What Is GPU as a Service? Billed Accelerator Hours for AI Work

EXPLAINED · Mechanism · Last reviewed: 24 Sep 2026 · Next review: after next hyperscaler AI-cloud earnings or a GPU instance / Capacity Blocks price-list update

GPU as a service is billed accelerator time: a cloud operator puts GPUs in a power-ready hall and sells hours — on-demand, reserved, or as a dated capacity block — instead of the customer buying the cards.

Who gets paid when AI teams buy hours instead of silicon, and which clock you are watching — list rate, reserved block, or megawatts that can light up. Compare the specialist seller in what a neocloud is and the demand dial in what a hyperscaler is.

What is GPU as a service?

GPU as a service (GaaS) is billed accelerator capacity. The operator owns or leases the servers, places them in a hall with power, cooling, and fabric, and bills the customer for time on those GPUs. The customer gets accelerator hours without buying the cards, the CDU, or the transformer.

The product is the hour, not the brand on the tray. A hyperscaler sells that hour as one line among storage, databases, and general compute. A neocloud sells little else. Same accelerators can appear in both stories. Who gets paid, and what a demand dip hits, are different — see neocloud vs hyperscaler.

Oracle’s Q1 FY27 print is the dated hyperscaler example. On 10 September 2026 it said Cloud Infrastructure revenue rose 121% to US$7.4 billion, that it delivered 850 MW of additional datacenter capacity and more than 300,000 GPUs since the end of Q4, booked more than US$30 billion of new AI cloud contracts, and lifted remaining performance obligations to US$664 billion (Oracle / PR Newswire, 10 Sep 2026). Capacity delivered is not hours billed. The slides add the utilisation clock: 97.9% in the quarter, and GPUs up for renewal repriced about 20% above prior contracts (Oracle Q1 FY27 slides, 10 Sep 2026 — company-reported). A fleet that stays nearly full after a record add means the short clock is still delivery. Wire: Oracle 850 MW — fleet still 97.9% full.

How do you buy a GPU hour?

Three commercial shapes matter. They are not the same product just because they share a GPU name.

On-demand. Start and stop when you need the card. The posted rate is a starting point. Large buyers negotiate off the page.

Reserved / committed. Book a block for months. You pay for access even if you do not use every hour. The operator gets a more predictable cash line; you get supply certainty. This is how most scaled fleets fill.

Capacity blocks. A dated reservation for a cluster that starts on a future day and ends on a known day. Amazon EC2 Capacity Blocks for ML let you reserve accelerated instances in UltraClusters for up to six months, in sizes from one to 64 instances (512 GPUs or 1,024 Trainium chips), up to eight weeks ahead (AWS product page, accessed 24 Sep 2026). You pay a reservation fee up front; the price is set at purchase. AWS says the current reservation rates are scheduled to update next in October 2026 (AWS pricing, accessed 24 Sep 2026).

Microsoft’s ND H100 v5 series is the hyperscaler instance shape: one VM, eight NVIDIA H100 80 GB GPUs (Microsoft Learn, accessed 24 Sep 2026). That is a SKU definition, not a price. You buy the instance hour, not the tray.

What does a GPU hour cost — and why is list not the earn?

There is no single GPU-hour price. The same card carries different effective rates by commitment, region, and whether the hall is live today.

AWS’s table is a reservation rate that updates with supply and demand; the next scheduled update is October 2026. Nebius’s October column is a posted on-demand hike — the second in about three months (Reuters, 17 Sep 2026, reported). Commitment discounts still apply; Nebius’s own list still offers up to 35% off for reserved clusters (Nebius price list, checked 23 Sep 2026).

List price ≠ realised price. A hyperscaler can discount inside a multi-year cloud deal. A neocloud can post a rise and still fill most of the book on older committed rates. The earn is billed hours × the rate after discounts, not the landing-page number.

A second unit has shown up next to the hour: revenue per megawatt. Power, not chip count, is the scarce envelope once the cards exist. Nebius’s Q2 2026 letter said its four largest new deals averaged more than US$1 billion each at US$20–25 million of annual revenue per megawatt, with a short-term opportunity of US$40–50 million per megawatt (shareholder letter, 12 Aug 2026). CoreWeave said it signed short-dated Q3 deals of about three to six months at approximately US$40 million per megawatt of annualised revenue (Business Wire, 17 Sep 2026). At US$20 million per megawatt, a 100 MW site is about US$2 billion a year; at US$40 million, about US$4 billion — our arithmetic on those company figures. The chips barely change. The price of the hour does.

Why does utilisation bind?

Utilisation is the share of the installed fleet that is actually earning a customer bill. A GPU that sits idle still burns depreciation or lease cost, hall space, and a share of power. High utilisation spreads that carry across more revenue hours. Low utilisation turns a fleet into a waiting room.

Oracle’s 97.9% utilisation after adding 850 MW is the hyperscaler version of that fact (earnings slides, 10 Sep 2026). The new capacity was absorbed, not parked. The opposite print — capacity arriving while utilisation falls — would mean hours are no longer scarce.

Upstream, CoWoS and HBM decide when the next tray can ship. Downstream, speed-to-power decides when a delivered tray becomes a billable hour. GaaS sits in the middle.

Who gets paid — and where is the stuck step?

On the hyperscalers layer, who gets paid tracks which step clears first when AI teams want hours.

Cash sticks at the operator when the fleet is full and realised prices hold — Oracle’s 97.9% utilisation and ~20% renewal premium on older GPUs is one dated print of that (company-reported, 10 Sep 2026). Cash sticks at the customer when they prepaid for a block that has not energised yet. Cash sticks upstream when CoWoS slots or HBM stacks bind and the next tray slips.

The US market door is where most public GaaS exposure sits: US speed-to-power when halls wait on watts, and CoreWeave neocloud contracts when the specialist book is the story. China is the other hyperscaler-layer door when domestic hours cannot use the same stack — China HBM access.

A useful test: if GPU contracts dominate growth, the GaaS line behaves like a neocloud even inside a hyperscaler brand. If accelerator revenue is a slice of a broad book, a soft hour is absorbed elsewhere. That is a concentration test, not a ticker call.

Common misconceptions

• People assume GPU as a service is a neocloud-only product. Actually hyperscalers sell the same hour as one line among many. The mechanism is billed time. The risk shape changes with revenue concentration — see neocloud vs hyperscaler.
• People assume the list GPU-hour is what the operator earns. Actually realised price after discounts, reserved mix, and utilisation decide the earn. A posted hike can sit next to older committed rates that still fill most of the book.
• People assume more delivered megawatts means more profit this quarter. Actually installs and billed hours run on different clocks. A hall can be contracted and still dark.
• People assume a capacity block is the same as on-demand. Actually a block is a dated window you pay for up front. AWS sets the reservation price at purchase; the public table can move later without changing a block you already bought.

Not investment advice. Do your own research.

FAQ

Related terms

• Neocloud — specialist seller of accelerator hours.
• Hyperscaler — the CapEx dial that sets how many hours can exist.
• Neocloud vs hyperscaler — same GPUs, different who gets paid.
• HBM — memory that must match the GPU before an hour can be sold.
• CoWoS — packaging attach that can hold the next tray.

Markets and stack doors

Layer: hyperscalers. Sibling layer: neoclouds. Markets: US speed-to-power · CoreWeave neocloud contracts · China HBM access.

Get the Wire — short notes when GPU hours or hall clocks move. Wire: Oracle 850 MW · Nebius prices. Analysis: where the stuck step moves.

Sources