SemiAnalysis Cuts 2030 AI Accelerator Capex to $923B | Magica
SemiAnalysis Cuts 2030 AI Accelerator Capex Forecast to $923 Billion
Editorial Team
••📖8 min read
SemiAnalysis has reportedly cut its 2030 global AI accelerator spending forecast by about 40% to $923 billion, with Nvidia facing the largest revisions. Lower price assumptions and financing constraints help explain the reduction, while customer deals show both the strength and financial cost of competing chip platforms.
The reported 2030 global AI accelerator capex forecast falls by $608 billion, from $1.531 trillion to $923 billion.
Nvidia faces the largest revisions; custom chips and AMD fare comparatively better in the reported model.
SemiAnalysis, a semiconductor research firm that sells models tracking chip demand and data-center construction, has cut its forecast for global AI accelerator capital expenditure in 2030 to $923 billion, according to a November 16 account of its latest model update. The roughly 40% reduction removes $608 billion from its previous estimate for that year.
The revision changes the outlook for the specialized processors that supply AI computing capacity, combining lower expectations for hardware demand with lower prices. It does not establish a year-over-year spending decline or show that customers have canceled $608 billion of orders: the comparison is between two forecasts for 2030.
SemiAnalysis’s reported revisions compare forecasts for the same years, rather than year-over-year spending declines. Source: EE Times China.
A smaller spending forecast has several moving parts
The reported cuts become especially large from 2028 onward. The 2028 global accelerator capex estimate falls 28%, from $1.054 trillion to $758 billion; the 2029 estimate is reduced 41%. The reported update also cuts cumulative accelerator logic-silicon shipment area for 2026–2030 by 17%.
Manufacturers, utilities and miners linked to data center construction have shed $4.2 trillion in market value since June. Strong supplier orders contrast with financing concerns, project delays and political resistance.
Editorial Team
Global AI accelerator capex, 2030
$1.531 trillion
$923 billion
Global high-bandwidth memory shipments, 2028
9.0 exabytes
6.4 exabytes
AI accelerator power demand, 2030
122.8 gigawatts
80.7 gigawatts
These measures describe different things. Silicon area measures a physical input, memory shipments measure memory capacity, and power demand measures an electrical requirement. None is interchangeable with spending or the number of chips shipped. Calculated from the displayed inputs, memory demand falls about 29% and power demand about 34%; the report describes the memory reduction as 30%.
SemiAnalysis also expects easing high-bandwidth memory supply constraints to normalize prices and reduce hardware bills of materials, according to the fuller report. That matters because a lower dollar forecast can reflect cheaper components as well as less hardware.
Revenue and financing constrain the buildout
The fuller account attributes three linked explanations to SemiAnalysis. First, third-quarter growth in annualized recurring revenue at frontier AI laboratories slowed, prompting concern that revenue from selling model tokens—the units of text processed or generated—would not keep pace with infrastructure costs.
Second, investors became less willing to finance AI projects without purchasing commitments from major cloud providers. In the model's explanation, that restrains spending on land, power and data-center shells before accelerators can be installed.
The financing mechanism predates this revision. In its October 19 construction analysis, SemiAnalysis said developers typically spend only 5%–10% of project capital before securing a customer. A pre-lease—typically a 15–20-year customer contract—helps them raise the debt needed to complete construction. A weaker appetite for uncommitted projects can therefore constrain chip demand before any chip order is placed.
Third, regulatory factors associated with the U.S. midterm elections are expected to delay roughly 5.8 gigawatts of data-center capacity. The report says SemiAnalysis treats revenue and financing constraints as the more important causes. It does not provide the underlying laboratory revenue series or enough project-level detail to independently test those causal weights.
The October analysis also limits the cooling narrative. SemiAnalysis said it tracked more than 24 gigawatts of data centers in advanced construction across 40 campuses, mostly intended for delivery in 2027. It argued that its proprietary data-center model still pointed to material acceleration in 2027 capacity despite construction and financing difficulties. That earlier physical-construction evidence and the new accelerator spending forecast cover different scopes and horizons; the latter cannot by itself establish that construction already underway has stopped.
SemiAnalysis’s company-reported October 2026 model projected accelerating datacenter capacity additions outside China. This earlier outlook measures IT capacity rather than accelerator spending. Source: SemiAnalysis.
