“From $0.01/hr” is not a price
Browse GPU comparison sites and you’ll see it everywhere: “from $0.01/hr.” It’s technically true — somewhere on a marketplace, someone’s idle GTX with a strange configuration listed at a penny for an hour. And it is useless as a price, for the same reason “houses from $1” would be: the minimum of a noisy distribution describes the noise, not the market.
Minimum-listing pricing fails in three specific ways. It’s phantom-prone: one mispriced, unrentable, or instantly-taken listing becomes the headline number. It’s volatile in a fake way: sites that track minimums show the same card swinging −90% and +40% month over month — that’s listing churn, not the market moving. And it selects for the worst hardware: the cheapest “H100 listing” is systematically the one with the catch — the odd interconnect, the unverified host, the config nobody wants.
So we never publish a minimum. Every number on this site is a two-stage median: each provider’s median across its listings first, then the median across providers — so no marketplace’s volume, and no single listing, can become “the price.” Minimums do exist here, labeled honestly: the market table’s Low (24h) column shows the cheapest observed listing with its provider, next to the median, so you can see both the floor and the truth at once. The gap between them is itself information.
There’s a second pattern that rides along with “from” pricing: comparison tables that are paid placements — referral links ranked by commission, promoted rows inside the data. We take affiliate links too (they’re how a free site eats), but under a published rule: links never affect prices, rankings, or the Index, and every one is marked. A reference has to choose whose side its numbers are on.
If you want the floor, we show the floor. If you want the price, the medians are the homepage — methodology public, history free back to 2023, every observation timestamped with its raw source stored. That’s the entire difference between a deal feed and a reference: one of them you can cite.
Data in this post is from Price of Compute’s open dataset — live at priceofcompute.com, free via the API. Cite it: see /cite.