What Is a Neocloud? Specialist GPU Clouds for AI Work

A neocloud is a specialist cloud that bills GPU and AI accelerator capacity — usually by the hour or under committed contracts — rather than offering a broad menu of cloud services.

The money link: a neocloud gets paid when a customer uses a working GPU and pays for that hour (or when committed capacity delivers billable hours). Backlog is not cash until delivery and use. Utilisation and realised price are the two dials.

EXPLAINEDLast reviewed: 02 Oct 2026Next review: after next CoreWeave (or major neocloud) earnings

Education only. Not investment advice. Do your own research.

What is a neocloud?

A neocloud concentrates on accelerators. Customers take GPU time by the hour, by the instance, or under reserved contracts.

The core loop is six steps:

1. Buy AI servers with GPUs (or finance them)
2. Place them in a hall with power, cooling, and network
3. Offer instances to customers
4. Bill for time, reservations, or committed use
5. Pay for hardware depreciation, power, space, staff, and software
6. Keep the gap if utilisation and pricing cover costs

That gap is the earn.

The stack layer for this business is neoclouds. Upstream clocks sit in HBM explained and CoWoS explained; contract shape sits in take-or-pay GPU contracts and remaining performance obligation.

What are the key differences between a hyperscaler and a neocloud?

A hyperscaler sells many services with GPUs as one line. A neocloud sells accelerator time and little else — that drives concentration, financing shape, and dip sensitivity.

Hyperscalers set the alternative supply. When they expand AI instances, neocloud pricing feels pressure. When their queues are long, neoclouds absorb overflow. Customers frequently use both. See also hyperscaler explained and the hyperscalers stack layer.

What are the main neocloud company groups?

Three groups earn differently. Ask which group, not which name — education only, not a buy list.

Examples illustrate group shape — not a preference list. Education only.

Dated scale markers (company-reported, as labelled):

• CoreWeave. FY2025 revenue US$5.131B; backlog US$66.8B as of 31 Dec 2025 (26 Feb 2026 results). Q2 2026 revenue US$2.575B; backlog about US$104B as of 30 Jun 2026, excluding more than US$25B early-Q3 commitments (11 Aug 2026). No newer quarterly print as of 2 Oct 2026. Backlog is company-defined under committed contracts — not cash.
• Nebius. Group revenue US$582.3M in Q2 2026; AI cloud US$574.9M (12 Aug 2026 PR). On 1 Oct 2026 Nebius said it acquired Inferize to cut idle GPU tax on cold starts — a utilisation bind, not a new revenue print (EQS; terms undisclosed).
• Crusoe. Raised US$3.9B in a Series F that pushed post-money valuation to US$30.9B (TechCrunch, 17 Sep 2026) — an example of private capital-markets-funded scale, not a valuation call.

When GPUs are scarce, secured supply is a moat and developer list pricing gets squeezed. Markets door: CoreWeave neocloud contracts.

What does a GPU hour cost?

GPU-hour prices move with accelerator scarcity, commitment length, and whether the hall is power-ready — not list rate alone. The same card can carry very different effective prices depending on who is asking.

List price ≠ realised price. Large committed buyers negotiate off-page. Educational snapshot of developer-first on-demand quotes only.

Accelerator type, commitment length, discounts, and power-ready capacity move the quote. Separate list price from realised price.

How does the revenue actually come in?

• On-demand. Customers start and stop GPUs at a posted rate. Revenue falls when machines sit idle.
• Reserved and committed capacity. Customers book blocks for months. Cash is predictable once delivery starts.
• Extras and prepaid credits. Storage, networking, and support add fees; prepaid pulls cash forward but still converts into usage later.

Why is utilisation the quiet boss?

Utilisation is the share of the fleet actively earning customer bills. A GPU that sits idle still burns space cost, power, and capital cost.

High utilisation spreads fixed hardware cost across more revenue hours. Drivers:

• Demand from AI training and inference
• Mix of reserved versus on-demand use
• Ability to place small jobs into gaps
• Scarcity of competing capacity elsewhere
• Cold-start and spare-capacity drag (idle GPU tax)

Volume and rate are separate — a sold-out fleet can still be soft on realised price.

What is the idle GPU tax — and why backlog is not cash?

The idle GPU tax is the cost of capacity assigned or held ready but not yet serving paid work — cold starts, demand spikes, weight reloads. On 1 Oct 2026 Nebius said it acquired Inferize to shorten cold starts and cut that tax (EQS; terms undisclosed). Lesson: filled racks are not paid hours. See Nebius Inferize idle GPU tax Wire. Separately, scaled neoclouds report large backlog or RPO under committed contracts — visibility, not cash, until power-ready capacity is live and hours are paid. CoreWeave’s ~US$104B backlog as of 30 Jun 2026 is not cash this quarter. See remaining performance obligation and RPO vs take-or-pay.

What does it cost to run?

Per GPU-month: billed hours × effective rate, minus hardware carry, power, cooling, space, and ops. Newer accelerators can age older fleets.

How does scarcity upstream affect the earn?

When finished GPUs are scarce because of memory or packaging queues, neoclouds with secured supply keep fleets full and rates firm — while buying the next tranche gets harder. Earn on the existing fleet can look strong while expansion stays constrained.

Power-ready capacity is the other gate — without watts, inventory is not billable. The market door where most listed neocloud exposure sits is CoreWeave neocloud contracts.

What are the failure modes?

1. Idle fleet after a demand dip or a wave of new supply
2. Power or cooling limits that prevent full clock speeds
3. A generation shift that strands older cards for top workloads
4. Customer concentration if a few buyers dominate revenue
5. Idle GPU tax that keeps assigned capacity from becoming paid hours

Debt or equipment financing speeds growth but adds covenants; equity buys runway but does not remove the need for utilisation.

Common misconceptions

• People assume a neocloud is just a small hyperscaler. Actually revenue concentration on accelerators changes the entire risk shape — a demand dip hits the whole business.
• People assume list GPU-hour prices are what providers earn. Actually realised price after discounts is the number that matters; sold-out fleets can still be soft on margin.
• People assume more GPUs automatically means more profit. Actually idle cards still cost power and capital — utilisation is the quiet boss.
• People assume backlog equals cash. Actually backlog and RPO are visibility under committed contracts — cash waits on delivery, power, and paid hours.

Practical takeaways

• The earn is billed GPU time minus hardware carry, power, cooling, space, and operations
• Utilisation is the quiet boss — including idle GPU tax on cold starts and spare capacity
• List price is not realised price; backlog and RPO are not cash until delivery and paid hours
• Ask whether GPUs are idle for lack of customers, lack of power, or cold-start drag
• Judge the group — scaled, developer-first, or sovereign — before judging the name

FAQ

Related terms

• Hyperscaler — broad cloud vs accelerator-focused earn.
• GPU as a Service — billed accelerator hours when capacity turns usable.
• Take-or-pay GPU contracts — committed capacity awaiting delivery.
• Remaining performance obligation — visibility that is not yet billed hours.
• Neocloud vs hyperscaler — versus page for the same boundary.
• GPU cloud providers — roles along the fleet, hours, offtake, and power bind.
• HBM — memory that gates finished AI GPUs.
• CoWoS — packaging that gates GPU supply.

Where this sits in the stack

Layer home: neoclouds, next to hyperscalers.

Further reading

Markets: CoreWeave neocloud contracts. Wire: Nebius Inferize idle GPU tax.

Sources

Get the Wire — short notes when the neocloud clock moves.