Original title: Revising Down Accelerator Demand and Shipment Outlook; Nvidia Bears the Brunt
Original authors: Myron Xie and Ivan Chiam
Editor's note: An Nvidia Feynman Ultra carries average selling prices that differ by about $77,000 across two versions of SemiAnalysis's model. On November 13, the semiconductor research firm cut its average selling price forecast for the chip from about $175,000 to $98,000, a reduction of about 44%. Falling high-bandwidth memory (HBM) costs explain part of that, and the rest of the adjustment reflects analysts' new judgment on Nvidia's pricing power. As major customers gain more alternatives, Nvidia may be able to charge a lower premium for its next-generation chips than previously expected.
The revision goes far beyond Nvidia. SemiAnalysis cut its forecast for global AI accelerator spending in 2028 from $1.054 trillion to $758 billion, a reduction of nearly $300 billion. Under the new model, annual spending will still exceed $900 billion by 2030, more than double the 2026 level. The increment remains substantial, but net new annualized recurring revenue at frontier AI labs slowed in the third quarter, and investors have tightened financing for data centers. Projects lacking purchase commitments from major customers often must first resolve who will pay to build the facilities, connect power and buy equipment before chip orders can materialize.
Nvidia faces pressure from two types of customers. Anthropic can allocate compute across Google TPUs, Amazon Trainium and two GPU vendors. OpenAI is advancing its own chip development and can also gain potential upside from its warrant deal with AMD. At the same time, Nvidia has the largest exposure among chip suppliers to emerging cloud service providers. These customers need to raise funds to buy GPUs first, then recover their investment by renting out compute. Top buyers have more choices, and buyers expanding on financing find it harder to place orders. With these two pressures combined, SemiAnalysis cut its forecast for Nvidia's compute product revenue corresponding to 2030 shipments by 66%.
The comparison between Broadcom and MediaTek reveals an often-overlooked calculation in this round of competition. Both participate in TPU projects; MediaTek can offer lower system costs, while Broadcom can support financing arrangements for land, power, facilities and equipment. For cash-strapped buyers, how much a system ultimately costs and how much cash must be put up today are two different procurement thresholds. As chip competition extends to the entire data center, suppliers that can help customers get projects built are also more likely to win orders.
The following is the original text:

·SemiAnalysis has cut its 2028 accelerator spending forecast from $1.054 trillion to $758 billion, with the 2030 forecast about 40% lower than the previous version.
·Net new annualized recurring revenue at frontier AI labs slowed in the third quarter, and debt and equity investors have also tightened financing for AI infrastructure, leaving projects without procurement commitments from major cloud providers particularly vulnerable.
·The downward revision to high-bandwidth memory (HBM) price assumptions is one of the main reasons for the lower accelerator spending expectations. Under the new model, HBM demand measured in bits still grows year over year.
·Both Nvidia's volume and price assumptions were lowered. Measured by compute product revenue corresponding to shipments, the 2027 to 2030 forecasts were cut by 22%, 39%, 56% and 66%, respectively.
·Broadcom's cumulative custom chip revenue forecast for 2026 to 2030 was cut by 11%, mainly due to lower HBM prices. MediaTek's Triggerfish designed for Anthropic saw its full-lifecycle shipment forecast cut by 42%.
·AMD's forward spending forecast was also revised down, but its warrant transactions enhanced its appeal to Meta and OpenAI. Amazon Trainium's 2027 and 2028 shipment forecasts remain unchanged.
We have lowered our forecasts for AI chip and server shipments, spending and power demand across the entire forecast period, with the 2028 to 2030 forecasts more heavily affected. The downgrade reflects adjustments in three areas: lower shipments, lower prices, and lower per-chip power consumption assumptions for future accelerator projects in 2029 and 2030.
Overall, we adjusted our forecasts for the following reasons.
1. We reassessed token revenue expectations across the ecosystem, particularly given the slowdown in net new annualized recurring revenue (ARR) at frontier AI labs in the third quarter. As a result, we lowered our estimates for new supply so that revenue can more sustainably cover overall compute costs, as we noted in a recent report. We remain more optimistic than consensus on these labs' revenue and see many areas where frontier AI capabilities can be deployed with high returns on investment. However, whether infrastructure buildout can obtain financing depends on the overall expectations of capital market participants, which leads to the next point.
2. Capital markets have already tightened, with both debt and equity investors becoming more reluctant to fund capital expenditures related to AI infrastructure buildout, especially projects without commitments from hyperscale cloud providers to purchase. We discussed this issue in our financing framework report. This will reduce data center land development, power infrastructure, and main building construction, with affected projects more concentrated among those lacking hyperscale cloud provider support, and the power available for deploying systems will also decrease accordingly.
3. The results of the U.S. midterm elections have caused 5.8 GW of U.S. data center capacity originally planned for deployment in 2H 2027 and 1H 2028 to be delayed. Further delays caused by regulatory factors are also lowering total data center capacity for 2028. We have also incorporated this factor into our analysis, and it has indeed led to a reduction in near-term compute equipment shipments, but we emphasize that the first two factors are more significant.
The magnitude of the downward revision is not uniform across chip platforms and suppliers. Under our analytical framework, customers will still prioritize chips with the best performance per total cost of ownership (perf/TCO) and performance per watt (perf/W), and then consider financing conditions on that basis. Platforms ranking lowest on these metrics will see larger cuts. As a result, Nvidia is hit the hardest, with the majority of spending cuts concentrated in Nvidia. In addition, among ASIC projects, we expect MediaTek's TPU project will also be affected, as its financing capability is weaker than Broadcom's. The chart below summarizes changes in total accelerator and HBM-related spending.

