GPU as a Service Market Size, Share & Forecast to 2035

The GPU as a Service Market covers the rental of graphics processing unit (GPU) compute capacity by the hour, by the month or under multi-year capacity contracts, sold as cloud instances, bare-metal servers or managed clusters by hyperscale clouds, specialist GPU clouds (neoclouds), sovereign and telecom clouds and GPU marketplaces. DC Market Insights values the global GPU as a Service Market at USD 38.58 billion in 2025 and forecasts USD 268.21 billion by 2035, a CAGR of 21.40%. The 2025 value is built from rented GPU-hours: an average rentable fleet of about 3.10 million H100-equivalent GPUs, running 8,760 hours at 74% billed utilization and a blended realized price of USD 1.92 per H100-equivalent GPU-hour, gives USD 38.58 billion. AI inference is the fastest-growing workload at 27.91% a year because every model that reaches production needs GPUs around the clock, and the Middle East and Africa is the fastest-growing region at 25.91% a year because sovereign AI programs there start from a small rented base.

Executive Summary

The Global GPU as a Service Market size was valued at USD 3.58 billion in 2020, reached USD 38.58 billion in 2025, and is anticipated to reach USD 268.21 billion by 2035, at a CAGR of 21.40% during the forecast period.

Report attribute	Details
Historical Period	2020-2024
Base Year	2025
Forecast Period	2026-2035
GPU as a Service Market Size 2025	USD 38.58 Billion
GPU as a Service Market, CAGR	21.40%
GPU as a Service Market Size 2035	USD 268.21 Billion

DC Market Insights estimates that GPU rental revenue grew at 60.88% a year between 2020 and 2025, and by 59.1% in 2025 alone, as large language model training moved from research clusters to rented capacity. Listed GPU cloud specialists confirm the pace: CoreWeave reported 2025 revenue of USD 5.13 billion, up from USD 1.92 billion in 2024, and a revenue backlog of USD 66.80 billion at the end of 2025. Nebius reported fourth-quarter 2025 revenue of USD 227.70 million, up 547% year over year, and annualized run-rate revenue (ARR) of USD 1.25 billion at year-end (Source: Nebius shareholder letter filed with the US Securities and Exchange Commission).

DC Market Insights forecasts growth of 65.5% in 2026, to USD 63.86 billion, as Blackwell-generation capacity contracted in 2025 goes live, then a gradual slowdown to 21.40% a year over the decade. Growth slows because the realized price of each H100-equivalent GPU-hour keeps falling as newer GPUs deliver more compute per dollar, even while the rented fleet grows about 13 times between 2025 and 2035.

Market Scope: What Does the GPU as a Service Market Cover?

The GPU as a Service Market covers every contract under which a buyer pays a provider for time on GPUs that the provider owns or leases and operates in a data center. DC Market Insights sizes it at USD 38.58 billion in 2025, measured as provider revenue in current US dollars, with 2025 as the base year and forecasts to 2035.

Included. On-demand GPU virtual machines and bare-metal instances, reserved instances and capacity blocks, multi-year committed capacity contracts (including capacity that one cloud provider rents from another), spot and interruptible GPU capacity, managed training and inference clusters billed by GPU-hour, GPU capacity sold through marketplaces and aggregators, and the storage and networking charges billed with the GPU instance.

Excluded. Sales of GPUs, servers and data center equipment, colocation of GPUs owned by the customer, AI model application programming interfaces (APIs) priced per token, software-as-a-service applications, and GPUs that a cloud company uses for its own internal products. The hardware and facilities behind GPU clouds are sized in the AI Infrastructure Market report, and the chips themselves in the Data Center GPUs Market report.

How the segments fit together. The service-model, GPU-type, provider-type, workload, end-user and regional splits each divide the full USD 38.58 billion. DC Market Insights converts all GPUs into H100-equivalents, counting one Blackwell B200 as about 2.5 H100-equivalents and one Ampere A100 as about 0.35, so that fleets of different generations can be added together.

