Why we're publishing GPU prices with receipts
If you have ever rented a GPU, you know the ritual. Five tabs open, five pricing pages that do not agree on what an H100 is, one of them quietly quoting a committed-use rate next to someone else's on-demand rate, and a comparison site in the middle that says one of them costs $1.99 - a number it copied from another comparison site, which copied it from a blog post, which the provider stopped honoring eight months ago.
We kept running into this ourselves, so we built the thing we wanted: one table, every provider, every number with a receipt attached.
What a receipt is
Open any listing on Compute Cafe and at the bottom you will find the receipt: the exact source the price was read from (a live API endpoint or the provider's own pricing page) and the timestamp of the observation. Not 'updated recently.' The actual time. If the source moves, the number moves. If the source dies, the row is carried over unchanged for at most 14 days, flagged as carried over, then dropped. Nothing is silently refreshed and nothing is estimated.
The collector runs every six hours and commits both the current data and a daily history snapshot to a public git repo. The repo is the audit trail. If we ever get a number wrong, the wrongness is itself versioned - you can see what we showed, when we showed it, and when we fixed it.
CATALOG is not AVAILABLE
Some providers run live APIs that report real, rentable capacity. Some publish a pricing page - a rate card, not a stock count. Both are honest sources, but they are not the same claim, so we label them differently. AVAILABLE means an endpoint said you can rent this right now. CATALOG means the provider publishes this rate, and we make no claim about whether the machine exists today. Comparison sites blur this line constantly. We would rather show you one honest row labeled CATALOG than ten rows implying stock that is not there.
This is also why some providers have thin panels. Voltage Park publishes exactly one on-demand configuration publicly. JarvisLabs lists five offers. We show what the source shows - the alternative is padding, and padding is how the $1.99 H100 happens.
The referral links, disclosed
Some providers on the panel run referral programs, and we participate in them where the terms are honest. When a provider's link is a referral link, the methodology page says so, names the program, and states the exact terms - including when a program pays credits instead of cash. The links never affect ordering: the table sorts by price, full stop. If a provider has no program, its panel row says that too.
We think this is the only way a price site earns the right to be read. The GPU market is young, the prices move fast, and the incentive to round in your own favor is everywhere. So: receipts. Every number, every time.