Custom AI Accelerator Market Size and Forecast 2025-2030: A $15.7 Billion Market in 2025 Where Chip Suppliers Pay Hyperscalers in Stock for Second-Source Seats

Chip suppliers now pay Amazon and Google in stock for second-source seats in a custom AI accelerator market worth $15.7 billion in 2025 and about $233.0 billion by 2030 on the base path, a 71.5% CAGR.

On September 8, 2026, Qualcomm announced a multi-generation custom-silicon and 1.6T optical collaboration with Amazon. It had already granted Amazon a warrant on up to 25 million Qualcomm shares that vests as Amazon commits to and makes purchases, up to $60 billion of payments. Thirty-six days before that warrant, Marvell signed the agreement behind a warrant to Google that vests one tranche per $500 million of custom revenue, up to $120 billion. The supplier pays the hyperscaler for its second-source seat in the supplier's own stock, a rebate the report prices as a share of the purchases it unlocks: insurance for the buyer, not a shift in share.

Custom AI accelerators, the chips a design partner builds for a single cloud provider or AI lab, were a $15.7 billion market in 2025 at supplier revenue. On a path that follows the leading supplier's own guidance through 2028, the market reaches about $233.0 billion by 2030, a 71.5% CAGR for 2025-2030 that is front-loaded.

The 80-page report segments the market by workload, supplier and supplier type. It profiles 12 companies, including suppliers Broadcom, Marvell, Qualcomm, MediaTek, Alchip and Global Unichip alongside program owners Google, Amazon Web Services, Meta, Microsoft and OpenAI. It carries 14 charts, 11 data tables, three scenarios and a year-by-year forecast to 2030.

Report Highlights:

• The Supplier-Pays Contract: Chip suppliers now pay hyperscalers in stock for second-source seats. Marvell's warrant to Google vests 240,042 shares for each $500 million of custom revenue, up to $120 billion through January 2033. Qualcomm's warrant to Amazon vests against up to $60 billion of purchases. Under US accounting rules, such warrants reduce the issuing supplier's own reported revenue.
• Market Sizing Built from Supplier Revenue: The custom AI accelerator market was $15.7 billion in 2025 at supplier revenue and reaches about $233.0 billion by 2030 on the base path, a 71.5% CAGR for 2025-2030. The report brackets the 2025 estimate with low and high bounds for every input and builds conservative and optimistic scenarios around the 2030 figure. Growth is concentrated in 2026-2028, which follow the leading supplier's own guidance, and slows after 2028, when that guidance ends. The report extends the base path to 2035 as a separate, explicitly speculative view.
• Inference and Training, Sized Separately: Inference is the larger workload by the report's estimate, and its share of market value rises through 2030; the report sizes inference and training separately for 2025 and 2030. The largest new programs, Google's TPU v8i, OpenAI's first custom accelerator and Meta's MTIA, are described by their supplier as optimized for, or outperforming GPUs on, inference workloads.
• Second Sources Without a Share Shift: On the base path, the second-source deals widen the bench without narrowing the leader's lead. One supplier, Broadcom, accounted for the large majority of 2025 supplier revenue by the report's estimate. It guides $115 billion and $230 billion of AI chips to its six custom-accelerator customers in fiscal 2027 and 2028, nearly six times Qualcomm's ten-year purchase ceiling.
• What Full Vesting Would Take: Warrant ceilings are options rather than forecasts. Full vesting of Marvell's warrant would need about $18.5 billion a year of Google custom revenue, more than twice Marvell's entire fiscal 2026 revenue of $8.2 billion. The head-on catalogue estimate for this market, $9.25 billion for 2025, is below Broadcom's estimated custom-accelerator revenue that year, although the catalogue's stated scope is wider.

This report will provide answers to the following questions:

• How large was the custom AI accelerator market in 2025 at supplier revenue, and how fast does it grow to 2030?
• What do the Marvell-Google and Qualcomm-Amazon warrants commit, and what is a second-source seat worth to the supplier that pays for it?
• How do customer warrants affect a supplier's reported revenue under US accounting rules?
• How is market value split between inference and training, and how does the split change by 2030?
• How is the market shared among full-service partners, turnkey design houses and platform entrants, and which second sources sit at each program owner?
• Why are warrant ceilings options rather than forecasts, and how does the supplier-revenue estimate compare with catalogue estimates?
• What limits deployment: power, data-center shells, memory and financing-dependent demand?

Methodology

Our analysis originates from primary research - direct interviews with executives, operators, and technical practitioners actively shaping these markets. This fieldwork provides access to perspective and data not available in secondary sources: what decision-makers are observing in real time, the problems driving purchasing behavior, and where they see value migrating. Every data point and claim undergoes human verification before inclusion; figures that cannot be substantiated or traced to credible sources are excluded.

Market sizing triangulates across multiple independent estimation methods, producing investment-grade estimates with assumptions documented explicitly so readers can evaluate the underlying logic, stress-test key inputs, and defend the numbers in boardrooms and diligence processes. We validate quantitative claims against peer-reviewed research, regulatory filings, and observable market signals - including systematic searches for contradicting evidence. Where methods produce divergent estimates, we investigate the source of variance and report ranges rather than false precision. Forecasts are constructed through scenario modeling anchored to base rates from comparable markets. (While every effort has been made to ensure accuracy, forward-looking statements reflect current expectations and are subject to risks, uncertainties, and assumptions that may cause actual results to differ materially.)

The result is thesis-driven analysis that delivers clear conclusions: specific enough to cite, transparent enough to verify, comprehensive enough to satisfy diligence requirements, and rigorous enough to withstand the follow-up question.

Target Audience

• Strategy and corporate development teams at semiconductor companies and ASIC design houses
• Silicon, infrastructure and procurement leaders at cloud providers and AI labs
• Equity analysts and investors covering semiconductors and AI infrastructure
• Foundry, advanced-packaging and memory suppliers planning capacity for custom accelerators
• Finance and accounting professionals assessing customer warrants and revenue recognition