Market Drivers: What Drives the GPU as a Service Market?

Multi-year capacity contracts worth tens of billions of US dollars, hyperscale clouds that report being supply constrained and an AI training and inference workload that few buyers can host on their own explain why DC Market Insights forecasts the GPU as a Service Market to grow 65.5% in 2026.

CoreWeave Backlog Reached USD 104.00 Billion by June 2026

Specialist GPU clouds now sell most capacity years ahead. CoreWeave’s revenue backlog rose from USD 66.80 billion at the end of 2025 to about USD 104.00 billion at 30 June 2026, its active power reached 1.5 GW and its contracted power about 3.7 GW (Source: CoreWeave second-quarter 2026 results). Second-quarter 2026 revenue was USD 2.58 billion, against USD 1.21 billion a year earlier. Meta agreed to source USD 14.20 billion of AI infrastructure from CoreWeave in a contract that runs through December 2031, and CoreWeave’s agreements with OpenAI reached a total contract value of up to about USD 22.40 billion in September 2025. DC Market Insights estimates that committed and reserved capacity made up 63.8% of GPU rental revenue in 2025.

Hyperscale Clouds Report Capacity Shortages

Demand for rented GPUs runs ahead of supply at the largest clouds. Google Cloud revenue grew 82% to USD 24.80 billion in the second quarter of 2026, its backlog reached USD 514.00 billion, and Alphabet said it continues to be supply constrained. Amazon Web Services (AWS) grew 36.7% in the same quarter, its fastest growth in 18 quarters, and Microsoft’s Azure revenue surpassed USD 100.00 billion for the first time in fiscal 2026. Oracle Cloud Infrastructure (OCI) revenue rose 121% to USD 7.40 billion in the first quarter of fiscal 2027, with remaining performance obligations of USD 664.00 billion. DC Market Insights values GPU rental by hyperscale clouds at USD 24.07 billion in 2025, 62.4% of the market.

Hyperscalers Rent Capacity from Neoclouds

Hyperscale clouds now buy GPU capacity from specialists to meet their own customers’ demand. Microsoft signed a five-year agreement with Nebius worth USD 17.40 billion, which could rise to USD 19.40 billion, starting at a Vineland, New Jersey site with 300 MW of initial capacity. Microsoft also signed a five-year contract worth about USD 9.70 billion with IREN for NVIDIA GB300 GPUs, an expanded deal with Nscale for about 200,000 NVIDIA GB300 GPUs and a multibillion-dollar agreement with Lambda. DC Market Insights estimates that this hyperscaler and big tech offtake was worth USD 7.18 billion in 2025, 18.6% of the market.

Market Trends: How Is GPU as a Service Changing?

The GPU as a Service Market is moving from Hopper to Blackwell GPUs, from training to inference and from single-provider contracts to marketplaces that pool capacity across many clouds.

Blackwell Reaches 86.0% of Rental Revenue by 2035

DC Market Insights estimates that NVIDIA Hopper GPUs (H100 and H200) earned 58.3% of GPU rental revenue in 2025 and Blackwell-generation GPUs 21.4%. The largest contracts announced since late 2025 specify Blackwell systems: the Microsoft agreements with IREN and Nscale both specify NVIDIA GB300 GPUs, and Lambda’s Microsoft agreement includes GB300 NVL72 systems. DC Market Insights forecasts Blackwell and later generations to grow at 39.51% a year and reach 86.0% of revenue by 2035.

Inference Overtakes Training

A frontier model training run rents a large cluster for weeks or months, while serving the model rents GPUs for as long as the product is live. DC Market Insights estimates that AI model training earned 54.2% of GPU rental revenue in 2025 and AI inference 33.1%. Inference grows at 27.91% a year and reaches 55.8% of the market by 2035, against 33.5% for training, in the DC Market Insights forecast.

