Semiconductor research firm SemiAnalysis has cut its 2028 forecast for global AI accelerator spending from $1.054 trillion to $758 billion. According to Coincamps, the forecast model the firm revised on November 13, 2026 reflects a downward adjustment of $296 billion, or about 28%, from the previous projection. Even after the revision, however, the annual forecast still shows 2028 spending rising about 21% year over year.

The key point of this revision is that a lower spending forecast does not necessarily mean fewer AI chip shipments. The cost of buying accelerators depends on both shipment volumes and selling prices, and price swings in high-bandwidth memory (HBM) in particular affect overall system costs. Securing data center power capacity also does not put computing resources into operation unless construction financing is raised and equipment is delivered. To understand this forecast revision, it is necessary to look separately at chip prices and shipment volumes, as well as at buyers' ability to raise funds and at data center operating status.

AD

2028 AI accelerator spending forecast cut by $296 billion

According to an annual table published by BigGo Finance, SemiAnalysis's revised forecast still shows global AI accelerator spending continuing to grow through 2030. Growth from 2028 to 2029 slows to about 5%, however, before recovering to about 16% in 2030. The 2030 forecast is $923 billion, roughly 2.35 times the 2026 figure of $392 billion.

Year Before revision ($100 million) After revision ($100 million)
2026 Not listed 3,920
2027 7,700 6,270
2028 10,540 7,580
2029 13,520 7,960
2030 15,310 9,230

The roughly 28% downward revision for 2028 is the $296 billion difference as a share of the old forecast of $1.054 trillion. The roughly 21% year-over-year growth rate is calculated from the revised 2027 forecast of $627 billion and the 2028 forecast of $758 billion. Both calculations are based on rounded figures published by the news outlets.

These figures are based on the SemiAnalysis revision as presented by Coincamps and BigGo Finance. SemiAnalysis's published description of the model shows that it tracks AI accelerator shipment volumes and average selling prices by manufacturer and product, and analyzes HBM supply, wafer starts, advanced packaging and system deployment. However, the company's public page does not include the revised figures or a detailed breakdown of the equipment costs included in the spending forecast.

Revisions to HBM prices and shipment volumes weigh on spending forecast

According to the revisions published by Coincamps, the biggest factor pushing down the AI accelerator spending forecast was a lower price assumption. SemiAnalysis estimated lower shipment volumes than before, which it expects to bring HBM supply and demand closer to balance and bring prices down. Reflecting this price change in the component cost of each accelerator, the result was that lower unit prices, in addition to lower shipment volumes, pushed down total spending.

The downward revisions to shipment-related indicators are not uniform, however. According to Coincamps, the cumulative forecast for the silicon area of logic chips for AI accelerators shipped over the five years from 2026 to 2030 was cut 17% from the previous forecast. HBM demand, measured in bits equivalent to memory capacity, was revised down about 30% for 2028 and 12% for 2030, but even in the revised model the volume of demand itself is expected to grow year over year. Note that calculating the number of wafers needed for manufacturing from silicon area also requires taking into account individual die sizes and manufacturing yields.

The outlook for power demand was also lowered. The 2030 power demand forecast for AI accelerators fell about 34% from the previous forecast, reflecting lower shipment volumes as well as a lower estimate of power consumption per future product. The power actually required depends not only on when accelerators ship but also on when the destination data centers begin operating.

By manufacturer, the impact on Nvidia is large. According to Coincamps, SemiAnalysis cut its revenue forecast for Nvidia's compute products corresponding to 2030 shipments by 66% from the old model. In addition to shipment volumes and HBM component costs, it also estimates lower average selling prices for next-generation products, taking into account that customers can choose competing products. The revisions released so far make clear that both volumes and prices were revised down, but they do not reveal how much each contributed to the lower revenue forecast.

This shift in the competitive environment is also visible in the procurement agreements of major AI companies. Anthropic has signed agreements with Google and Broadcom for multiple gigawatts of next-generation TPU capacity expected to come online from 2027. Under its agreement with Amazon, it also plans to secure up to 5GW of computing resources for training and inference, including multiple generations of Trainium. The company also uses Nvidia GPUs to train and serve Claude, and its approach is to use multiple accelerators depending on the purpose.

OpenAI has also announced an agreement to deploy a total of 6GW of AMD GPUs over multiple years. The equity warrants AMD granted vest in stages according to deployment and purchase milestones, as well as AMD's share price targets and technical and commercial conditions. OpenAI has also announced a collaboration with Broadcom to deploy 10GW of its own custom-designed AI accelerators, although that was announced as the signing of a term sheet toward rack deployment. Even when expressed in the same unit of gigawatts, the certainty of actual orders and operation differs between a ceiling on usage under a contract, a multi-year deployment plan, and the scale targeted in a collaboration.

