Data Center Accelerator Market Size, Share & Forecast to 2035

The Data Center Accelerator Market covers the graphics processing units (GPUs), custom AI application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs) and other AI accelerator chips sold for installation in data center servers, valued at first sale as packaged modules, boards or cards with their on-package memory. DC Market Insights values the global Data Center Accelerator Market at USD 169.77 billion in 2025 and forecasts USD 605.42 billion by 2035, a CAGR of 13.56%. The 2025 value is built from units and prices: about 6.30 million data center GPUs at an average of USD 22,060 give USD 138.98 billion, about 5.18 million custom AI ASICs at USD 5,120 give USD 26.52 billion, and FPGAs and other accelerators add USD 4.27 billion. Custom AI ASICs are the fastest-growing segment at 19.89% a year because the largest cloud providers and AI labs now design their own chips for inference, and the Middle East and Africa is the fastest-growing region at 16.45% a year as Gulf AI programs install accelerators from a small base (DC Market Insights estimate).

Executive Summary

The Global Data Center Accelerator Market size was valued at USD 9.61 billion in 2020, reached USD 169.77 billion in 2025, and is anticipated to reach USD 605.42 billion by 2035, at a CAGR of 13.56% during the forecast period.

Report attribute	Details
Historical Period	2020-2024
Base Year	2025
Forecast Period	2026-2035
Data Center Accelerator Market Size 2025	USD 169.77 Billion
Data Center Accelerator Market, CAGR	13.56%
Data Center Accelerator Market Size 2035	USD 605.42 Billion

The market grew at 77.58% a year between 2020 and 2025, and 2025 alone added 57.1% over 2024 as Blackwell-generation systems ramped. NVIDIA, the largest supplier in the DC Market Insights estimate, reported data center revenue of USD 193.70 billion for its fiscal year ended January 25, 2026, up 68% (Source: NVIDIA filing with the US Securities and Exchange Commission). Growth slows to 13.56% a year from 2025 to 2035 because unit prices stop rising as fast as they did in the Hopper and Blackwell generations, custom ASICs take share at lower prices per chip, and the market grows from a base that is already 17.7 times its 2020 size.

Market Scope: What Does the Data Center Accelerator Market Cover?

The Data Center Accelerator Market covers every compute accelerator sold for use in a data center server, valued at first sale in current US dollars, with 2025 as the base year and forecasts to 2035. DC Market Insights sizes it at USD 169.77 billion in 2025.

Included. Merchant data center GPUs sold as SXM or OCP Accelerator Module (OAM) modules, baseboards and PCIe cards; the accelerator content of rack-scale systems; custom AI ASICs designed by cloud providers and AI labs and built with design partners, valued at the price paid to the design partner or foundry chain; data center FPGAs used for AI, networking offload and financial workloads; and other accelerators such as neural processing units, language processing units and wafer-scale engines. High-bandwidth memory (HBM) is included when it sits on the accelerator package.

Excluded. Server central processing units (CPUs), data processing units and network interface cards, Ethernet and InfiniBand switches, NVLink and other scale-up switch trays, server chassis, power and cooling hardware, software licenses, cloud rental revenue, and accelerators built into client PCs, phones and vehicles.

How the segments fit together. The accelerator type, workload, form factor, memory and end-user splits each re-cut the same USD 169.77 billion total for 2025. Regions and countries are allocated by where the accelerators are installed, not where they are designed or made.

Market Drivers: What Drives the Data Center Accelerator Market?

Three drivers explain most of the growth in the Data Center Accelerator Market, which DC Market Insights forecasts at 13.56% a year to 2035. They are AI capital spending that BloombergNEF expects to rise from a little less than USD 450 billion to close to USD 750 billion among the 14 largest public data center operators, multi-gigawatt accelerator supply agreements signed by AI labs, and a shift of spending from training to inference.