Marketplaces Pool GPUs Across Clouds

NVIDIA announced DGX Cloud Lepton in May 2025, a compute marketplace that connects developers with tens of thousands of GPUs from a global network of cloud providers, with partners including CoreWeave, Crusoe, Firmus, Foxconn, GMI Cloud, Lambda, Nebius, Nscale, SoftBank Corp. and Yotta Data Services. DC Market Insights values GPU marketplaces and aggregators at USD 1.35 billion in 2025 and forecasts USD 9.12 billion by 2035, with startups and developers taking 48.9% of marketplace revenue in 2025.

Market Challenges: What Holds the GPU as a Service Market Back?

Falling price per GPU-hour, customer concentration and the power needed for each new cluster are the three constraints that DC Market Insights expects to shape the GPU as a Service Market through 2030.

AWS Cut H100 Instance Prices by up to 45%

AWS cut on-demand prices for its P5 (H100) instances by up to 45%, its P5en instances by up to 26% and its P4d and P4de (A100) instances by up to 33% in June 2025. DC Market Insights models the blended realized price per H100-equivalent GPU-hour falling from USD 2.25 in 2023 to USD 1.92 in 2025 and USD 1.05 in 2035. Each new GPU generation resets the price of older capacity.

A Few Buyers Hold Most Committed Capacity

AI labs and model developers bought 41.5% of rented GPU capacity in 2025 by DC Market Insights estimates, and hyperscaler and big tech offtake another 18.6%. For neocloud specialists, offtake by hyperscalers and big tech made up 52.2% of 2025 revenue in the DC Market Insights matrix. A shift in one hyperscaler’s build-versus-rent decision moves the revenue of a whole provider group.

Power Limits How Fast Capacity Comes Online

The International Energy Agency (IEA) estimates that data centers used about 415 TWh of electricity in 2024, around 1.5% of world consumption, and projects about 945 TWh by 2030. BloombergNEF forecasts US data center power demand rising from almost 35 GW in 2024 to 78 GW by 2035. DC Market Insights ties its fleet forecast to power that providers have already contracted.

Market Opportunities: Where Are the New Revenue Pools?

DC Market Insights identifies the two largest new revenue pools in the GPU as a Service Market as enterprises, growing at 26.67% a year, and sovereign and telecom GPU clouds, growing at 26.88% a year.

Enterprises Reach USD 87.44 Billion by 2035

DC Market Insights estimates that enterprises spent USD 8.22 billion on rented GPU capacity in 2025, 21.3% of the market, and forecasts USD 87.44 billion by 2035, when they nearly match AI labs. In the DC Market Insights view, enterprises rent rather than build because inference workloads are uneven and GPU generations turn over quickly. Hyperscale clouds took 86.1% of enterprise GPU spending in 2025 in the DC Market Insights matrix.

Sovereign and Telecom GPU Clouds Grow Tenfold

Governments want domestic AI capacity under local law. The European Union’s InvestAI initiative aims to mobilize EUR 200.00 billion for AI, including a EUR 20.00 billion fund for AI gigafactories. DC Market Insights values sovereign and telecom GPU clouds at USD 2.43 billion in 2025 and USD 26.28 billion by 2035, with government and research buyers taking 26.7% of their 2025 revenue.

Pricing: How Much Does a GPU-Hour Cost?

An on-demand NVIDIA H100 GPU lists at USD 3.99 to about USD 6.16 per GPU-hour at Lambda and CoreWeave, and an AWS H100 capacity block at USD 5.19 per GPU-hour, while DC Market Insights estimates the blended realized price across all contracts at USD 1.92 per H100-equivalent GPU-hour in 2025. The gap reflects multi-year contracts, which are priced well below list, and older GPUs rented at lower rates.

Published price (USD per GPU-hour)	Pricing basis	NVIDIA H100	NVIDIA B200
Lambda	On-demand, per GPU	3.99	6.69
CoreWeave	On-demand, 8-GPU instance divided by 8	6.16	8.60
AWS	EC2 Capacity Blocks, US East (N. Virginia)	5.19	12.36

CoreWeave lists an 8-GPU HGX H100 instance at USD 49.24 per hour and an 8-GPU HGX B200 instance at USD 68.80 per hour; AWS lists the p5.48xlarge capacity block at USD 41.53 per hour and the p6-b200.48xlarge at USD 98.84 per hour.

