GPU rental prices · explained

What a GPUprice indexmeasures.

GPU rental prices now have indices, announced futures and prediction markets of their own. Behind the word “price” sit three different numbers: what a provider posts, what a marketplace offers, and what renters actually paid. This page says which is which, who publishes what, and where the numbers on this site fit.

as of 13 October 2026

Three kinds of GPU price

List price

The hourly rate a provider posts for an instance or a GPU, before any discount.

Examples: RunPod's posted on-demand price; the AWS, Azure and Oracle Cloud price lists.
Good for: Exact and reproducible: anyone can read the same number from the same page.
Blind spot: Moves rarely and says nothing about discounts, reserved terms or whether capacity is free.

Cloud list prices per class →

Marketplace median

The middle of the offers a renter could take at one moment on an open marketplace, with the number of offers behind it.

Examples: Vast.ai and Clore.ai, read every hour on this site.
Good for: Moves with supply and demand within hours, and the offer count shows how thin the market is.
Blind spot: An offer is not a deal: it shows what was asked, not what was paid, and a median over a handful of offers is fragile.

Marketplace medians per class →

Transaction index

A daily rate computed from rentals that were actually executed, usually weighted by volume and cleaned of outliers.

Examples: Ornn's Compute Price Index describes itself this way; Silicon Data standardises observations from many sources into one daily rate.
Good for: Closest to what compute really changed hands for, which is why futures reference such indices.
Blind spot: Needs access to deal data, so the inputs cannot be checked from outside, and a blended rate hides which provider moved.

Who publishes what

The indices below are their publishers' products; we describe them from their own announcements and quote none of their values.

Silicon Data

Launched the Silicon Data H100 Rental Index (Bloomberg ticker SDH100RT) on 20 May 2025: a daily rate built from 3.5 million pricing data points across rental platforms.
Standardises each observation for machine specification, rental terms and location, and publishes separate neocloud and hyperscaler readings.
CME Group announced the Silicon Data H100 and B200 Rental Index Futures on 11 August 2026: listed on NYMEX, each contract a month's worth of rent for one GPU, planned for 5 October 2026 pending regulatory review.

Sources:Silicon Data press release, 20 May 2025 ↗Silicon Data H100 index page, 13 October 2026 ↗CME Group press release, 11 August 2026 ↗

Ornn

Publishes the Ornn Compute Price Index (OCPI), described by ICE as the first compute index built only from printed transactions, for H100, H200, B200, RTX 5090 and other GPUs.
Ornn's B200 page describes each daily value as a volume-weighted, winsorised average of transacted rental prices, settled once per trading day at 4:00 p.m. New York time.
Intercontinental Exchange announced cash-settled GPU compute futures on OCPI on 19 May 2026, to launch subject to regulatory approval.

Sources:ICE press release, 19 May 2026 ↗Ornn B200 market page, 13 October 2026 ↗

Compute Desk

Nodal Exchange plans GPU compute futures with Compute Desk; the contracts are to settle against Compute Desk's GPU rental price indexes, as reported on 3 September 2026.

Sources:Futures & Options World, 3 September 2026 ↗

Prediction markets

Kalshi lists yes-or-no contracts on rental prices for the A100, H100, H200, B200 and RTX 5090 that reference the Ornn Compute Price Index, as reported on 17 August 2026.
Business Insider reported on 10 October 2026 that users of Kalshi and Polymarket bet on what a B200 rental will cost at the end of October.

Sources:PokerNews on Kalshi's compute markets, 17 August 2026 ↗Business Insider Deutschland, 10 October 2026 ↗

Where the numbers on this site fit

We publish what each source shows, one source at a time: marketplace medians with their offer counts and quartiles, RunPod's posted floor and the big clouds' list prices, read hourly or four times a day, each with its timestamp.
We never blend sources into one rate and never compute a transaction index: we see offers and price lists, not deals.
The Rent Index says what each source's daily median did over 7 to 90 days. It describes; it does not forecast.
Two derived numbers sit beside the readings. The cloud premium divides the cheapest big-cloud list price by the same day's marketplace median for the same card, so a reader sees what a reserved instance cost against one marketplace GPU that day. Market depth shows how many offers stood behind each day's median, and a median over fewer than five offers is flagged as a thin market.
Our series are not designed or offered as a reference rate for settling financial contracts, and nothing here is investment advice.

Seven questions to ask of any GPU price

01Which GPU exactly?
An H100 SXM and an H100 PCIe are different machines; so are 80 GB and 94 GB versions. A price without the variant compares unlike things.
02Which term?
On-demand, spot, interruptible and reserved capacity are priced differently; a monthly commitment is not an hourly rate.
03Posted, offered or paid?
A list price, a marketplace offer and an executed rental are three different numbers, and they rarely agree.
04One seller or a blend?
A blended rate can move because the mix of sellers changed, not because any price did.
05How many observations?
A median over three offers is a reading, not a market. Ask for the count.
06Where?
Region and data centre change both the price and the latency; cloud prices differ by region.
07When?
Rental prices move within a day. A number without a timestamp cannot be checked.