SemiAnalysis Cuts AI Compute Forecasts: 2028 Capex Projection Down Nearly 30%, Five-Year Wafer Shipments Down 17%, Nvidia Hit Hardest

Deep News
昨天

New signs of cooling are emerging in the AI infrastructure investment boom.

Semiconductor research firm SemiAnalysis, in its latest model update released on November 13, significantly lowered its forecasts for AI accelerator shipments, capital expenditure, HBM demand, and power demand over the coming years. The adjustments for 2028 through 2030 were particularly pronounced: the 2028 global AI accelerator capex forecast was cut by 28%, while 2029 and 2030 were reduced by 41% and 40%, respectively.

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

Among the revisions, NVIDIA took the biggest hit. 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, arguing 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 capital expenditure.

According to its latest model, the global AI accelerator capex forecast for 2026 is essentially unchanged, remaining at around $392 billion. But starting in 2027, the previous high-growth expectations were notably revised downward.

The 2027 capex forecast was cut from $770 billion to $627 billion, a 19% reduction; 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 high-speed expansion in AI accelerator capex that SemiAnalysis had previously expected is being clearly revised. In particular, for 2029 and 2030, annual investment scale that was originally projected to exceed $1 trillion has now been lowered 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, but the scale of growth is significantly lower than previously expected.

Behind the capex downgrade are dual changes in AI industry revenue growth and the financing environment.

SemiAnalysis noted that the pace of new annual recurring revenue (ARR) growth at frontier AI labs slowed in the third quarter, prompting it to re-examine the scale of token revenue the entire ecosystem can generate. To allow 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, and that slowing AI revenue growth and a tightening financing environment are the more important drivers.

Silicon Wafer Shipment Area Cumulatively Cut 17%, HBM Demand Also Sharply Revised

Beyond the capex forecast downgrade, SemiAnalysis's judgment on actual hardware demand has also become more cautious.

Under the latest model, the cumulative forecast for global AI accelerator logic silicon wafer shipment area from 2026 to 2030 was lowered by 17%. Among that, the forecasts for 2027, 2028, 2029, and 2030 were cut by 13%, 16%, 24%, and 15%, respectively.

It is worth noting that NVIDIA accounts for 66% of the reduction in silicon wafer shipment area, meaning this round of AI hardware demand expectation adjustments is not pressuring all suppliers equally, and NVIDIA is the primary affected party.

HBM demand forecasts also saw significant downgrades. The 2028 global HBM shipment forecast 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 the HBM supply-demand balance is improving, and therefore 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, this price forecast downgrade also means the supply chain is returning to a more normal state.

In addition, AI accelerator power demand was also sharply revised down. Under the latest forecast, AI accelerator power demand in 2030 will fall from the previously projected 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 shipment volumes and from adjustments to future per-chip power consumption assumptions. SemiAnalysis noted that power remains a persistent constraint, and the highest chip power consumption in future accelerator product roadmaps may also moderate.

NVIDIA Under the Most Pressure, Future Chip Shipment Forecasts Sharply Cut

Among major AI chip suppliers, NVIDIA became the company most affected by this round of forecast adjustments.

Using a compute revenue basis, SemiAnalysis cut NVIDIA's shipment forecasts for 2027 through 2030 by 22%, 39%, 56%, and 66%, respectively; measured on a total power consumption basis including the entire system, the reductions were 14%, 35%, 50%, and 59%.

On pricing, SemiAnalysis not only considered the impact of lower HBM costs but also argued that 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.

With both shipment volumes and selling prices lowered, there will be even greater pressure on NVIDIA's future revenue forecasts. This adjustment involves not only GPUs but also Groq LPU 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 GPU, and NVIDIA GPU. SemiAnalysis believes that in terms of performance-to-cost ratio, NVIDIA GPUs are in a marginal position among these four platforms, and Anthropic prefers other platforms.

On OpenAI's side, although GPUs still occupy a relatively important position in its infrastructure, AMD GPUs have become an attractive option. In addition, OpenAI will prioritize deploying self-developed ASIC projects such as Jalapeno and the next-generation Serrano.

New cloud service providers (Neoclouds) constitute another risk for NVIDIA. SemiAnalysis pointed out that NVIDIA has the highest exposure to Neocloud customers, especially smaller long-tail customers. Such companies currently face greater project execution and financing risks.

It is worth noting that SemiAnalysis emphasized that this model adjustment did not incorporate any potential impact from Jensen Huang resigning as NVIDIA CEO, and that even if Jensen Huang continues as CEO, the firm would still make the same forecast downgrades. However, the firm believes leadership changes could further increase downside risk to NVIDIA's market share loss.

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 reductions.

The firm believes that Broadcom's major ASIC projects, including Google TPU, OpenAI, and Anthropic-related projects, have advantages in performance-to-cost ratio and energy efficiency, and that 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 clear downgrades. SemiAnalysis, on a total power consumption basis, cut AMD's forecasts for 2028 through 2030 by 23%, 55%, and 57%, respectively; on an expenditure basis, the reductions were 29%, 35%, and 32%. The 2027 forecast was relatively stable, essentially flat on a power basis and down 11% on an expenditure basis.

AMD was hit less than NVIDIA, mainly thanks to its positioning with major 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 option with the second-best performance-to-cost ratio in inference computing after TPU.

In addition, AMD's recent product deployment pace has actually accelerated. 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 Humufish lifecycle shipments cumulatively by 470,000 units for 2028 to 2029 and lowered the lifecycle shipments for the Triggerfish project for Anthropic by 42%.

Amazon Trainium was one of the few platforms whose near-term forecasts were not downgraded in this adjustment. SemiAnalysis did not cut Trainium's shipment forecasts for 2027 and 2028 because Amazon has a relatively ample pipeline of data center projects, Trainium's manufacturing cost is lower, 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 labs' revenue growth slows and capital market financing tightens, competition among chip suppliers will shift further from a pure performance contest to a comprehensive contest of performance-to-cost ratio, 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.

免责声明:投资有风险,本文并非投资建议,以上内容不应被视为任何金融产品的购买或出售要约、建议或邀请,作者或其他用户的任何相关讨论、评论或帖子也不应被视为此类内容。本文仅供一般参考,不考虑您的个人投资目标、财务状况或需求。TTM对信息的准确性和完整性不承担任何责任或保证,投资者应自行研究并在投资前寻求专业建议。

热议股票

  1. 1
     
     
     
     
  2. 2
     
     
     
     
  3. 3
     
     
     
     
  4. 4
     
     
     
     
  5. 5
     
     
     
     
  6. 6
     
     
     
     
  7. 7
     
     
     
     
  8. 8
     
     
     
     
  9. 9
     
     
     
     
  10. 10