DC Market Insights model	2020	2025	2030	2035
Realized price (USD per H100-equivalent GPU-hour)	2.20	1.92	1.36	1.05
Average rentable fleet (million H100-equivalents)	0.30	3.10	19.00	40.50
Billed utilization	62%	74%	73%	72%

The realized price falls about 5.9% a year after 2025 as Blackwell and later GPUs deliver more H100-equivalents of compute for each dollar of rent, while billed utilization stays near 72%.

Standards and Regulation: Which Rules Shape GPU as a Service?

Four sets of rules shape who can rent GPUs and on what terms: the European Union Data Act, the EU Artificial Intelligence Act (AI Act), the EU InvestAI program and United States export controls on advanced AI chips.

EU Data Act (since 12 September 2025). The Data Act requires that customers of cloud and edge services be able to switch from one provider to another, and it will entirely remove switching charges, including data egress charges, from 12 January 2027.

EU AI Act (general-purpose AI obligations from 2 August 2025). The obligations for general-purpose AI (GPAI) models became applicable on 2 August 2025, adding documentation duties for model developers who train on rented GPU clusters.

InvestAI and AI gigafactories (February 2025). The EU initiative aims to mobilize EUR 200.00 billion for AI investment, including a new EUR 20.00 billion European fund for AI gigafactories.

US export controls. The US Department of Commerce announced the rescission of the AI Diffusion Rule in May 2025. DC Market Insights treats export rules as a regional allocation risk rather than a global size risk, because rented capacity can be placed in the countries where GPUs can be shipped.

Market Segmentation: Which Segment Leads the GPU as a Service Market?

DC Market Insights estimates that committed and reserved capacity leads the GPU as a Service Market with 63.8% of 2025 value, NVIDIA Hopper GPUs lead GPU types with 58.3%, hyperscale cloud providers lead provider types with 62.4%, AI model training leads workloads with 54.2% and AI labs lead buyers with 41.5%. AI inference is the fastest-growing segment among workloads at 27.91% a year, because inference demand rises with every user and application rather than with each new model.

By Service Model: Committed and Reserved Capacity Leads with 63.8%

Service model	2025 (USD billion)	Share 2025	2035 (USD billion)	CAGR 2025-2035
Committed and Reserved Capacity	24.61	63.8%	156.10	20.29%
On-Demand Instances	9.49	24.6%	73.76	22.76%
Spot and Marketplace Capacity	4.48	11.6%	38.35	23.95%

Committed capacity loses 5.6 points of share by 2035 as supply catches up with demand and inference buyers prefer flexible terms, but it remains the largest service model in 2035 at 58.2%.

By GPU Type: NVIDIA Hopper Leads with 58.3%

GPU type	2025 (USD billion)	Share 2025	2035 (USD billion)	CAGR 2025-2035
NVIDIA Hopper (H100, H200)	22.49	58.3%	16.09	-3.29%
NVIDIA Blackwell and Later	8.26	21.4%	230.66	39.51%
NVIDIA Ampere and Older	5.63	14.6%	1.61	-11.77%
AMD Instinct and Other GPUs	2.20	5.7%	19.85	24.60%

Hopper revenue declines to USD 16.09 billion by 2035 in the DC Market Insights forecast as its rental price falls and its fleet retires, while AMD Instinct and other GPUs gain 1.7 points of share as buyers add a second GPU supplier for inference.