Operator Capex Near USD 750 Billion Lifts Accelerator Demand

BloombergNEF expects the capital expenditure of the 14 largest publicly owned data center operators to reach close to USD 750 billion, up from a little less than USD 450 billion a year earlier, and counted more than 23 GW of data center capacity under construction worldwide, with 15.9 GW in the United States. DC Market Insights estimate: accelerator spending equals about 38% of that USD 450 billion base. NVIDIA’s data center revenue reached USD 62.30 billion in the fourth quarter of its fiscal 2026, up 75% from a year earlier and up 22% from the previous quarter, which shows demand still accelerating into 2026.

Multi-Gigawatt Supply Agreements Lock In Volumes Through 2029

AI labs now buy accelerators by the gigawatt. OpenAI signed a letter of intent with NVIDIA for at least 10 GW of NVIDIA systems, with the first phase on the Vera Rubin platform in the second half of 2026; agreed to deploy 6 GW of Advanced Micro Devices (AMD) GPUs, starting with 1 GW of Instinct MI450 in the second half of 2026; and agreed with Broadcom to deploy 10 GW of OpenAI-designed custom accelerators between the second half of 2026 and the end of 2029. Anthropic agreed to use up to one million Google Cloud tensor processing units (TPUs), with well over a gigawatt online in 2026. DC Market Insights estimates the accelerator content of one gigawatt of AI capacity at USD 17 billion to USD 22 billion at 2025 prices, so these four programs alone represent several years of 2025 market value.

Inference Overtakes Training by 2028

AI inference accounted for 41.3% of 2025 accelerator spending (USD 70.11 billion), against 51.2% (USD 86.92 billion) for training. DC Market Insights forecasts inference to grow at 18.00% a year to USD 366.88 billion, 60.6% of the 2035 market, because every deployed model serves tokens for years after it is trained. NVIDIA states that its Rubin platform will deliver up to a 10x reduction in inference token cost compared with Blackwell, which lowers the price of a token and widens the set of workloads worth running on accelerators.

Market Trends: How Is Data Center Acceleration Changing?

The Data Center Accelerator Market is moving to rack-scale systems, custom silicon and HBM4, with rack-scale systems rising from 28.7% of value in 2025 to 61.8% by 2035.

Rack-Scale Systems Reach 61.8% of Value by 2035

Accelerators are increasingly sold as complete racks in which dozens of chips share one scale-up fabric. Grace Blackwell systems accounted for roughly two-thirds of NVIDIA’s data center revenue in the fourth quarter of fiscal 2026, and the Trainium3 UltraServer from Amazon Web Services (AWS) places up to 144 Trainium3 chips in a single system. DC Market Insights values the accelerator content of rack-scale systems at 28.7% of 2025 value (USD 48.72 billion), growing at 22.61% a year, the fastest form factor. Rack-scale systems are increasingly liquid-cooled, which ties this trend to the Data Center Liquid Cooling Market.

Custom Silicon Rises from 15.6% to 26.9% of Spending

Google, AWS and Microsoft all now deploy their own accelerators. Google’s Ironwood TPU scales to 9,216 chips in one pod, with 4.6 petaFLOPS of FP8 performance per chip, four times the prior Trillium generation. Trainium3, the first 3 nm AI chip from AWS, delivers up to 4.4 times the compute of Trainium2 UltraServers. Microsoft’s Maia 200 inference accelerator is built on TSMC’s 3 nm process with more than 140 billion transistors and 216 gigabytes (GB) of HBM3e. Broadcom guided its AI semiconductor revenue to double year over year to USD 8.20 billion in the first quarter of its fiscal 2026, driven by custom AI accelerators and Ethernet AI switches.