On the demand side, the revenue growth of leading AI developers and the financing environment for data center construction lie behind the forecast revision. According to Coincamps, SemiAnalysis's analysis is that the pace of net additions to annualized recurring revenue (ARR) at leading AI companies has slowed. It also sees financing, whether through debt or equity, becoming harder for projects without purchase commitments from major cloud providers and others. The firm also built into its forecast that some data centers will be delayed in coming online after the U.S. midterm elections, but it cites slower revenue growth and a tougher financing environment as the main factors.

AD

Anthropic's compute and infrastructure costs are growing faster than revenue

The burden on AI companies of securing computing resources can also be seen in Anthropic's financial information. According to the company's IPO fact sheet (PDF), revenue in the July–September 2026 quarter was $17 billion, up about 47% from $11.6 billion in the April–June quarter. Compute and infrastructure costs, meanwhile, rose about 77%, from $5.6 billion to $9.9 billion, growing faster than revenue.

As a result, revenue per dollar of compute and infrastructure cost fell from $2.08 in the previous quarter to $1.72. According to Anthropic, the July–September quarter included $5 billion of computing resources procured under short-term contracts. These cost more than securing equivalent resources under multi-year contracts, partially offsetting the cost savings from improved compute efficiency.

What this financial information shows is that while Anthropic rapidly expanded revenue, the cost of procuring computing resources rose even faster. Judging the slowdown in net ARR additions across leading AI companies that SemiAnalysis points to would require comparing the revenue trends of multiple companies on the same basis. Anthropic's results do show, however, that securing computing resources under short-term contracts to meet growing demand can make the cost burden heavy depending on procurement terms, even as revenue rises.

CoreWeave's results show the gap between contracted and operating power

CoreWeave's capital spending offers a reference point for judging whether AI data center construction plans translate into actual supply of computing resources. According to the company's July–September 2026 earnings materials filed with the SEC, contracted power capacity exceeded 4.6GW, while active power capacity exceeded 1.7GW. Both figures are as of the same reporting date, and "exceeded" indicates the lower bound of the disclosed capacity. Active power capacity tells us the scale of equipment operation, but confirming the number of GPUs delivered or their utilization requires other data.

Operating a data center requires not only securing power but also upfront investment in land, buildings, power infrastructure and servers. CoreWeave said it is using financing backed by customer contracts and operating cash flow, and that it expects to receive about $9 billion in prepayments and deposits from customers over the next three months. If it can secure such funds, they can be applied to construction and equipment procurement costs that are needed before service begins and revenue is earned.

The company put its 2026 capital expenditure forecast at $37 billion to $39 billion, raising the low end by $2 billion from the previous forecast. The reason is its outlook that infrastructure buildout and handover to customers will proceed faster than originally planned. For companies able to turn large customer contracts into financing, increasing capital spending to bring computing resources online sooner may also occur. Whether contracted power is translating into actual supply of computing resources can be confirmed by future increases in active power capacity.

AD

Alphabet's higher capital spending and falling AI chip prices can coexist

While AI accelerator spending forecasts are being cut, major cloud providers continue to plan enormous capital expenditures. According to Alphabet's July–September 2026 earnings call, capital expenditure in the quarter was $51.4 billion, with the majority directed to technical infrastructure centered on AI. About 65% of technical infrastructure investment went to servers and networking equipment, and the remaining 35% to data center-related spending. The 65% figure uses technical infrastructure investment as its denominator, and the servers and networking equipment it covers include items other than AI accelerators.

Alphabet expects full-year 2026 capital expenditure of $195 billion to $205 billion, and expects a significant further increase in 2027. If spending on data center construction and networking equipment grows, a company's overall capital expenditure can rise even as the average selling price of accelerators falls. The company explained that as technical infrastructure investment expands, depreciation and operating costs such as electricity will also increase in subsequent accounting periods.

On the other hand, the cash flow in the business of selling TPU systems to outside customers differs from capital spending on its own data centers. According to Alphabet, when a sales contract is signed with a customer, it is first recorded in the backlog, and the company then builds up inventory of systems that incorporate not only compute equipment but also storage and networking. Because payment is collected after delivery to the customer and invoicing, equipment procurement spending occurs before payment from the sale. In fact, in the July–September quarter, the inventory buildup in preparation for external sales of TPU systems weighed on operating cash flow. Capital spending for building out its own data centers and the cash burden from procuring inventory of systems for external sale appear in different line items in the financial results.

To evaluate this SemiAnalysis forecast revision, it is important to track not only the amount invested in AI infrastructure but also whether contracts are progressing into actual orders, equipment deliveries and operation. In particular, a key factor is whether companies that have built up long-term contracts and backlogs are securing the necessary funds through loans and customer prepayments. Combining this with AI accelerator shipment volumes and average selling prices would allow a more accurate understanding of whether changes in spending come from demand volume or from price changes. If equipment prices fall while deliveries and operation advance, the computing resources actually available could grow faster than spending on AI accelerators.