HomeArticle

Expectations for AI chips have been sharply lowered.

半导体行业观察2026-11-17 09:41
A clear shift is taking place in the growth expectations for AI infrastructure investment.

There is a noticeable shift underway in the growth expectations for AI infrastructure investment.

On November 13, semiconductor research firm SemiAnalysis released its latest model adjustment, sharply downgrading forecasts for global AI accelerator capital expenditure, chip shipments, HBM demand and power consumption in the coming years. Among them, the 2028 global AI accelerator capital expenditure expectation was cut from the previous $1.054 trillion to $758 billion, a 28% reduction; the 2030 figure was reduced from $1.531 trillion to $923 billion, a roughly 40% drop.

In this round of adjustments, NVIDIA has become the company most affected. SemiAnalysis not only significantly lowered NVIDIA's GPU shipment expectations for the next few years, but also cut the price of its next-generation Feynman series products. In comparison, Broadcom is less affected, AMD also shows certain resilience, and Amazon's Trainium short-term shipment expectations have not even changed.

This may mean that the AI computing power market is entering a new stage: demand is still growing, but the previous market's optimistic expectations for the scale of computing power investment, chip prices and customer procurement capabilities are beginning to face reassessment.

Global AI computing power investment expectations see an adjustment of nearly one trillion US dollars

Let's first look at the overall market.

According to the latest forecast from SemiAnalysis, global AI accelerator capital expenditure in 2027 will drop from the previously estimated $770 billion to $627 billion, a 19% reduction; the 2028 figure will fall from $1.054 trillion to $758 billion, a 28% cut; the 2029 number will drop from $1.352 trillion to $796 billion, a 41% reduction; the 2030 figure will be reduced from $1.531 trillion to $923 billion, a roughly 40% drop.

According to this forecast, from 2027 to 2030, the cumulative expected global AI accelerator capital expenditure will be about $1.6 trillion less than previously estimated. The market previously expected that the annual AI accelerator capital expenditure would exceed $1 trillion in 2029, but now this timeline has been significantly delayed.

BigGo Finance pointed out in related reports that this adjustment does not mean that AI computing power demand has disappeared, but that the previously expected rapid expansion is cooling down. According to the adjusted model, global AI accelerator capital expenditure will still grow from $392 billion in 2026 to $923 billion in 2030, but the growth slope has slowed significantly. There may even be only about 5% year-on-year growth in 2029.

SemiAnalysis gives three main reasons.

First, the revenue growth of leading AI companies shows signs of slowing down. Especially for some cutting-edge AI labs, the growth rate of new annualized recurring revenue (ARR) in the third quarter fell short of previous expectations. This led SemiAnalysis to reassess the revenue that AI model commercialization can bring, and what scale of computing power investment these revenues can actually support.

Second, the financing environment is tightening. Previously, a large number of AI data center projects were advanced through debt financing and equity financing, but now investors are more cautious about projects that lack procurement commitments from large cloud service providers. Land, power, main data center buildings and subsequent equipment procurement may all be affected.

In addition, regulatory factors related to the US midterm elections may also cause some data center projects to be delayed. SemiAnalysis expects that about 5.8GW of data center capacity originally planned to be put into operation from the second half of 2027 to the first half of 2028 may be postponed. However, the agency emphasized that slowing revenue growth and tightening financing environment are the main reasons for this adjustment.

This adjustment has also been transmitted to the semiconductor industry chain.

SemiAnalysis lowered its forecast for the cumulative shipment area of global AI accelerator logic wafers from 2026 to 2030 by 17%. Among them, the single-year reduction in 2029 reached 24%.

In terms of HBM, the 2028 global shipment forecast was reduced from 9.0EB to 6.4EB, a 30% cut; the 2029 figure was lowered from 10.3EB to 7.9EB, a 23% reduction; the 2030 number was reduced from 10.9EB to 9.6EB, a 12% drop.

At the same time, the 2030 global AI accelerator power consumption demand forecast was reduced from 122.8GW to 80.7GW, a 34% cut.

There is a notable detail here: the reduction in capital expenditure is not exactly equal to the reduction in actual chip demand.

Taking 2030 as an example, the AI accelerator capital expenditure forecast is reduced by about 40%, but the logic wafer shipment area is only reduced by 15%, and the HBM bit demand is reduced by 12%. This shows that in addition to the reduction in shipments, a considerable part of this adjustment comes from the expected decline in chip selling prices and memory costs.

SemiAnalysis believes that as the supply-demand relationship of HBM gradually improves, memory prices will normalize in the future. A considerable part of the previous growth in AI infrastructure investment came from rising memory prices, and now this part of the price premium is being readjusted.