HBM4 Raises Memory Bandwidth to 2 Terabytes per Second per Stack

Memory bandwidth limits many inference workloads. The HBM4 standard from the Joint Electron Device Engineering Council (JEDEC) allows transfer speeds up to 8 gigabits per second (Gb/s) across a 2048-bit interface, for up to 2 terabytes per second (TB/s) per stack. HBM-based accelerators hold 89.8% of 2025 value, but non-HBM designs that use LPDDR, GDDR or on-chip SRAM grow faster, at 19.30% a year, as inference cards trade bandwidth for memory capacity and cost.

Market Challenges: What Holds the Data Center Accelerator Market Back?

Export controls and power availability are the two constraints that DC Market Insights expects to cap the Data Center Accelerator Market below its unconstrained demand in 2026 and 2027.

Export Controls Cost AMD About USD 440 Million

US export rules decide which accelerators can be sold into China, the second-largest country market in the DC Market Insights estimate. AMD’s 2025 results included about USD 440 million in net inventory and related charges because of US export controls on its Instinct MI308 products. In May 2025, the US Department of Commerce rescinded the AI Diffusion Rule, which had been issued in January 2025, and issued guidance warning industry of the risks of using Chinese advanced computing chips, including specific Huawei Ascend chips (Source: Bureau of Industry and Security announcement). DC Market Insights forecasts China at 12.62% a year, below the global rate, because domestic accelerators replace imported ones at lower prices.

Power Limits Decide Where Accelerators Can Be Installed

Qualcomm’s rack-scale inference systems draw 160 kW per rack, and the gigawatt-scale agreements above need new grid connections at campus scale. DC Market Insights estimates that each gigawatt of AI capacity absorbs about 1.0 million to 1.3 million accelerators at Blackwell-generation power levels, so delays in grid connections translate directly into delayed accelerator shipments. The low scenario below models a two-year slip in AI power delivery.

Market Opportunities: Where Are the New Revenue Pools?

The two largest new revenue pools in the Data Center Accelerator Market are inference-optimized accelerators, growing at 18.00% a year, and sovereign and enterprise AI buyers outside the hyperscale clouds.

Inference Accelerators Open the Market to New Suppliers

Inference rewards memory capacity and cost per token more than peak training throughput, which gives new suppliers an entry point. Qualcomm plans AI200 rack-scale inference systems for 2026 and AI250 for 2027, with 768 GB of LPDDR per card in the AI200 rack, and Saudi AI company Humain plans to deploy 200 MW of them in Saudi Arabia and other locations. Intel expects customer sampling of its Crescent Island data center GPU, with 160 GB of LPDDR5X, in the second half of 2026. NVIDIA signed a non-exclusive license for Groq’s inference technology and hired its founder Jonathan Ross. DC Market Insights values other accelerators, the segment these new suppliers sell into, at 1.3% of 2025 value (USD 2.26 billion), growing at 18.29% a year to USD 12.11 billion.

Enterprises and Sovereign AI Buyers Add USD 115.09 Billion of Annual Spending

Enterprises accounted for 12.9% of 2025 accelerator spending (USD 21.90 billion) and grow fastest among buyer groups at 15.13% a year, as private AI clusters move into colocation halls. Government, sovereign AI and research buyers grow from USD 18.00 billion to USD 65.39 billion at 13.77% a year, and the two groups together add USD 115.09 billion of annual spending between 2025 and 2035. Suppliers that sell complete, liquid-cooled racks with software support win these buyers, who lack hyperscale engineering teams.

Pricing: How Much Does a Data Center GPU or AI ASIC Cost?

A data center GPU sold for about USD 22,060 on average in 2025 (DC Market Insights estimate). A custom AI ASIC sold for about USD 5,120, and both figures are blended average selling prices across all models, including on-package memory and excluding the server around it.