The new model cuts 2028 accelerator capex forecast by 28%. Total accelerator capex, new vs. old model comparison, in billions of dollars. From 2027 forecast onward, red numbers indicate changes in the new model versus the old model. Yellow is the new version, blue is the old version.
The decline in total spending is primarily driven by lower prices. As shipment forecasts are revised down, HBM supply-demand dynamics become more balanced, so we have also lowered our HBM price assumptions. For more details, see the memory model update published concurrently. To this end, we have incorporated HBM price changes into forecasts for each chip to reflect changes in bill of materials (BOM) costs. For Nvidia's future products starting from Feynman, the average selling price (ASP) reductions also reflect lower markup percentages and changes in HBM costs. This reflects that competition in the AI hardware market has significantly intensified.

Accelerator average selling price (ASP) adjustments, in dollars. Product model, supplier, HBM type, old model ASP, new model ASP, change (%). Source: SemiAnalysis.
A significant portion of the previously forecast 2027 spending increase came from price increases, especially memory price hikes. Therefore, the subsequent price decline means the supply chain is returning to a more "normal" state.
Price changes also explain why the silicon area corresponding to shipments and HBM demand measured in bits are still rising, and growing faster than nominal spending.

Forecast shipments from 2026 to 2030 have been revised down by 17% in total silicon area, with Nvidia accounting for 66% of the reduction.

The 2028 HBM bit demand forecast has been revised down by 30%, and 2030 by 12%.
AI power demand forecasts have also been revised down. This is the combined result of lower shipment volumes and reduced per-chip power consumption assumptions. Since power remains a constraint, we expect future product roadmaps to lower the maximum power consumption of chips. The decline in power demand also means that, under our revised data center model for global (excluding China) power supply forecasts from 2027 to 2029, there will be sufficient power for these shipped devices to complete deployment.

The new model revises down the 2030 accelerator power demand forecast by 34%.

Changes in AI accelerator power demand by platform, in MW.