Therefore, for wafer fabrication, semiconductor equipment, and packaging and testing enterprises, the 40% reduction in AI accelerator capital expenditure cannot be simply interpreted as a 40% reduction in demand across the entire industry chain. The actual impact also depends on changes in wafer area, HBM bit demand, and specific product shipments.

NVIDIA is under the greatest pressure, with both future shipments and product selling prices lowered

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

According to the caliber of computing power product revenue, SemiAnalysis lowered NVIDIA's shipment forecasts for 2027 to 2030 by 22%, 39%, 56% and 66% respectively; calculated according to the total power consumption caliber including the entire system, the reduction ranges are 14%, 35%, 50% and 59% respectively.

What is more noteworthy is that the price expectation for NVIDIA's future products has also changed significantly.

SemiAnalysis believes that in addition to the impact of falling HBM costs, NVIDIA also needs to reduce product markup and accept lower gross margins to maintain competitiveness.

Specifically, the estimated selling price of Feynman was reduced 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.

Among them, the expected price reduction of Feynman Ultra exceeds 40%.

This means that the pressure NVIDIA faces in the future is not only the reduction in the number of customer purchases, but also the challenge to its product pricing power. Moreover, this adjustment is not limited to GPUs, but also involves Groq LPU and standalone CPUs.

Why is NVIDIA hit the hardest?

SemiAnalysis explains that large AI customers are having more choices of computing power platforms, and NVIDIA GPUs are no longer the most attractive solution in all scenarios.

Taking Anthropic as an example, the four major computing power platforms it currently uses include Google TPU, Amazon Trainium, AMD GPU and NVIDIA GPU. SemiAnalysis believes that from the combined performance and cost performance, NVIDIA GPUs are in a relatively marginal position among these four platforms, and Anthropic prefers the other three solutions.

OpenAI is also changing its procurement strategy. Although its infrastructure is still dominated by GPUs, the warrant agreement between AMD and OpenAI gives AMD GPUs additional financial attractiveness. In addition, OpenAI will also prioritize deploying its self-developed ASIC Jalapeno and the next-generation Serrano.

Another factor that cannot be ignored is Neocloud (emerging cloud service providers).

NVIDIA has the highest business exposure to Neocloud customers among major suppliers, especially a large number of smaller long-tail customers. These companies usually need to purchase GPUs through financing first, and then recover their investment by renting out computing power. When the capital market tightens, projects that lack procurement commitments from large customers are more vulnerable to impact.

In addition, some of Neocloud's GPU lessees involve Chinese end customers. If the US RASA Act is passed, remote use of GPUs may also be regarded as an export act, and Chinese customers will face the risk of export control when remotely using advanced GPUs.

SemiAnalysis also specifically stated that this forecast adjustment does not take into account the possible impact of Jensen Huang leaving his position as NVIDIA CEO. Even if Jensen Huang remains in office, the agency will make the same downward revision of the forecast. However, it believes that leadership transition may further increase the risk of NVIDIA's market share loss.

Judging from recent delivery status, NVIDIA's product rhythm has also been adjusted.

The Vera Rubin L6 board was postponed to the fourth quarter, and about 100,000 Rubin GPUs were moved from the third quarter to the fourth quarter, but the full-year shipment expectation of Vera Rubin in 2026 still remains at about 980,000 units.

To fill the delivery gap in the third quarter, NVIDIA will increase shipments of about 130,000 GB300 units; in the fourth quarter, as Rubin accelerates volume release, shipment expectations for GB300 and H200 will be reduced by about 260,000 units and 150,000 units respectively.

After adjustment, NVIDIA's total revenue forecasts for the third and fourth quarters of 2026 are about $109.1 billion and $125.3 billion respectively. In addition, the number of Vera Rubin cabinet orders for Microsoft in 2027 has also decreased by 200 units.

Broadcom is relatively stable, MediaTek's TPU business faces reassessment

In contrast to NVIDIA, Broadcom has shown greater resilience in this round of adjustments.

SemiAnalysis lowered its forecast for the cumulative revenue of Broadcom's ASIC business from 2026 to 2030 by 11%, mainly due to the product selling price adjustment brought about by the decline in HBM costs, rather than large-scale cuts in shipments.

Broadcom's advantage is that its main ASIC projects, including Google TPU, OpenAI and Anthropic-related projects, are competitive in terms of performance-to-cost ratio and energy efficiency. More importantly, Broadcom can also assist customers in providing financing support to creditors, covering data center land, power, main building and system construction.

As AI data center construction increasingly relies on capital investment, the ability to help customers solve financing and infrastructure construction problems has become an important competitiveness for chip suppliers.