Average selling price (USD thousand)	2020	2025	2030	2035
GPUs	7.80	22.06	31.40	34.90
Custom AI ASICs	2.90	5.12	8.30	10.60
FPGAs	1.60	2.21	2.50	2.70
Other Accelerators	3.10	4.96	6.20	7.10

GPU prices nearly tripled between 2020 and 2025 as each generation added HBM capacity, larger dies and two-die packages. DC Market Insights forecasts GPU average prices to rise more slowly, by about 4.7% a year to 2035, because rack-scale buyers negotiate volume pricing and custom ASICs cap what merchant suppliers can charge. Custom ASIC prices rise faster, at about 7.5% a year, as cloud providers move their own designs to 3 nm and 2 nm processes and add HBM4.

Standards and Regulation: Which Rules Shape Data Center AI Chips?

Three sets of rules shape the Data Center Accelerator Market: the JEDEC HBM4 memory standard at up to 2 TB/s per stack, the UALink scale-up specification for up to 1,024 accelerators, and US export controls.

JEDEC HBM4 (JESD270-4). The JEDEC JESD270-4 HBM4 standard supports 4-high, 8-high, 12-high and 16-high DRAM stacks and up to 2 TB/s per stack, and DC Market Insights expects it to become the main memory interface for new accelerators from 2026.

UALink 200G 1.0. The first specification of the Ultra Accelerator Link Consortium enables 200G-per-lane scale-up connections for up to 1,024 accelerators in one AI computing pod.

US export controls. US export controls, administered by the Bureau of Industry and Security, already led to about USD 440 million of AMD charges on its MI308 products, and the rescission of the AI Diffusion Rule, covered in the challenges above, cancelled its pending compliance requirements.

Market Segmentation: Which Segment Leads Data Center AI Chip Spending?

GPUs lead the Data Center Accelerator Market with 81.9% of 2025 value, hyperscale cloud providers lead buyers with 61.7%, and AI training leads workloads with 51.2%. Custom AI ASICs are the fastest-growing segment at 19.89% a year, against 11.86% for GPUs, because cloud providers and AI labs use their own chips to cut the cost of inference at scale.

By Accelerator Type: GPUs Lead with 81.9%

Accelerator type	2025 (USD billion)	Share 2025	2035 (USD billion)	CAGR 2025-2035
GPUs	138.98	81.9%	426.27	11.86%
Custom AI ASICs	26.52	15.6%	162.68	19.89%
FPGAs	2.01	1.2%	4.36	8.04%
Other Accelerators	2.26	1.3%	12.11	18.29%

GPUs lose 11.5 points of share by 2035 but still grow more than threefold in value, because the software ecosystem around GPUs keeps them the default for training and for buyers without in-house chip teams. FPGAs grow slowest, at 8.04% a year, as AI workloads move to purpose-built chips.

By Workload: AI Training Leads with 51.2%

Workload	2025 (USD billion)	Share 2025	2035 (USD billion)	CAGR 2025-2035
AI Training	86.92	51.2%	204.63	8.94%
AI Inference	70.11	41.3%	366.88	18.00%
HPC and Scientific Computing	7.81	4.6%	18.77	9.16%
Data Analytics, Video and Other	4.92	2.9%	15.14	11.89%

AI inference overtakes training by 2028 in the DC Market Insights model. HPC and scientific computing keeps a stable but shrinking share as national laboratories buy the same accelerators as AI clouds.

By Form Factor: Multi-Accelerator Baseboards Lead with 54.2%

Form factor	2025 (USD billion)	Share 2025	2035 (USD billion)	CAGR 2025-2035
Rack-Scale Systems	48.72	28.7%	374.15	22.61%
Multi-Accelerator Baseboards (SXM and OAM)	92.01	54.2%	171.94	6.45%
PCIe Cards	29.03	17.1%	59.33	7.41%

Multi-accelerator baseboards remained the largest form factor in 2025, but rack-scale systems grow at 22.61% a year and become the largest form factor by 2030. PCIe cards serve enterprise inference servers and grow at 7.41% a year.