By Provider Type: Hyperscale Cloud Providers Lead with 62.4%

Provider type	2025 (USD billion)	Share 2025	2035 (USD billion)	CAGR 2025-2035
Hyperscale Cloud Providers	24.07	62.4%	147.79	19.90%
Neocloud GPU Specialists	10.73	27.8%	85.02	23.00%
Sovereign and Telecom GPU Clouds	2.43	6.3%	26.28	26.88%
GPU Marketplaces and Aggregators	1.35	3.5%	9.12	21.05%

CoreWeave’s 2025 revenue of USD 5.13 billion equals 47.8% of the neocloud segment in DC Market Insights estimates. Neocloud specialists gain 3.9 points of share by 2035 on the strength of contracted backlogs.

By Workload: AI Model Training Leads with 54.2%

Workload	2025 (USD billion)	Share 2025	2035 (USD billion)	CAGR 2025-2035
AI Model Training	20.92	54.2%	89.85	15.69%
AI Inference	12.77	33.1%	149.67	27.91%
Fine-Tuning and Other AI	2.85	7.4%	18.77	20.74%
HPC and Rendering	2.04	5.3%	9.92	17.14%

High-performance computing (HPC) and rendering keep growing at 17.14% a year but lose share to AI workloads.

By End-user: AI Labs and Model Developers Lead with 41.5%

End-user	2025 (USD billion)	Share 2025	2035 (USD billion)	CAGR 2025-2035
AI Labs and Model Developers	16.01	41.5%	88.50	18.65%
Enterprises	8.22	21.3%	87.44	26.67%
Hyperscaler and Big Tech Offtake	7.18	18.6%	34.33	16.94%
Startups and Developers	4.32	11.2%	33.26	22.64%
Government and Research	2.85	7.4%	24.68	24.09%

DC Market Insights forecasts hyperscaler and big tech offtake to grow slowest, at 16.94% a year, because the largest buyers bring more capacity in-house once their own data centers are built.

DC Exclusive: The GPU-Hour Model and Provider × Buyer Matrix

At USD 11.08 billion in 2025, GPU capacity that hyperscale clouds rent to AI labs and model developers forms the single largest cell of the matrix and 28.7% of all GPU as a Service revenue.

2025, USD billion	AI Labs	Enterprises	Hyperscaler Offtake	Startups	Government and Research	Total
Hyperscale Cloud Providers	11.08	7.08	1.20	2.75	1.96	24.07
Neocloud GPU Specialists	3.87	0.42	5.60	0.66	0.18	10.73
Sovereign and Telecom GPU Clouds	0.64	0.61	0.28	0.25	0.65	2.43
GPU Marketplaces and Aggregators	0.42	0.11	0.10	0.66	0.06	1.35
Total	16.01	8.22	7.18	4.32	2.85	38.58

Neocloud GPU specialists earn 52.2% of their revenue from hyperscaler and big tech offtake, and supply 78.0% of all offtake capacity, so their growth depends on how long the largest clouds keep renting rather than building. Enterprises buy 86.1% of their GPU capacity from hyperscale clouds, where their data already sits.

The matrix is a DC Market Insights model, not a survey of contracts. It is built from 77 inputs: 16 years of rentable fleet, billed utilization and realized price (48 inputs), 4 provider-type shares, 5 end-user shares and 20 provider-buyer affinity weights, checked against the disclosed revenue and backlog of 2 listed GPU cloud specialists. The model and its assumptions are explained on our research methodology page.

Regional Insights: Which Region Leads the GPU as a Service Market?

DC Market Insights estimates that North America leads the GPU as a Service Market with 58.6% of 2025 value, and the Middle East and Africa is the fastest-growing region at 25.91% a year. Asia Pacific gains the most share of any region, rising from 20.4% to 24.2% by 2035.

North America Leads with 58.6% Share

DC Market Insights values North American GPU rental at USD 22.61 billion in 2025 and USD 136.51 billion by 2035, a CAGR of 19.70%. The region hosts the frontier AI labs, the three largest hyperscale clouds and the largest neocloud campuses, including Nebius’s Vineland site and IREN’s Childress, Texas campus. North America loses 7.7 points of share by 2035 as other regions build domestic capacity, not because its volume falls.

