On-demand, reserved and spot: how GPU rental prices form

There is no single price for a GPU-hour. The same chip, in the same week, is offered at several prices at once, and each one is the price of a different promise: how soon the capacity is available, how long it is committed for, whether it can be taken back, and who carries the risk of the hours nobody uses.

Our price board shows the clearest of them every business day. For an H100, a neocloud's on-demand rate reads

$3.89 per GPU-hour

Compute Terminal price board (public rate cards and marketplaces)As of 11 Nov 2026Source

(as of 11 Nov 2026), the hyperscalers' list price

$10.04 per GPU-hour

Compute Terminal price board (public rate cards and marketplaces)As of 11 Nov 2026Source

(as of 11 Nov 2026) and the marketplace and spot price

$2.41 per GPU-hour

Compute Terminal price board (public rate cards and marketplaces)As of 11 Nov 2026Source

(as of 11 Nov 2026). For the newer B200, the neocloud rate reads

$7.85 per GPU-hour

Compute Terminal price board (public rate cards and marketplaces)As of 11 Nov 2026Source

(as of 11 Nov 2026) and the marketplace rate

$5.42 per GPU-hour

Compute Terminal price board (public rate cards and marketplaces)As of 11 Nov 2026Source

(as of 11 Nov 2026). This piece explains what each of those layers is, the layers the board shows only to subscribers, and what moves each one.

On-demand: the price of flexibility

An on-demand price is the published hourly rate for capacity that can be started and stopped at will, with no commitment, and billed by the hour or the second. The provider carries the risk that the GPU sits idle, so on-demand is usually the highest price a customer pays for a given chip from a given provider.

Neocloud on-demand prices vary widely between providers, even for the same chip, because providers differ in their networks, their support, their regions and how much capacity they have left to fill. That is why the board does not quote any one provider. For each chip it takes each provider's lowest qualifying listing, normalises it to one GPU for one hour, and publishes a trimmed mean across providers, with the number of providers beside it. Our piece on how to read the price board explains the rules.

The hyperscaler premium

The four hyperscalers on the board, Amazon Web Services, Microsoft Azure, Google Cloud and Oracle Cloud Infrastructure, list the same chips at higher on-demand prices than the neoclouds. Their price bundles a great deal that a neocloud's does not: the rest of the cloud, enterprise agreements, compliance and a global footprint. Few large customers pay those list prices for long; most commit to a term and pay a discount. The hyperscaler list price is best read as a ceiling.

For the newest chips the census is often incomplete. Some hyperscalers do not publish an on-demand price for them at all, and offer the capacity only through reservations or commitments. The board then shows how many of the four report, rather than inventing a figure for the others.

Spot and marketplaces: spare capacity

Spot capacity is capacity the provider can take back at short notice, usually when a customer paying a higher price needs it. Marketplaces let owners of GPUs list spare hours for others to rent. Both sit below on-demand prices, because the capacity comes with fewer guarantees, and the floor is set by what an owner will accept rather than leave a GPU idle. Marketplace prices also move more than list prices, because they respond to spare capacity rather than to a published rate card. The board's marketplace class collects spot and marketplace prices whoever quotes them, the hyperscalers included.

Reserved: paying for certainty

A reserved price buys a fixed term, usually one or three years, at a discount to on-demand in return for the commitment. Hyperscalers offer it as reservations, savings plans or committed-use discounts, and some neoclouds publish their own term rates. Some providers also offer short reservations of guaranteed capacity on fixed dates, which can cost more than on-demand, because they promise availability rather than a discount.

Reserved prices show how much a customer will pay for certainty, and how much a provider will give up to fill its fleet for years ahead. The board's one-year and three-year reserved columns are available to Plus and Pro subscribers.

Contracts: the price of a megawatt

The largest buyers do not rent by the hour at all. They sign multi-year take-or-pay contracts for whole clusters, priced per megawatt per year or as a total value, often with a prepayment. These contracts usually cover the newest rack-scale systems rather than the four standard chips on the board, and they rarely disclose the number of GPUs, so a price per GPU-hour usually cannot be worked out from them.

Colocation leases are a different thing again. They are priced per kilowatt-month and cover the powered, cooled building, not the GPUs, so they sit far below any GPU price for the same megawatt. Our piece on converting $/kW-month to $/GPU-hour shows how the units connect.

What moves the layers

Four forces move rental prices, and they work on the layers differently.

• Supply of the newest chips. When a generation is scarce, its hourly price holds up and supports the price of older chips with it. When supply catches up, that support fades.
• Power-limited capacity. A GPU earns only once it sits in a powered, cooled hall. Grid connections and construction now set the pace more than chip deliveries, which limits how fast new capacity can reach the market.
• Capacity taken by contracts. Providers that commit most of their capacity to multi-year contracts have less left to offer on demand, which supports on-demand prices.
• Generation changes. Each new chip generation resets the top of the market and, over time, pulls down what buyers will pay for the one before.

These forces do not always point the same way. In October 2026 some providers raised their published GPU prices, against the long-held expectation that rental prices only fall.

How to read a price you see quoted

Before comparing two prices, ask a few questions. Which layer is it: on-demand, spot, reserved or contract? Which variant of the chip, since the same name covers different memory and form factors? Is it per GPU or per server? Is it a list price, or a price someone actually paid? On what date, and across how many providers? Two figures that look alike can measure very different things, and most disagreements about what a GPU-hour costs come from comparing different layers.

What the board shows, and who sees what

Visitors see today's on-demand cells for the four chips across the three classes, each with its sample size and its label. Compute Open adds the last seven days of on-demand history, the indicative cells, the hyperscaler breakdown by vendor and the latest weekly medians. Plus adds the reserved columns, the full history and each provider's rate card. Pro adds the data as CSV.

Sources

• Compute Terminal price board, from public rate cards and marketplaces; methodology.
• Providers' public pricing pages, read for Compute Terminal's rental-price research of 2 November 2026; provider-level detail is for subscribers.
• Company filings (8-K and 6-K) for disclosed compute contracts, as recorded in our research.