Methodology

Every number on Compute Cafe carries a receipt: the source it was read from and the time it was observed. Open any listing and the receipt is at the bottom. This page is the standing explanation of where the data comes from, how it is normalized, and what each label means.

Panel value

The headline stat strip shows a panel value: the sum of every listed config's hourly rate, annualized at 8,760 hours. It answers one question - what would it cost to rent every config on the panel, at its listed rate, around the clock for a year. It is a booking-value ceiling for the panel, not revenue, not market size, and not hardware value. Hyperscalers publish the same VM in many regions, and each regional row counts, the same way a listing on any marketplace counts. The number moves with the data on every refresh and is computed from the same public rows, so it can be recomputed by anyone from the rows published across the site.

The panel

The panel is fixed and published here. A provider gets added when we can read its prices from its own API or its own published pricing page. We do not copy numbers from other aggregators, and we do not estimate. If a provider's source fails, its rows are carried over unchanged for up to 14 days and flagged, then dropped. Nothing is silently refreshed.

The collector runs every 6 hours on GitHub Actions and commits both the current data and a daily history snapshot to the public repo. The repo is the audit trail: every change to every number is a commit.

Availability labels

Available means the provider's live endpoint reported the offer as rentable at observation time. Catalog means the provider publishes this rate but no live stock data, so we show the rate without claiming you can deploy it this second. Waitlist and sold out come from the provider directly.

Normalization

Prices are hourly, in USD. Several Indian providers (E2E Networks, InHosted, AIC Cloud, Shakti Cloud) publish in rupees; those rows keep their INR list price on the receipt and are converted at the observed rate from open.er-api.com. Multi-GPU rows are priced per instance, with the GPU count shown on the row. Committed-use and reserved rates are excluded; everything here is on-demand.

Sorting is by price, cheapest first. The color scale ranks each listing against other listings of the same GPU model: green is the cheapest quintile for that GPU. Nothing else affects ordering.

first seen / last seen

Each listing has a stable identity (provider, GPU, count, region). First seen is when that offer first appeared in our data; last seen is the most recent refresh that observed it. A price change does not reset either. Daily history snapshots are the raw material for the price-movement notes we publish; per-listing series are available through the MCP get_history tool.

Money

The site is free. Some outbound links may become referral links as provider programs are set up; when one does, it is marked here and on the provider, given a rel="sponsored nofollow" attribute, and it never changes what we show or how we sort. Ranking is by price. We do not take placement money.

Panel status above is generated from the live refresh metadata, so what you read here is what the collector is actually doing right now.