Asia Pacific Holds 20.4% Share

DC Market Insights values Asia Pacific at USD 7.87 billion in 2025 and USD 64.91 billion by 2035, a CAGR of 23.49%. China is the largest market in the region at USD 2.93 billion, and India grows fastest at 29.99% a year from a small base. Japan follows at USD 1.31 billion in the DC Market Insights country allocation.

Europe Holds 14.1% Share

DC Market Insights values Europe at USD 5.44 billion in 2025 and USD 42.38 billion by 2035, a CAGR of 22.79%. The InvestAI gigafactory fund and the Data Act’s switching rules support domestic GPU clouds, and Nscale’s Microsoft deal includes GPUs in Narvik, Norway, Loughton, United Kingdom, and Sines, Portugal.

Middle East and Africa Holds 3.9% Share

DC Market Insights values the Middle East and Africa at USD 1.50 billion in 2025 and USD 15.02 billion by 2035, a CAGR of 25.91%, the fastest of any region. Saudi Arabia accounts for about a third of regional revenue in the DC Market Insights country allocation. Saudi capital is also flowing into GPU clouds: Aramco Ventures led US-based Together AI’s USD 800.00 million Series C round in July 2026.

Latin America Holds 3.0% Share

DC Market Insights values Latin America at USD 1.16 billion in 2025 and USD 9.39 billion by 2035, a CAGR of 23.26%. Brazil accounts for close to half of regional revenue, with demand mostly from enterprises and startups served by hyperscale cloud regions.

Country Analysis: Which Countries Spend the Most on GPU as a Service?

DC Market Insights estimates that the United States is the largest country market for GPU as a Service at USD 21.19 billion in 2025, 54.9% of global value, followed by China at USD 2.93 billion (7.6%) and the United Kingdom at USD 1.58 billion (4.1%). The 12 largest countries account for 86.9% of 2025 revenue.

Rank	Country	Region	2025 (USD billion)	Share 2025	2035 (USD billion)	CAGR 2025-2035
1	United States	North America	21.19	54.9%	126.27	19.54%
2	China	Asia Pacific	2.93	7.6%	20.64	21.56%
3	United Kingdom	Europe	1.58	4.1%	11.10	21.53%
4	Canada	North America	1.42	3.7%	10.24	21.84%
5	Japan	Asia Pacific	1.31	3.4%	10.26	22.85%
6	Germany	Europe	1.12	2.9%	8.39	22.31%
7	France	Europe	0.81	2.1%	6.82	23.75%
8	India	Asia Pacific	0.81	2.1%	11.16	29.99%
9	South Korea	Asia Pacific	0.73	1.9%	6.36	24.17%
10	Australia	Asia Pacific	0.61	1.6%	5.39	24.34%
11	Brazil	Latin America	0.52	1.3%	4.39	23.78%
12	Saudi Arabia	Middle East and Africa	0.48	1.2%	5.29	27.12%

India is the fastest-growing country market at 29.99% a year in DC Market Insights estimates, ahead of Saudi Arabia at 27.12%, and India passes Japan, Canada and the United Kingdom by 2035. The country figures are a DC Market Insights allocation of the regional totals, weighted by each country’s estimated share of rentable GPU capacity and of AI lab, enterprise and public-sector demand.

Forecast Scenarios: How Could the 2035 Forecast Change?

DC Market Insights forecasts the GPU as a Service Market at USD 268.21 billion by 2035 in the base case, with a range of USD 197.41 billion to USD 326.95 billion across the low and high scenarios.

Scenario	Assumption from 2026	2035 (USD billion)	CAGR 2025-2035
Low	Rentable fleet 20% below base in 2035 as more buyers build their own clusters; realized price 8% below base	197.41	17.73%
Base	Rentable fleet reaches 40.50 million H100-equivalents; realized price falls to USD 1.05 per GPU-hour	268.21	21.40%
High	Rentable fleet 15% above base as inference moves to rented capacity; realized price 6% above base	326.95	23.83%

Competitive Insights: Who Leads the GPU as a Service Market?

