AI Hardware Market Size and Share (2025)

How big is the AI hardware market?

The global AI hardware market was worth $182.0 billion in 2025, according to Global Market Insights (GMI). GMI projects it will reach $323.6 billion in 2026 and $943.7 billion by 2035, growing 12.6% a year from 2026. That is 5.2 times the 2025 figure. GMI also estimates that 15.0 million data center AI accelerators were shipped in 2025.

GMI's figure is a compute-only measure. It counts accelerator systems and components at their first external sale, including the high-bandwidth memory, packaging and board content built into them. It leaves out standalone memory, storage, networking, general-purpose CPUs, software, cloud services and chips that companies design for their own use, such as Google's TPUs. Other research firms define the market differently, so their AI hardware and AI chip market sizes for the same years vary widely.

How fast is the AI hardware market growing?

GMI forecasts 12.6% compound annual growth from 2026 to 2035, after a 78% jump from $182.0 billion in 2025 to $323.6 billion in 2026 (derived from GMI's two figures). GMI does not explain the size of that one-year step on its summary page. Growth is uneven across segments: GMI expects NPUs and edge inference chips to grow 35.9% a year and edge infrastructure 29.2% a year, from small bases of $2.5 billion and $3.6 billion.

Figure 1 sets the forecast against the chip industry's 2025 sales. In 2025, AI compute hardware was smaller than global memory chip sales ($223.1 billion) and logic chip sales ($301.9 billion), as reported by the Semiconductor Industry Association. GMI's 2026 figure is already larger than both, and its 2035 forecast is larger than all semiconductor sales in 2025 ($791.7 billion). The chip industry is growing too: SIA expects global sales of roughly $1 trillion in 2026.

What is NVIDIA's share of the AI hardware market?

NVIDIA held 85.05% of global AI hardware revenue in 2025, according to GMI. Broadcom, which designs custom AI chips for cloud companies, had 6.59%, and AMD had 4.01%. The top five vendors, including Marvell and Qualcomm, held 96.8% combined, leaving 3.2% for everyone else.

Company filings show the same imbalance. NVIDIA's data center revenue was $193.7 billion in its fiscal year ended January 2026, up 68% and 12.9 times its $15.0 billion three years earlier. AMD's Data Center segment, which also includes server CPUs, brought in $16.6 billion in 2025. Broadcom said its AI revenue grew 65% to $20 billion in its fiscal 2025, and Marvell's data center revenue was $6.1 billion.

GMI's share implies NVIDIA earned about $154.8 billion of the $182.0 billion market in 2025 (derived). NVIDIA itself reported $162.4 billion of data center compute revenue for a fiscal year that runs one month later. The two do not match exactly because the periods and definitions differ, but they are close.

Which AI chips sell the most?

GPUs. They generated $162.1 billion in 2025, or 89% of the market, and GMI projects they will still be 69.01% of revenue in 2035. Merchant custom AI chips (ASICs and XPUs) were 9.07% of revenue in 2025 and are projected to reach $231.2 billion by 2035. NPUs and edge inference chips made $2.5 billion, and FPGAs used for AI were 0.27% of the market.

Which country buys the most AI hardware?

The United States, at $111.1 billion in 2025, or 61% of the global market. China was second at $21.8 billion, followed by Japan at about $9.1 billion (derived from GMI's share of Asia Pacific). North America as a region was $114.7 billion and Asia Pacific $47.3 billion. GMI projects Asia Pacific to grow fastest, at 15.3% a year, and the U.S. to reach $515.6 billion by 2035.

Cloud data centers bought $132.9 billion of AI hardware in 2025, 73% of the total. On-premises and enterprise deployments made up 25.00%, and edge infrastructure $3.6 billion. By industry, technology and cloud services companies accounted for $123.8 billion.

Methodology & caveats

All values are nominal U.S. dollars as published. Market figures come from the free summary of GMI report GMI14378 (published November 2026, page updated November 11, 2026), read on November 12, 2026. GMI publishes some segment values in dollars and others as percentage shares. Where a card shows a dollar value for a share (custom ASICs, FPGAs, on-premises deployments, Japan, the United Kingdom), it is computed from GMI's $182.0 billion total and marked as derived. GMI's 2026 and 2035 values are forecasts, and GMI does not publish its full model.

GMI's market is defined at the first external sale of accelerator systems, so it is not directly comparable with SIA's chip sales, which come from WSTS and are measured at the chipmaker level. The two overlap: the memory inside an accelerator counts in both. Figure 1 holds SIA's 2025 values fixed to give a sense of scale, not to predict when one market will pass another, and the lines between GMI's 2026 and 2035 values are straight segments, not yearly forecasts.

Company revenue comes from SEC filings and is reported on each company's own fiscal year and definitions, so it is not a like-for-like ranking. NVIDIA's data center figure includes networking, AMD's includes server CPUs, and Broadcom's AI figure, stated on its earnings call rather than in a filing, includes networking chips. Values in stat tiles and chart labels are rounded for display; tables and CSV downloads carry the published figures.