How GPU rental prices work, and how we measure them

There are two different things people call a GPU rental price, and they behave differently. RunPod publishes a catalogue: a fixed hourly rate per card type, in two tiers, secure cloud and community cloud. The number does not move from day to day unless RunPod decides to change it. Vast.ai is a marketplace: independent hosts list their own machines at their own prices, and what you see is whatever is available at that moment.

We collect both once a day. For RunPod we store the catalogue price of each tier. For Vast.ai we store, for each card, the cheapest listing and the median of all listings, together with how many listings were behind that median. That last number matters more than it looks.

A marketplace minimum is one machine, from one host, on one day. It can be a card in a garage with a slow uplink, or a listing that disappears in an hour. Quoting it as the price of that GPU would be misleading, so the numbers you see on the front page and in the daily reports use the median, and only when enough offers back it. When a card has a single listing, we say so rather than dress it up as a market.

The same logic applies to changes. A card that goes from three cents to thirty cents in a week on a marketplace has usually not become ten times more expensive; one cheap machine left and the next cheapest one was pricier. That is inventory turnover, not a price move, and our reports are told to leave it out.

Vast.ai lists machines, and a machine can carry eight cards. Its hourly price is for the whole box. We divide by the number of cards so that every number on the site is the cost of one GPU for one hour, on every provider. Without that step an eight-card H100 node would look eight times more expensive than a single-card listing of the same chip.

Sorting GPUs by price puts a GTX 1050 Ti at the top. Nobody rents a card because it is cheap; they rent the cheapest card their model fits on. So the ladder on the front page answers a different question: for each amount of video memory, what is the cheapest card that has at least that much? A 24 GB rung, a 32 GB rung, up to 141 GB. Pick the rung your model needs and read the price.

Whether a model fits is arithmetic we can do: parameters times bytes per parameter, plus the attention cache for the context you intend to run, plus a safety margin. Whether it runs fast on that card is a measurement we have not made, and you will not find a claim about it anywhere on this site.

It cannot tell you availability at the minute you click, only what was listed when we collected. It cannot compare a RunPod secure-cloud instance with a Vast.ai host as if they were the same product; one comes with a datacenter and an SLA, the other with whatever the host promised. And it cannot tell you throughput. When you see a price here, read it as the entry ticket, not the total cost of the job.