• Amazon Web Services (AWS)
• Microsoft Azure
• Google Cloud
• Oracle Cloud Infrastructure
• CoreWeave
• Nebius
• Lambda
• Crusoe
• Nscale
• IREN
• Together AI
• Fluidstack
• Firmus
• GMI Cloud
• Yotta Data Services
• SoftBank Corp.

Hyperscale cloud providers together hold 62.4% of the GPU as a Service Market in 2025, and CoreWeave, a listed GPU cloud specialist, holds about 13.3% (DC Market Insights estimate: CoreWeave’s 2025 revenue of USD 5.13 billion set against the USD 38.58 billion market; the hyperscale clouds do not report GPU rental revenue separately). CoreWeave’s revenue backlog reached about USD 104.00 billion in June 2026. Nebius ended 2025 with ARR of USD 1.25 billion and targets ARR of USD 7.00 billion to USD 9.00 billion and 800 MW to 1 GW of connected power by the end of 2026; it also signed a USD 3.00 billion, five-year agreement with Meta. Oracle’s infrastructure revenue rose 121% in the first quarter of fiscal 2027. Lambda raised USD 1.50 billion in Series E funding in November 2025, Crusoe raised over USD 3.00 billion at a valuation of about USD 30.00 billion in 2026, and Fluidstack raised USD 830.00 million at a USD 7.50 billion valuation and plans USD 50.00 billion of US AI infrastructure with Anthropic. Firmus, GMI Cloud, Yotta Data Services and SoftBank Corp. were named as NVIDIA DGX Cloud Lepton partners.

Recent Developments

• In September 2025, Microsoft signed a five-year GPU capacity agreement with Nebius worth USD 17.40 billion, which could rise to USD 19.40 billion (Source: Data Center Dynamics news report).
• In September 2025, CoreWeave expanded its agreement with OpenAI by up to USD 6.50 billion, bringing the total contract value to about USD 22.40 billion (Source: CoreWeave press release).
• In October 2025, Nscale signed an expanded deal with Microsoft for about 200,000 NVIDIA GB300 GPUs across Europe and the United States (Source: Nscale press release).
• In November 2025, IREN signed a five-year contract worth about USD 9.70 billion, including a 20% prepayment, to give Microsoft access to NVIDIA GB300 GPUs (Source: IREN press release).
• In August 2026, CoreWeave reported second-quarter revenue of USD 2.58 billion and a revenue backlog of about USD 104.00 billion (Source: CoreWeave earnings release).
• In September 2026, Crusoe raised over USD 3.00 billion in a funding round that values the company at about USD 30.00 billion (Source: Bloomberg Government news report).

We follow these contracts, price changes and capacity announcements as they happen in our data center industry updates.

Methodology: How We Built the GPU as a Service Market Model

DC Market Insights built this market bottom-up from rented GPU-hours. The rentable GPU fleet, converted into H100-equivalents, averages 0.30 million in 2020, 3.10 million in 2025, 19.00 million in 2030 and 40.50 million in 2035. Each year’s fleet is multiplied by 8,760 hours, by a billed utilization rate that rises from 62% in 2020 to 74% in 2025 and settles at 72%, and by a blended realized price per H100-equivalent GPU-hour of USD 2.20 in 2020, USD 1.92 in 2025 and USD 1.05 in 2035. The total was split by 3 service models, 4 GPU types, 4 provider types, 4 workloads, 5 end-user groups and 5 regions, and every split sums to its parent total. The result was checked top-down against the reported revenue of CoreWeave (USD 5.13 billion in 2025) and Nebius (USD 1.25 billion of ARR at the end of 2025), against published on-demand and capacity-block prices from Lambda, CoreWeave and AWS, and against the cloud revenue disclosures of Microsoft, Alphabet, Amazon and Oracle. Country values allocate each regional total by country weights for rentable capacity and demand. The analyst named on this page built the model, and the reviewer checked the arithmetic, the sources and the dates before publication.