Changes in total power consumption by chip model, in W.
Next, we will review these changes one by one by supplier and platform.
Measured by compute product revenue, we have significantly revised down Nvidia's shipment forecasts for 2027, 2028, 2029, and 2030 by 22%, 39%, 56%, and 66%, respectively. Measured by total power consumption, they have been revised down by 14%, 35%, 50%, and 59%, respectively. Starting with Feynman, in addition to incorporating HBM cost changes, we have further lowered average selling price assumptions because we believe Nvidia will need to accept lower margins to remain competitive relative to other projects. The forecast price for Feynman has been reduced from approximately $110,000 to approximately $78,000, Feynman Ultra from approximately $175,000 to approximately $98,000, and Feynman Next from approximately $191,000 to approximately $118,000. The simultaneous reduction in prices and shipment volumes means Nvidia's forecast revenue will decline significantly. The reductions cover GPUs, Groq LPUs, and standalone CPUs.
We have made the largest downward revision to Nvidia demand, as Nvidia GPUs rank lower in attractiveness within the compute systems of frontier AI labs. Among the four major compute platforms used by Anthropic—TPU, Trainium, AMD GPUs, and Nvidia GPUs—Nvidia GPUs are the most marginal. In terms of performance per unit cost, Anthropic prefers other platforms.
Although GPUs play a more important role in OpenAI's infrastructure, AMD GPUs have become an attractive option when factoring in the substantial financial benefits from the AMD warrant transaction. In addition, OpenAI will prioritize deployment of its ASIC projects Jalapeno and next-generation Serrano, as these chips already lead in total cost of ownership performance per unit.
Furthermore, Nvidia has the greatest exposure among suppliers to emerging cloud service providers (Neocloud) customers, especially smaller long-tail customers. Today, Neoclouds face much greater risks in both project execution and financing capability. The scale of compute that Chinese end customers have committed to purchase from Neoclouds is also considerable. The likelihood of the RASA Act becoming law appears to be rising. Once in effect, remote access to GPUs will be treated as an export, and Chinese customers' remote access to advanced GPUs will therefore constitute a violation of export controls.
Finally, we note that this model adjustment does not account for the impact of Jensen Huang resigning as Nvidia CEO. Even if Jensen Huang remains CEO, we would make the same adjustment. Given Jensen Huang's importance to Nvidia, we believe this leadership transition will increase the risk of market share decline, potentially exceeding the forecasts we list here.
Because shipments of Vera Rubin L6 boards were delayed from Q3 to Q4, we also adjusted our Q3 and Q4 2026 forecasts. We moved approximately 100,000 Rubin GPUs from Q3 2026 to Q4 2026, maintaining full-year 2026 Vera Rubin shipments at approximately 980,000 units. GB300 adds approximately 130,000 units of shipments in Q3 to fill the gap. As Rubin ramps up, we reduced approximately 260,000 GB300 units and approximately 150,000 H200 units in our Q4 2026 forecast. Combining these adjustments, Nvidia's total revenue for Q3 2026 is approximately $109.1 billion, and Q4 is approximately $125.3 billion.

Shipment forecast (Nvidia). The table is updated as of November 13, 2026, listing shipments by model, average selling price, and revenue by quarter, with percentage changes versus the prior version. Summary items include Nvidia data center GPU revenue included in the compute business and Nvidia total company revenue. Source: SemiAnalysis.
Broadcom is least affected, with cumulative ASIC revenue forecasts for 2026 to 2030 revised down 11%, mainly due to lower average selling prices driven by reduced HBM costs, with shipment cuts having a smaller impact. We believe Broadcom's demand will not decline, because its three major ASIC projects (TPU, OpenAI, and Anthropic) can provide end customers with higher performance per unit cost and performance per unit power, while Broadcom also has the ability to backstop financing arrangements provided to creditors, raising funds for land, power, main structures, and systems. For OpenAI, we reduced the 2029 forecast for second-generation Serrano by 1 GW, while significantly raising shipments for the third-generation product in 2030.
We have slightly lowered our TPU shipment forecasts, affecting Google and Anthropic. As we expect data center development in Texas to pause, we have also pushed back the delivery of 700 MW of TPUs originally scheduled for Anthropic in 2027 to 2028. Our data center team expects that approximately 275 MW of capacity from the Crusoe/Goodnight project will be delayed from Q4 2027, and another 330 MW from the Nexus project in Hubbard has not yet been approved and is also at risk. Both projects are intended for deploying Anthropic's TPUs.
Another Broadcom project for which we have lowered our forecast comes from the startup MatX. We believe MatX faces higher uncertainty and greater difficulty in obtaining financing, so we have lowered the estimated demand from Anthropic for MatX chips.

