How to read the price board

The price board is Compute Terminal's daily table of GPU rental prices. Each morning it is rebuilt from the public rate cards and marketplace listings we store, and each cell answers one question: what does an hour of this chip cost, from this kind of provider, today? This piece explains how to read a cell, the labels that tell you how much weight it can bear, and what each plan sees.

Rows and columns

The rows are chips. The headline shows the four that matter most to the market today, the H100, H200, B200 and B300, and the full board adds the A100, the GB200 and AMD's MI300X.

The columns are classes of provider, because the same chip is sold on very different terms.

• Neocloud. On-demand prices from GPU clouds whose main business is renting accelerators.
• Public rate-card average (list). The on-demand list prices of the four hyperscalers: Amazon Web Services, Microsoft Azure, Google Cloud and Oracle Cloud Infrastructure. This is a census, not a sample: there are only four, and the board reports them all.
• Marketplace and spot. Spare and interruptible capacity, including any provider's spot price.
• Further cells. All on-demand providers together; reserved prices for one-year and three-year terms; and serverless prices, billed by the second, which sit in their own cell and are never mixed into the others.

Our piece on how rental prices form explains why these layers differ.

What a cell's value is

Each provider counts once in a cell, at its lowest qualifying price. Every price is put per GPU per hour, so an eight-GPU server's hourly price is divided by eight, and prices in other currencies are converted at the European Central Bank's reference rate of the day. Variants are mapped to the chip family and recorded, and the H100 cell reads the SXM version only. A listing is left out if it is a fraction of a GPU, if its source marks it unavailable, if it has not been seen recently, or if it sits so far below the rest of its class that a reviewer must look at it first.

The value is a trimmed mean. The board cuts the same share of providers from the top and the bottom of the range, then averages the rest, so one unusually high or low quote cannot move the cell on its own. With only a few providers, nothing is cut. The hyperscaler cell is a plain average of the four list prices.

On the latest board, the H100 reads

n: the sample behind the value

Beside every value is n, the number of providers in the cell before the tails are cut. It is the first thing to read. A value from many providers is a fair picture of the market; a value from a few is a fair picture of those few. A cell keeps its providers and their range even when it has no value of its own, so the sample can still be inspected.

The labels

Every cell carries one of four labels, and the reader never sees a bare "n/a".

• Published. The cell has at least the minimum number of providers set in the methodology, or, for the hyperscaler cell, all four hyperscalers report. Only published cells feed our indices.
• Indicative. The cell has fewer providers than the minimum. It shows their trimmed mean, marked indicative, with n beside it. Where there are too few for even that, it shows the latest weekly median from the GetDeploying GPU price dataset, labelled with its source and licence.
• Census. A hyperscaler cell where fewer than all four report. It shows the average of those that do, with the count of reporting hyperscalers and each one named. For the newest chips this is common, because some hyperscalers do not publish an on-demand price for them.
• Not offered. No public price anywhere: "not offered" for a hyperscaler, "no public price" for any other class.

An indicative or census value is real information, but it is not a market median, and it should be quoted with its label.

Reading a move

A cell is not an index. Each day's value is the trimmed mean of the providers present that day, so a cell can move because prices moved or because a provider joined or left. Before reading a change as a price move, check whether n changed too. Our indices are built to remove that effect: they measure each day's change only over the providers present on both days, so a provider joining or leaving does not move them. The methodology page explains the rules.

Common misreadings

• Comparing different layers. An on-demand cell set against a contract price, a reserved rate or a quoted transacted price will always look wrong, because they price different promises.
• Comparing per-server and per-GPU prices. Rate cards often quote a whole server; the board is always per GPU.
• Treating an indicative cell as a market price. Read the label and n first.
• Reading the hyperscaler cell as what large customers pay. It is the published list price; committed customers usually pay less.

What each plan sees

The board is the same data for everyone; the plans differ in how much of it they open.

• Visitors see today's twelve on-demand headline cells, neocloud, hyperscaler list and marketplace for the H100, H200, B200 and B300, each with n and its label.
• Compute Open, free with an email address, opens the last seven days of on-demand history, the indicative cells, the hyperscaler breakdown by vendor, the latest weekly medians and the implied-contract panel to read.
• Standard sees the same board as Compute Open.
• Plus opens the reserved one-year and three-year columns, the board's full history with image export, the history of the weekly medians, each provider's rate card, and the implied-contract panel with assumptions it can change.
• Pro has all of that as CSV, with the raw rate cards for each chip.

The methodology is public in full, and the free headline values may be quoted by anyone under CC BY 4.0, with the credit "Source: Compute Terminal". Plans and prices are on the pricing page.

Sources

• Compute Terminal price board and its methodology; Compute Terminal Index methodology, version 1.0-rc (2 November 2026).
• GetDeploying GPU price dataset, CC BY 4.0, for the indicative fallback.
• Compute Terminal plans, as set on 5 November 2026; pricing.