Nvidia faces pricing pressure as well as supplier substitution
Nvidia, the established data-center graphics processing unit (GPU) supplier competing with AMD and cloud companies' custom AI processors, accounts for 66% of the cumulative silicon-area reduction in the reported update. That is its share of the reduction, not a 66% cut to global shipments. The fuller account separately gives reductions of 22%, 39%, 56% and 66% for Nvidia's 2027–2030 shipment forecasts measured on a compute-revenue basis. Those percentages should not be read as equivalent reductions in physical chip counts.
The reported pricing assumptions also fall sharply. SemiAnalysis lowers the assumed average selling price of Feynman from $110,000 to $78,000 and Feynman Ultra from $175,000 to $98,000. These are analyst assumptions about future products, rather than announced transaction prices.
The report attributes Nvidia's greater exposure to competition from Google's TPU and Amazon's Trainium custom accelerators and AMD GPUs, alongside its exposure to smaller specialist cloud operators with greater financing risk. It attributes to SemiAnalysis the assessment that major AI developers including Anthropic and OpenAI increasingly favor alternatives on cost and performance grounds. The accounts do not supply like-for-like workload benchmarks or customer operating costs to establish the size of that advantage.
That competitive mechanism has a documented precedent, although it does not prove the magnitude of the new forecast. Anthropic, the AI developer behind Claude, announced in October 2025 plans to use up to one million Google Cloud TPUs. It valued the expansion at tens of billions of dollars and expected well over a gigawatt online in 2026. Google Cloud said the decision reflected price-performance and efficiency Anthropic had seen over several years; that is the supplier's explanation, not an independent cost comparison.
Anthropic described using Google TPUs, Amazon Trainium and Nvidia GPUs together, and retained Amazon as its primary training partner and cloud provider. The announcement documents an existing multi-supplier strategy and a planned expansion, rather than a completed replacement of Nvidia.
The reported model lowers custom-chip supplier Broadcom's cumulative 2026–2030 custom-chip revenue forecast by 11%, primarily because of cheaper memory rather than fewer shipments. It attributes relative resilience to energy efficiency and financing support for projects such as Google TPU. Amazon Trainium's 2027 and 2028 shipment estimates remain unchanged, supported by available data-center capacity and cost advantages.
AMD, the semiconductor company selling Instinct AI GPUs alongside server and PC processors, also has its 2028–2030 forecasts reduced. The report describes a smaller impact than Nvidia's, citing warrant agreements with Meta and OpenAI and earlier volume-production assumptions for MI400 and Meta's customized MI450-MC, in the fourth quarter of 2026. Those production timings are part of the reported model assumptions.
AMD's August 2026 quarterly filing shows the scale and incentives behind those relationships. Under agreements reached in October 2025 and February 2026 respectively, OpenAI and Meta each intended to deploy up to six gigawatts of AMD data-center GPUs, with each initial gigawatt using Instinct MI450 products. Each customer received warrants to buy up to 160 million AMD shares for one cent each, subject to purchase milestones, stock-price requirements and other conditions. No tranches had vested or become exercisable as of June 27. These are conditional equity incentives and deployment intentions, not delivered capacity.
AMD also disclosed $4.1 billion of maximum gross exposure at June 27 from guarantees of commercial partners' data-center lease obligations, with terms up to 15 years. Such guarantees typically become payable on a partner's default. Financing support can help secure infrastructure, while leaving the supplier exposed if a partner cannot pay.
Nor does the filing portray competition as settled. AMD said Nvidia's data-center GPU position, financial resources and proprietary software ecosystem can limit customers' ability to choose alternatives. AMD also warned that its own customers may lack capital, power or data-center capacity. These constraints qualify the model's relative optimism about Nvidia's challengers.
SemiAnalysis’s reported average selling price assumptions fall for both products. These are analyst estimates, rather than announced Nvidia prices or observed transactions. Source: Anue.
What would establish whether the revision is right
The central unresolved question is how much of the spending reduction comes from fewer deployments, how much from lower prices, and how much from buyers choosing different suppliers. The available accounts do not provide a complete reconciliation of those effects or the full underlying model.
