Just as the global AI computing power sector is undergoing a drastic correction, the semiconductor research organization SemiAnalysis has stepped in again.
In its latest model update, significant downgrades were made to the global AI accelerator capital expenditure, shipments, HBM demand, and power demand.
Among these, the adjustments for the years 2028 to 2030 are the most pronounced: the capital expenditure forecast for 2028 has been cut by 28%, and those for 2029 and 2030 have been reduced by 41% and 40%, respectively.
According to the new model, the global AI accelerator capital expenditure for 2030 will drop from the previously predicted $1.531 trillion to $923 billion, a reduction of $608 billion.
Meanwhile, the cumulative forecast for the shipment area of accelerator logic silicon chips from 2026 to 2030 has been downgraded by 17%, with NVIDIA accounting for 66% of the reduction.

The impact of this set of numbers is primarily reflected in the shattered narrative of the "unlimited expansion" of AI infrastructure previously held by the market.
Over the past two years, the capital market generally assumed that cloud vendors and AI labs would continue to invest heavily in computing power, thereby driving the revaluation of GPU, HBM, optical modules, data center electricity, and liquid cooling sectors.
However, SemiAnalysis's revisions indicate that the expectations for capital expenditure in 2026 remain relatively stable, with a significant turning point emerging only after 2027, particularly for the annual investment scale that was originally over $1 trillion for 2029 and 2030, now adjusted to a range of $800 billion to $900 billion.
Three main reasons lie behind the downgrade:
First, the growth of AI revenue is slowing. The annual recurring revenue growth rate for leading AI labs fell in the third quarter, prompting organizations to reassess whether Token revenue can continuously cover the high costs of computing power.
Second, the financing environment has tightened. Debt and equity investors are becoming more cautious about projects lacking long-term procurement commitments from cloud service providers, which will affect the development of data center land, electricity, and main constructions.
Third, U.S. regulations and grid constraints have intensified, with an estimated 5.8 GW of data centers expected to be delayed in coming online between the second half of 2027 and the first half of 2028.
The impact on the industrial chain is not uniformly distributed.
NVIDIA is hit the hardest; not only has its shipment forecast been significantly downgraded, but the average selling price assumption for the new generation of Feynman series products has also been significantly lowered.
For example, the average selling price of Feynman has been cut from about $110,000 to about $78,000, while Feynman Ultra has dropped from about $175,000 to about $98,000.
This indicates that competition in AI chips is shifting from "who has the capacity" to "who offers better cost performance," with alternative platforms like TPU, Trainium, and AMD GPU weakening NVIDIA's pricing power.
The simultaneous downgrade of HBM and power demand is also noteworthy.
The global HBM shipment forecast for 2028 has been reduced by 30%, and the AI accelerator power demand forecast for 2030 has been decreased by 34%.
This does not imply that the storage and power sectors have lost value; rather, it indicates that the industrial chain is transitioning from competing for capacity and power to allocating capacity based on returns.
HBM still holds a relatively high proportion in the AI accelerator bill of materials, and HBM4, advanced packaging, and customized storage continue to present structural opportunities; the bottlenecks on the power side will also not completely disappear, but the order rhythm will rely more on the specific project grid progress.
For the market, this round of downgrades feels more like an early calibration towards de-bubbling.
Previously, the market was inclined to interpret AI computing power investment as a one-sided upward curve, but in reality, it is influenced by revenue verification, financing constraints, power grid access, chip efficiency, and competitive dynamics.
Moving forward, key observations should focus on cloud vendors' capital expenditure guidance, visibility of accelerator orders, changes in HBM contract prices, data center grid access waiting times, and the share of inference load in total computing power.
If inference demand continues to grow, unit computing costs decline, and effective utilization improves, the industrial chain will still gain structural support; however, if long-term project returns are difficult to verify, capital expenditure forecasts may continue to be subject to revision.
Currently, the global AI computing power sector is undergoing a painful phase of valuation downgrades and bubble bursting.
Even so, it remains unclear when this phase will be complete; while valuations may appear to have declined significantly, the question lies in whether the predictive parameters within the valuation models have been adjusted downwards as well. If not, then the valuations may not seem undervalued.
The best approach is still to patiently await the market's natural clearing.
What can be expected is that market inertia often leads to excessive responses; although many negative factors have emerged and the AI computing power sector has significantly corrected, if the market over-corrects and truly pushes valuations down to extremely low levels, combined with the clearing of various leveraged funds and selling pressure while short positions remain at historically high levels, it may instead lead to a sharp rebound.
As for when this opportunity will arise, we can only leave that question to the market for an answer.
免责声明:本文章仅代表作者个人观点,不代表本平台的立场和观点。本文章仅供信息分享,不构成对任何人的任何投资建议。用户与作者之间的任何争议,与本平台无关。如网页中刊载的文章或图片涉及侵权,请提供相关的权利证明和身份证明发送邮件到[email protected],本平台相关工作人员将会进行核查。