Of course, some of Broadcom's projects have also been adjusted. The expected scale of OpenAI's second-generation Serrano in 2029 has been reduced by about 1GW, but the shipment expectation of the third-generation product in 2030 has been significantly raised.

There have also been minor downward revisions to Google and Anthropic's TPU projects. Among them, the 700MW TPU originally planned to be delivered to Anthropic in 2027 was postponed to 2028. The reason is that there is uncertainty in the construction progress of some data centers in Texas: the capacity of about 275MW for the Crusoe/Goodnight project may be delayed, and the about 330MW Nexus project in Hubbard has not yet obtained a permit.

Compared with Broadcom, MediaTek's adjustments are more obvious.

SemiAnalysis lowered MediaTek's TPU business expectations, mainly due to reduced demand caused by changes in Anthropic's financing capacity, and the reassessment of overall TPU market demand.

Among them, the TPU v9 series is the most affected. The lifetime shipments of Humufish from 2028 to 2029 have been reduced by about 470,000 units, which mainly reflects the yield risk of the EMIB-T process; the lifetime shipments of TPU v9i (Triggerfish) have been lowered by 42%.

Triggerfish has a special meaning. This chip was designed according to Anthropic's opinions, and was originally expected to become MediaTek's first TPU product delivered to external customer Anthropic.

SemiAnalysis believes that under the circumstance that financing has become the main limiting factor, Broadcom's ability to provide financing support for the overall construction of data centers is more attractive to Anthropic.

MediaTek's advantage lies in the lower overall system cost and lower gross margin than Broadcom, which still makes it competitive when capital is limited. But this requires Google to take more financing responsibilities, or Anthropic to pay more in advance.

It is worth noting that SemiAnalysis remains optimistic about MediaTek's future share in the TPU market, especially the TPUs used internally by Google. Even after this downward adjustment, its forecast for MediaTek's TPU shipments from 2029 to 2030 is still higher than that of Broadcom and AMD/Marvell.

AMD and Amazon Trainium become relatively benefited platforms

AMD's forecast has also been revised downward, but the magnitude is smaller than that of NVIDIA.

Calculated by total power consumption, AMD's expectations for 2028 to 2030 are lowered by 23%, 55% and 57% respectively; calculated by expenditure, they are lowered by 29%, 35% and 32% respectively. The 2027 power consumption expectation is basically flat, and expenditure is reduced by 11%.

AMD is relatively less affected, which is related to its customer structure. Compared with NVIDIA's high exposure to Neocloud customers, AMD's customers are more concentrated in large cloud service providers and AI labs, including Meta, OpenAI and Anthropic.

Among them, both Meta and OpenAI have warrant agreements with AMD, and purchasing AMD systems has additional financial attractiveness. Although Anthropic has no similar agreement, AMD's products are competitively priced. In terms of performance per dollar in inference scenarios, Anthropic ranks AMD after TPU.

At the same time, AMD has also advanced the volume release time of some MI400 products. About 25,000 MI400 units, and about 30,000 Meta-customized MI450-MC units, have been advanced from the fourth quarter of 2027 to the fourth quarter of 2026, reflecting AMD's enhanced expectation for recent mass production ramp-up.

As for Amazon Trainium, shipment forecasts for 2027 and 2028 have not been lowered at all.

SemiAnalysis believes that Amazon has sufficient reserves of data center construction projects and is less affected by changes in the financing environment. In addition, Trainium has low manufacturing costs, and the cost for end customers to use is relatively cheap. Amazon is also willing to provide strategic subsidies for key customers such as Anthropic and OpenAI.

After 2029, Amazon will shift part of Trainium4's shipments to Trainium5, extend the deployment cycle of Trainium5 to 2030, and reduce the volume release scale of Trainium6 at the same time.

Judging from this round of adjustments, the AI computing power market has not lost its growth momentum, but the previous market's expectations for the scale of future investment were obviously too optimistic.

More importantly, the evaluation criteria for competition are changing.

In the past few years, NVIDIA has dominated the AI computing power market with its GPU performance, CUDA ecosystem and supply chain advantages. But as large AI enterprises begin to deploy more self-developed ASICs, and introduce AMD GPUs, Google TPUs and Amazon Trainium, customers' dependence on a single supplier is decreasing.

Under the circumstance that AI infrastructure investment is becoming larger and larger, chip performance is no longer the only determining factor. Product cost, energy efficiency, financing capacity and data center construction progress all begin to affect the final procurement choice.

For the semiconductor industry chain, this round of adjustments also sends a noteworthy signal: the growth of the AI chip market is shifting from simply pursuing the scale of computing power to a stage that pays more attention to return on investment and overall system cost.

The real challenge NVIDIA faces may not just be the downward revision of shipment expectations, but that customers are starting to have more choices.

This article is from the WeChat official account