By Memory: HBM-Based Accelerators Lead with 89.8%

Memory	2025 (USD billion)	Share 2025	2035 (USD billion)	CAGR 2025-2035
HBM-Based	152.45	89.8%	504.31	12.71%
Non-HBM (GDDR, LPDDR and SRAM)	17.32	10.2%	101.10	19.30%

By End-user: Hyperscale Cloud Providers Lead with 61.7%

End-user	2025 (USD billion)	Share 2025	2035 (USD billion)	CAGR 2025-2035
Hyperscale Cloud Providers	104.75	61.7%	352.96	12.92%
GPU Cloud and Neocloud Providers	25.13	14.8%	97.47	14.52%
Enterprises	21.90	12.9%	89.60	15.13%
Government, Sovereign AI and Research	18.00	10.6%	65.39	13.77%

Hyperscale cloud providers buy both merchant GPUs and nearly all custom ASICs. GPU cloud and neocloud providers buy almost only GPUs and grow at 14.52% a year as AI labs rent capacity before building their own.

DC Exclusive: The Accelerator Type × Buyer Matrix

GPUs bought by hyperscale cloud providers are the single largest cell of the matrix at USD 77.44 billion, 45.6% of all 2025 accelerator spending, while hyperscalers also account for 93.7% of all custom ASIC value. The matrix below cross-tabulates what is bought and who buys it.

2025, USD billion	GPUs	Custom AI ASICs	FPGAs	Other Accelerators	Total
Hyperscale Cloud Providers	77.44	24.85	1.26	1.20	104.75
GPU Cloud and Neocloud Providers	24.50	0.22	0.04	0.37	25.13
Enterprises	20.62	0.44	0.49	0.35	21.90
Government, Sovereign AI and Research	16.42	1.01	0.22	0.34	18.00
Total	138.98	26.52	2.01	2.26	169.77

Hyperscale cloud providers spend 23.7% of their accelerator budget on custom ASICs, while GPU cloud and neocloud providers spend 97.5% on GPUs; a supplier without a custom silicon business is therefore shut out of nearly a quarter of the largest buyer group. Enterprises carry the highest FPGA share of any buyer, through network offload and trading systems. The matrix is a DC Market Insights model, not a shipment count, built from 36 inputs: 16 annual market totals (2020 to 2035), 4 accelerator unit counts and 4 average prices for 2025, 4 buyer totals, 5 regional splits and 3 supplier disclosures used as a cross-check. The model and its assumptions are explained on our research methodology page.

Regional Insights: Which Region Leads the Data Center Accelerator Market?

North America leads the Data Center Accelerator Market with 59.4% of 2025 value. The Middle East and Africa is the fastest-growing region at 16.45% a year, because Gulf AI programs are installing accelerators from a small base (DC Market Insights estimate).

North America Leads with 59.4% Share

North American data centers installed accelerators worth USD 100.84 billion, 59.4% of the 2025 total, and the region is forecast to grow at 12.71% a year to USD 333.71 billion. The United States hosted 15.9 GW of the more than 23 GW of data center capacity under construction worldwide, according to BloombergNEF, and NVIDIA, AMD and Broadcom are all named in the gigawatt-scale programs announced in 2025. North America loses 4.3 points of share by 2035 as other regions build their own AI capacity.

Asia Pacific Holds 24.6% Share

Asia Pacific is worth USD 41.76 billion in 2025 and USD 158.50 billion by 2035, a CAGR of 14.27%. China accounts for 45.2% of the region, increasingly supplied by domestic chips: Huawei plans its Ascend 950 in 2026, Ascend 960 in 2027 and Ascend 970 in 2028, with an Atlas 950 SuperPoD of 8,192 chips in the fourth quarter of 2026. Japan, South Korea, India and Australia add most of the growth outside China.

Europe Holds 9.2% Share

Europe is worth USD 15.62 billion in 2025 and USD 61.15 billion by 2035, a CAGR of 14.62%. Public AI programs and national supercomputing centers add demand, while grid connection queues in the largest hubs cap how fast private clusters can grow.