Broadcom demand. The table is dated November 13, 2026, and lists quarterly shipment volumes, revenue, and percentage changes versus the previous version for projects including TPU, OpenAI, and Anthropic. The summary item is Broadcom AI ASIC revenue. Source: SemiAnalysis.
We have cut MediaTek's shipment forecasts by a larger margin.
This reflects both a decline in Anthropic's demand due to financing factors and our readjustment of total TPU demand, as in our model MediaTek will handle most TPU shipments in the future, especially Google's internally used TPUs. Even after this cut, MediaTek's TPU shipments in 2029-2030 remain higher than those of other partners such as Broadcom and AMD/Marvell. In addition, because MediaTek's customers procure HBM themselves, HBM price declines have no impact on MediaTek, so the change in revenue forecasts comes entirely from the shipment cut.
The cuts are mainly concentrated in two TPU v9 products, Humufish and Triggerfish. We have lowered Humufish's total lifecycle shipments for 2028 and 2029 combined by 470,000 units, mainly reflecting continued risks to EMIB-T yields.
Given that financing constraints are one of the main reasons for this adjustment, we believe Broadcom's ability to provide financing support for the entire construction project is more important to Anthropic. Therefore, the project we have cut the most is TPU v9i (Triggerfish), whose total lifecycle shipments have been lowered by 42%. Triggerfish was designed with input from Anthropic and was originally planned to be MediaTek's first TPU sold externally to Anthropic.
However, we also recognize that MediaTek's overall profit margin is lower than Broadcom's, resulting in a lower total system cost that is attractive when capital is constrained. This would require Google to bear more financing responsibility, or Anthropic to increase upfront payments.
For the subsequent TPU v10t (Icefish), we have slowed the 2029 ramp and shifted more shipments to 2030.

MediaTek demand. The table, updated as of November 13, 2026, lists quarterly shipments, revenue, and percentage changes from the previous version for each TPU model and the ByteDance ASIC project. The summary item is MediaTek's total revenue. Source: SemiAnalysis.
On a comprehensive power consumption basis, we have cut our AMD forecasts for 2028, 2029, and 2030 by 23%, 55%, and 57%, respectively, and on an expenditure basis by 29%, 35%, and 32%, respectively. The 2027 power forecast is roughly flat, with the expenditure forecast cut by 11%. Although AMD is also a supplier selling commercial chips to the open market, its revenue cut is smaller than Nvidia's because AMD has greater exposure to hyperscaler and frontier AI lab customers, including Meta, OpenAI, and Anthropic. For Meta and OpenAI, the warrant deals make purchasing AMD systems an almost no-brainer choice. Anthropic has no warrant deal, but we believe AMD's pricing remains very favorable for Anthropic. In terms of performance per dollar for inference, AMD is second only to TPU in attractiveness to Anthropic.
In addition, for the near term, we have moved up the MI400 ramp to reflect stronger production ramp in the fourth quarter of 2026. We have shifted approximately 25,000 MI400 units and approximately 30,000 Meta-customized MI450-MC units from the fourth quarter of 2027 to the fourth quarter of 2026.

AMD demand. The table, updated as of November 13, 2026, lists quarterly shipments, revenue, and percentage changes from the previous version for each model, along with the percentage change in average selling prices. The summary item is AMD data center GPU revenue. Source: SemiAnalysis.
We maintain our Trainium shipment forecasts for 2027 and 2028. For 2029, we shift the shipments expected to be borne by Trainium4 to Trainium5, further extend Trainium5's lifecycle to 2030, and lower the Trainium6 ramp scale. We maintain our near-term Trainium shipment forecasts because Amazon's pipeline of data center projects under construction and in preparation is very ample and largely unaffected. In addition, Trainium's manufacturing cost is much lower, so the cost to end customers is also low. Amazon is also willing, for strategic reasons, to subsidize projects using Trainium for key customers such as Anthropic and OpenAI.
We have renamed TPU v9i (Whalefish) to TPU v8i Max (Whalefish).
We have renamed "B300 1200W" to "B300 1100W" and updated the thermal design power (TDP) accordingly. The chip's overall power consumption remains unchanged.
In the ODM model, we have cut Microsoft's 2027 Vera Rubin rack orders by 200 units.
Welcome to join the official BlockBeats community:
Telegram Subscription Group: https://t.me/theblockbeats
Telegram Discussion Group: https://t.me/BlockBeats_App
Official Twitter Account: https://twitter.com/BlockBeatsAsia