GPU Cloud Providers

EXPLAINED · Players · Last reviewed: 30 Sep 2026 · Next review: after next major neocloud offtake or power-capacity print

GPU cloud providers are the firms and counterparties that put AI accelerators into billed hours: who owns and operates the GPU fleet, who sells reserved or on-demand hours, who buys multi-year offtake, and who supplies the power and halls that make those hours possible.

For investors on the Neoclouds layer, the useful question is which role gets paid when offtake and power bind — not which ticker “wins.” Mechanism siblings: GPU as a Service and take-or-pay GPU contracts. Versus: neocloud vs hyperscaler.

What are GPU cloud providers?

GPU cloud providers split into four roles along one bind. Fleet owners operate the accelerators. Hour sellers clear reserved or on-demand hours. Offtake buyers commit multi-year dedicated capacity. Power and hall suppliers deliver the megawatts and shells. Missing any one role leaves CapEx waiting — even when backlog looks full.

Who owns the GPU fleet?

Fleet owners design, buy, and run GPU clusters. CoreWeave’s Form S-1, filed 3 March 2025, reported that as of 31 December 2024 its 32 active data centers were running more than 250,000 GPUs with more than 360 MW of active power, and about 1.3 GW of total contracted power. That is one labelled fleet print — company-reported scale, not a ranking. For the layer map, see what a neocloud is.

Other specialist operators sit in the same fleet role when they own or control accelerator clusters sold as cloud. Names change with each CapEx cycle; the job stays ownership and operation of the metal.

Who sells the billed hours?

Hour sellers turn fleet capacity into GPU as a Service — reserved blocks and on-demand instances. Cash clears when delivered hours are used and paid, not when a rack looks full. Many neoclouds hold both fleet-owner and hour-seller roles; keep the jobs separate when you read an IR slide. See GPU as a Service Explained.

Who buys offtake?

Offtake buyers are usually hyperscalers and AI labs. On 8 September 2025, Nebius announced a multi-billion agreement to deliver AI infrastructure to Microsoft from its Vineland, New Jersey data center (Nebius Ex. 99.1 / SEC). On 3 November 2025, Lambda announced a multibillion multi-year Microsoft agreement to deploy tens of thousands of NVIDIA GPUs, including GB300 NVL72 systems (Lambda). Those dates show the offtake-buyer role — committed capacity — beside hyperscalers who also own their own fleets.

Microsoft’s FY25 Q4 remarks (30 July 2025) put numbers on the buyer side of the same bind: capital expenditures of $24.2 billion in the quarter (including $6.5 billion of finance leases), more than two gigawatts of new capacity stood up over the prior twelve months, and Azure demand still higher than supply. For how committed capacity shows up in revenue, see take-or-pay GPU contracts and remaining performance obligation.

Who supplies power and halls?

Power and hall suppliers gate every fleet. CoreWeave’s S-1 cites an agreement with Core Scientific for more than 500 MW of capacity as of 31 December 2024 — a labelled landlord/power print, not a share table. Crusoe announced on 28 September 2026 that it is developing Google’s data center campus in Armstrong County, Texas, with construction broken ground in June 2025 — an energy-first hall developer role beside the offtake buyer.

Wire that uses GPU-cloud / neocloud offtake in anger: Crusoe building Google’s Armstrong County campus.

Dated ranges — power, fleets, CapEx

These are labelled company prints, not industry “estimates.” Use them as ranges for roles — not as a league table.

Who gets paid — and where is the stuck step?

On the Neoclouds layer, who gets paid tracks which role clears first. Fleet owners clear when CapEx becomes live clusters. Hour sellers clear when reserved or on-demand hours are delivered and paid. Offtake buyers bind cash early through multi-year commitments. Power and hall suppliers clear when megawatts and shells are contracted and energised. The stuck step is usually offtake-plus-power before billed hours — backlog and filled racks are not the same as paid hours.

Common misconceptions

• People assume “GPU cloud providers” means one tip sheet of best clouds. Actually four roles share the bind: fleet owners, hour sellers, offtake buyers, and power+hall suppliers. Ranking tickers skips the stuck step.
• People assume backlog or contracted megawatts equal cash. Actually cash clears when a customer uses a working GPU and pays for that hour. Contract prints and power runways are labelled commitments — not revenue.
• People assume hyperscalers and neoclouds are the same role. Actually hyperscalers often buy offtake and own fleets; specialist GPU cloud providers concentrate on accelerator clusters. Versus page: neocloud vs hyperscaler.
• People treat a market-share table as investment advice. Actually this page maps roles and stuck steps. Share or capacity prints from filings and press are labelled and are not buy/sell calls.

FAQ

Related terms

• Neocloud — what a specialist GPU cloud operator is.
• GPU as a Service — how billed accelerator hours clear.
• Neocloud vs hyperscaler — specialist fleets versus broad cloud platforms.
• Take-or-pay GPU contracts — offtake that binds idle hours.
• Hyperscaler — who sets AI CapEx and often buys offtake.

Markets and stack doors

Layer: Neoclouds. Buyer layer: Hyperscalers.

Further reading

Wire: Crusoe — Google Armstrong County campus. Mechanism: take-or-pay GPU contracts.

Get the Wire — short notes when offtake, power, or fleet clocks move.

Sources