commitments for procurement support. This will directly impact the development of data center land, power, and shell construction, leading to less power availability for deploying future AI computing systems.
The results of the US midterm elections further intensified this pressure. SemiAnalysis forecasts that of the US data center capacity planned for completion from the second half of 2027 to the first half of 2028, 5.8GW will be delayed, with regulatory factors further squeezing 2028 capacity.
However, the agency emphasizes that delays in data center construction due to the midterm elections are only an additional factor, with the slowdown in AI revenue growth and tightening financing environments being the more important drivers of downward revisions.
Cumulative Wafer Shipment Area Down by 17%, HBM Demand Heavily Revised
Beyond capex forecasts, SemiAnalysis is also adopting a more cautious view of actual hardware demand.
According to the updated model, the cumulative forecast for global AI accelerator logic wafer shipment area from 2026 to 2030 has been revised downward by 17%. The reductions for 2027, 2028, 2029, and 2030 are 13%, 16%, 24%, and 15%, respectively.
Notably, Nvidia accounts for 66% of the reduction in wafer shipment area, meaning that the adjustments to AI hardware demand forecasts are not evenly distributed among vendors, but primarily impact Nvidia.
HBM demand forecasts have also been clearly revised downward. The global HBM shipment forecast for 2028 falls from the previous 9.0EB to 6.4EB, a 30% reduction; 2029 drops from 10.3EB to 7.9EB, down 23%; and 2030 goes from 10.9EB to 9.6EB, a 12% reduction.

SemiAnalysis believes that the supply-demand balance for HBM is improving, therefore it has lowered future HBM price assumptions, passing these changes downstream to AI chip bill of materials (BOM) costs.
Previously, a considerable portion of the AI infrastructure spend growth for 2027 was driven by rising prices, particularly for memory. The downward price forecasts now suggest, according to SemiAnalysis, the supply chain is returning to a more normal state.
In addition, AI accelerator power demand has been drastically revised down. The latest forecast sees AI accelerator power consumption in 2030 dropping from 122.8GW to 80.7GW, down 34%; the 2029 projection decreases from 108.7GW to 72.5GW, down 33%.
This change is due to both lower chip shipment volumes and adjustments to the assumed power draw per chip in the future. SemiAnalysis points out that power remains a limiting factor, and maximum chip power on future accelerator roadmaps may moderate.
Nvidia Faces Greatest Pressure with Significant Downward Revision of Future Chip Shipments
Among the major AI chip vendors, Nvidia is the most significantly affected by these forecast updates.
By compute revenue, SemiAnalysis has revised down Nvidia’s shipment forecasts for 2027 to 2030 by 22%, 39%, 56%, and 66%; by total system power, the reductions are 14%, 35%, 50%, and 59%, respectively.
On pricing, SemiAnalysis incorporated the impact of declining HBM costs and expects Nvidia to reduce markups on Feynman and subsequent products to cope with intensifying competition.
Specifically, the average selling price estimate for Feynman was lowered from about $110,000 to $78,000, Feynman Ultra from about $175,000 to $98,000, and Feynman Next from about $191,000 to $118,000.
The dual downward adjustments to shipment volume and selling price will put greater pressure on Nvidia’s future revenue projections. These changes affect not only GPUs, but also Groq LPU and standalone CPU products.
SemiAnalysis explains that Nvidia is suffering more because its chips have become less attractive components of compute procurement portfolios at some large AI labs versus other platforms.
For example, at Anthropic, four computing platforms are used: TPU, Trainium, AMD GPU, and Nvidia GPU. According to SemiAnalysis, from a performance-to-cost perspective, Nvidia GPUs are at the margin among these four, and Anthropic prefers other platforms.
For OpenAI, although GPUs still play an important role in its infrastructure, AMD GPUs have become an increasingly attractive choice. OpenAI will also prioritize deploying self-developed ASICs such as Jalapeno and the next-gen Serrano.
New generation cloud providers (Neoclouds) are another risk for Nvidia. SemiAnalysis notes Nvidia is most exposed to Neocloud customers, especially smaller long-tail clients, who are currently facing greater project execution and financing risks.