The next consequential decisions are whether lenders finance projects against firm customer commitments and whether AI developers convert planned capacity into installations. Laboratory revenue disclosures, completed data centers, realized accelerator prices and the supplier mix will test the assumptions. A model reconciliation separating price, volume and substitution would show which part of the $608 billion revision represents less hardware and which represents a lower bill for computing capacity.
RIKEN has stopped Fugaku’s research jobs after an intrusion into its operational network. Up to 4,597 users’ account information may be affected, and intruders tried acquired credentials on other systems; research-result leakage remains unconfirmed.
DeepSeek users report daily caps, shorter inputs and restrictions on resuming inactive conversations. Reported financing and a planned data center do not establish when more capacity will reach the free chat service, whose quota rules remain unclear.
Google found an attacker framework configured to compare software patches, develop exploits and execute them without human intervention. The finding sharpens the pressure on defenders to patch quickly, but does not establish that entire attacks now run autonomously.
Corma’s $125 million Series A brings total funding to $185 million as the defensive cybersecurity AI lab confronts a computing bottleneck. Its specialized-model strategy faces existing security agents and an unresolved test of performance and deployment costs.
Ask Maps is adding shareable travel schedules, saved-list availability searches and U.S. appointment finding. The update extends an existing AI planning service rather than introducing trip planning to Maps.
Federal Reserve Governor Michael S. Barr says banks report AI agents attempting account actions without verifiable permission. Autonomous deposit transfers could accelerate future bank runs, but their scale and speed remain unmeasured in his speech.
Amazon-owned Zoox is opening its Las Vegas robotaxis to Uber riders while keeping its own app. The November launch follows an earlier summer target; a joint MGM Resorts agreement addresses resort pickups, while fleet size and financial terms remain undisclosed.
Germany’s reported military AI shortlist excludes Palantir, while France and the Netherlands pursue European alternatives. The challenge is to secure national control over data and infrastructure while delivering usable systems that work with allies already adopting the US supplier.
China aims to make school AI education broadly available by 2030. National guidance already calls for pupil assessment, but an investigation finds common measures of AI literacy still lag an uneven rollout spanning courses, tutoring and student monitoring.
Drive Capital led Zus Health's $33 million Series B, extending a shared-record strategy backed by earlier financing and EHR partnerships. The company is expanding AI workflows and pursuing a more direct role in national data exchange, while disclosed activity figures leave customer returns unmeasured.
Atomarine’s $19 million seed round and Mocean Energy’s wave-powered computing proposal target land, water and electricity constraints. Earlier underwater projects show technical promise, but do not establish the costs or performance of these new floating designs.
Tencent’s new Hunyuan model charges 2 yuan per million output tokens and targets everyday office work. Its claimed 95% task-cost saving against Hy4 comes from an internal benchmark, with a trade-off on long coding tasks.
The Indian chip startup is funding programmable image sensors and targeting production in 2027. Its promised system savings remain unproven, while Sony's earlier designs challenge broad claims of novelty.
Baidu’s AI cloud infrastructure revenue nearly doubled in the third quarter, cushioning a 22% advertising decline. But AI applications were flat, app users fell, and the return to reported profit obscured lower adjusted earnings and a wider annual free-cash-flow deficit.
The Trump administration reportedly plans to designate more than two dozen AI companies to help win overseas business. Existing export rules support integrated technology packages, but guarantee neither financing nor licenses and leave room to support competing suppliers.
Singular leads the Danish accounting software company's seed round as it builds a London commercial hub and targets Benelux. Its pitch rests on automating preparation while accountants retain sign-off, with claimed efficiency gains still awaiting independent verification.
Alibaba says Qwen4-Max reduces average task costs by 62% on its workplace benchmark. The hosted launch builds on its predecessor’s agent capabilities as Alibaba expands a growing cloud business and bears the expense of AI infrastructure.
The FBI warns that ARTEX can accelerate intrusions, with some reaching privileged access in under three hours. Microsoft also documents unsuccessful mass targeting, while compromised backup-service credentials opened a separate route into customer environments.
Jensen Huang has reportedly discussed returning to Nvidia three days after resigning. Supporters want reinstatement during an investigation, while directors have installed interim leadership and begun a permanent CEO search.