How the prices are produced
One automated check a day, every figure tied to the page or API it came from, and rules that keep a bad reading from reaching the site. This page is the full method; the short version is on every table.
What a listing is
A listing is one price for one combination of provider, instance or plan, GPU model, GPU count, region or service tier, and pricing type. The pricing types are on-demand (the hourly list price with no commitment), spot (discounted capacity the provider can reclaim), reserved for one or three years (shown as the reservation total divided by the hours in the term, 8,760 per year), and committed (cluster prices that need a minimum order and term). Each listing has a stable id, so its history can be followed day after day.
Where the figures come from
- Microsoft Azure: Linux pay-as-you-go, spot and one- and three-year reservations from the public retail prices API. Reservation totals are divided by the hours in the term. Prices differ by Azure region; the table lists each region the SKU is sold in. Read from https://prices.azure.com/api/retail/prices.
- RunPod: Per-GPU hourly prices for Secure Cloud (RunPod-run data centers) and Community Cloud (vetted third-party hosts) from the public GraphQL API. RunPod prices a GPU type the same in every data center; the API does not price by region. Read from https://api.runpod.io/graphql.
- Lambda: On-demand per-GPU prices for 1x, 2x, 4x and 8x instances and committed 1-Click Cluster prices read from the pricing page tables. Lambda publishes one price per instance type regardless of region. Read from https://lambda.ai/pricing.
Only public list prices are used. Each provider is read in the order of preference official API, then the JSON behind the price page, then the price page itself. A price that is not public, such as a quote or a negotiated rate, is never shown.
Normalisation
Prices are kept in US dollars as listed. Whole-instance prices are divided by the GPU count to get the per-GPU figure; per-GPU listings are multiplied by the GPU count to show the instance figure. Windows, low-priority and dev/test rows are excluded because they are not comparable with other providers. Instance sizes that carry a fraction of a GPU are excluded from per-GPU comparison. Azure regions marked as government or edge zones are read but not published.
Checks before a price is published
- Range. A per-GPU hourly price below $0.05 or above $200 is treated as a parsing error, not a price. The previous figure stays and the reading is logged.
- Large moves. A price that moved more than 30% since the previous check is held: the previous figure stays on the site with an "under review" mark until the new figure has held for three consecutive daily checks or a reviewer accepts it.
- Missing sources. When a provider does not answer, its previous listings stay on the site marked unverified, with the date they were last read. Listings a provider no longer shows are removed the same day and noted in the change log.
- Coverage. Every expected configuration is listed in advance. One that disappears is reported; one that appears but is not yet mapped to a GPU model is logged and not published until it is.
Timing
The check runs daily around 06:00 UTC. Every table shows the time each listing was read. A listing older than 48 hours is marked unverified wherever it appears. The most recent check finished on October 23, 2026 at 01:06 UTC and recorded 2,626 listings.
History
Each listing's per-GPU price is recorded whenever it changes. Charts show the lowest on-demand price per day for the last 90 days; a flat line means the price did not move. History starts on October 22, 2026, the site's first check.
Specifications
GPU memory, bandwidth and power figures come from the vendor's product page or datasheet linked on each GPU page. Figures marked "second check pending" were typed from the vendor page and are awaiting a second review.
Corrections and data access
If a figure disagrees with what a provider shows, the provider is right and this site needs a fix. The latest data is published as JSON and summarised for AI assistants in llms.txt; reuse is welcome with a link back.