Middle East and Africa Holds 4.9% Share

The Middle East and Africa is worth USD 8.32 billion in 2025 and USD 38.14 billion by 2035, the fastest regional growth at 16.45% a year. Saudi Arabia and the United Arab Emirates account for most of the region in the DC Market Insights allocation, and Humain plans part of its 200 MW of Qualcomm inference systems for Saudi Arabia.

Latin America Holds 1.9% Share

Latin America is worth USD 3.23 billion in 2025 and USD 13.92 billion by 2035, a CAGR of 15.75%, led by Brazil, Mexico and Chile.

Country Analysis: Which Countries Install the Most Data Center Accelerators?

The United States is the largest country market for data center accelerators at USD 93.08 billion, 54.8% of 2025 global value, followed by China at USD 18.88 billion (11.1%) and Canada at USD 7.76 billion (4.6%). The 12 largest countries account for 88.3% of 2025 spending.

Rank	Country	Region	2025 (USD billion)	Share 2025	2035 (USD billion)	CAGR 2025-2035
1	United States	North America	93.08	54.8%	303.34	12.54%
2	China	Asia Pacific	18.88	11.1%	61.97	12.62%
3	Canada	North America	7.76	4.6%	30.37	14.61%
4	Japan	Asia Pacific	5.89	3.5%	20.13	13.08%
5	South Korea	Asia Pacific	3.93	2.3%	15.53	14.74%
6	United Kingdom	Europe	3.86	2.3%	13.51	13.36%
7	Saudi Arabia	Middle East and Africa	3.09	1.8%	15.29	17.33%
8	Germany	Europe	3.01	1.8%	11.07	13.89%
9	India	Asia Pacific	3.01	1.8%	19.18	20.36%
10	United Arab Emirates	Middle East and Africa	2.65	1.6%	12.82	17.09%
11	Australia	Asia Pacific	2.55	1.5%	9.99	14.64%
12	France	Europe	2.25	1.3%	9.29	15.24%

India is the fastest-growing large country market at 20.36% a year, as hyperscale operators and domestic cloud providers build AI capacity from a small installed base, ahead of Saudi Arabia at 17.33%. The country figures are a DC Market Insights allocation of the regional totals, weighted by each country’s estimated share of installed and new AI data center capacity.

Forecast Scenarios: How Could the 2035 Forecast Change?

The Data Center Accelerator Market reaches USD 605.42 billion by 2035 in the base case, with a range of USD 470.39 billion to USD 736.91 billion across the low and high scenarios.

Scenario	Assumption from 2026	2035 (USD billion)	CAGR 2025-2035
Low	Two-year slip in AI power delivery; AI capital spending flat in real terms after 2028; average prices 10% below base	470.39	10.73%
Base	Announced gigawatt programs delivered by 2030; custom ASIC share reaches 26.9%	605.42	13.56%
High	Grid and HBM supply keep pace with announced plans; inference demand 20% above base	736.91	15.81%

The low case is what a pause in AI capital spending looks like: deliveries slip, buyers sweat existing fleets for longer and suppliers cut prices to clear inventory. The high case assumes that HBM, advanced packaging and grid connections all keep pace with the gigawatt agreements signed in 2025. Supplier mix is the main swing factor inside the total: a faster move to custom ASICs lowers average prices per chip but raises unit volumes.

Competitive Insights: Who Leads the Data Center AI Chip Market?

NVIDIA leads the Data Center Accelerator Market with about 77% of 2025 value (DC Market Insights estimate), and the 14 suppliers below cover the merchant GPU, custom silicon and inference chip companies we track.