It is noteworthy that SemiAnalysis emphasizes that this model adjustment does not factor in the potential impact of Jensen Huang’s resignation as Nvidia CEO. Even if he remains CEO, the firm would make the same downward projections. However, it believes leadership changes may further increase downside risk for Nvidia’s market share.
Broadcom Is Relatively Resilient; AMD and MediaTek Face Different Degrees of Adjustments
In stark contrast to Nvidia, Broadcom has demonstrated stronger resilience in this forecast adjustment.
SemiAnalysis revised down Broadcom’s cumulative ASIC business revenue forecast for 2026–2030 by 11%, mainly due to lower average selling prices resulting from reduced HBM costs, rather than large reductions in shipment volumes.
The firm argues that Broadcom’s key ASIC projects, including Google TPU, OpenAI, and Anthropic-related efforts, have performance-to-cost and power efficiency advantages. Broadcom also collaborates with creditors to provide financing support for data center land, power, shell, and system deployment.
AMD has also shown some resilience but still faces notable downward revisions. By total power, SemiAnalysis cut AMD’s forecasts for 2028, 2029, and 2030 by 23%, 55%, and 57%, respectively; by expenditure, the cuts are 29%, 35%, and 32%. The 2027 forecast remains relatively stable, with power flat and expenditure down 11%.
AMD is less impacted than Nvidia, mainly thanks to its deployments with major clients such as Meta, OpenAI, and Anthropic. SemiAnalysis notes that the warrant transaction arrangements between Meta/OpenAI and AMD make AMD systems very attractive; Anthropic regards AMD as second only to TPU in terms of performance-to-cost for inference computing.
Additionally, AMD's recent product deployment schedules have actually advanced. SemiAnalysis moved up the shipment of around 25,000 MI400 chips and about 30,000 Meta-customized MI450-MC chips from Q4 2027 to Q4 2026 to reflect stronger production ramp-up.
MediaTek faces more obvious shipment adjustments. SemiAnalysis cut the Humufish lifecycle shipments by 470,000 units for 2028–2029 and reduced Triggerfish project lifecycle shipments for Anthropic by 42%.
Amazon Trainium is one of the few platforms whose near-term forecasts were not revised down in this update. SemiAnalysis did not cut Trainium shipment forecasts for 2027 and 2028 because Amazon has a large reserve of data center projects. Trainium’s manufacturing costs are lower, and Amazon is willing to strategically subsidize key customers.
Overall, this SemiAnalysis revision does not mean AI compute demand will vanish, but rather that previous growth expectations based on continuous financing, rapid data center expansion, and high-price hardware are being recalibrated.
As AI lab revenue growth slows and capital market financing tightens, competition among chip suppliers will move beyond pure performance to a more comprehensive contest involving performance-to-cost, power efficiency, and financing capabilities.
Nvidia still retains a strong position in the AI chip market, but SemiAnalysis believes its future growth faces greater pressure than some ASIC vendors.
Disclaimer: The content of this article solely reflects the author's opinion and does not represent the platform in any capacity. This article is not intended to serve as a reference for making investment decisions.
You may also like
Wednesday Nvidia Earnings Focus: Without Jensen Huang, Who Else Can Tell a Good AI Story?
Nvidia’s Q3 report is expected to show revenue of $109 billion and net profit YoY growth of about 90%, but Jensen Huang’s sudden departure has shifted attention away from the financial results. HSBC downgraded its rating from “Buy” to “Hold” and slashed its price target to $170, while Morgan Stanley is concerned the company may postpone issuing guidance for next year. The market is now focused on the selection of a permanent CEO, whether the target of 70% revenue growth by FY2028 can be achieved, as well as risks related to Rubin mass production, AI client financing, and export controls. This earnings report will be a crucial test of whether Nvidia can break free from reliance on Jensen Huang.
U.S. Stock Market Preview | The three major stock index futures show mixed movements; Nvidia earnings, Anthropic IPO, and Federal Reserve minutes to be released this week; Intel sued by shareholders
Before the U.S. stock market opens on Monday, November 16th, the three major U.S. stock index futures showed mixed performance.
Gold steadies after hitting lowest since October 2025, hawkish Fed caps gains
XRP recovers 7%, ETF outflows hit $19.13 million as $1 level tested