• NVIDIA
• AMD
• Broadcom
• Google (TPU)
• Amazon Web Services (Trainium)
• Microsoft (Maia)
• Intel
• Marvell Technology
• Huawei (Ascend)
• Qualcomm
• Groq
• Cerebras Systems
• Cambricon Technologies
• Alchip Technologies

The 77% share is a DC Market Insights estimate built from NVIDIA’s fiscal 2026 data center revenue of USD 193.70 billion, less more than USD 31 billion of networking revenue and an estimated share for CPUs and system components, set against the USD 169.77 billion market. AMD’s Data Center segment, whose fourth-quarter growth came from EPYC processors and the ramp of Instinct GPU shipments, reached a record USD 16.60 billion for full-year 2025, up 32%, and the OpenAI agreement for 6 GW of AMD GPUs comes with a warrant for up to 160 million AMD shares tied to deployment, share-price and other milestones. Broadcom is the largest custom AI ASIC supplier in the DC Market Insights estimate, with fourth-quarter fiscal 2025 AI semiconductor revenue up 74% year over year and a 10 GW agreement to deploy OpenAI-designed accelerators. Google, AWS and Microsoft deploy their own TPU, Trainium and Maia chips, Huawei plans new Ascend generations each year to 2028, and Qualcomm and Intel target inference workloads, as does the Groq technology licensed to NVIDIA; Cerebras Systems, Cambricon Technologies, Marvell Technology and Alchip Technologies complete the list of suppliers we track.

Recent Developments

Three AI lab supply agreements for at least 26 GW of accelerators, plus Anthropic’s deal for up to one million Google TPUs, are the largest recent developments in the Data Center Accelerator Market.

• In September 2025, NVIDIA and OpenAI signed a letter of intent to deploy at least 10 GW of NVIDIA systems, with NVIDIA intending to invest up to USD 100 billion in OpenAI progressively as each gigawatt is deployed (Source: NVIDIA newsroom announcement).
• In October 2025, AMD and OpenAI announced a multi-year agreement to deploy 6 GW of AMD GPUs, starting with 1 GW of Instinct MI450 in the second half of 2026 (Source: AMD press release).
• In October 2025, OpenAI and Broadcom agreed to deploy 10 GW of OpenAI-designed custom AI accelerators, with racks targeted to start in the second half of 2026 and complete by the end of 2029 (Source: OpenAI announcement).
• In October 2025, Anthropic agreed to expand its use of Google Cloud TPUs to up to one million chips, worth tens of billions of dollars, with well over a gigawatt of capacity online in 2026 (Source: Google Cloud press release).
• In December 2025, AWS made Trainium3 UltraServers generally available, with up to 144 Trainium3 chips per system and up to 4.4 times the compute of Trainium2 UltraServers (Source: Amazon press release).
• In January 2026, Microsoft introduced its Maia 200 inference accelerator, built on TSMC’s 3 nm process with 216 GB of HBM3e, and deployed it in its US Central region near Des Moines, Iowa (Source: Microsoft blog post).

We follow these deals, launches and capacity announcements as they happen in our data center industry updates. Accelerator clusters also set the demand for high-speed optics covered in the Optical Transceiver Market report.

Methodology: How We Built the Data Center Accelerator Market Model

DC Market Insights built this market bottom-up for 2025 from unit shipments and average selling prices for 4 accelerator types: about 6.30 million GPUs, 5.18 million custom AI ASICs, 0.91 million FPGAs and 0.46 million other accelerators, about 12.85 million units in all. The 2020 to 2024 history and the 2026 to 2035 forecast apply annual growth rates to the 2025 base, checked against announced gigawatt programs at an estimated 1.0 million to 1.3 million accelerators per gigawatt. The total was split by 4 accelerator types, 4 workloads, 3 form factors, 2 memory types, 4 buyer groups and 5 regions, and every split sums to the market total. The result was checked top-down against the disclosed data center revenue of 3 listed suppliers (NVIDIA, AMD and Broadcom) and against BloombergNEF capital spending figures for the largest operators. Country values allocate each regional total by country weights for installed and new AI capacity. The analyst named on this page built the model, and the reviewer checked the arithmetic, the sources and the dates before publication.