The cheapest cloud GPU providers, GPU by GPU.

It depends on the GPU. For each of the eight models above, the five lowest published on-demand prices come from different providers, and the ones that appear most often are Verda, Vast.ai, Lium. On every one of the eight, SpotGPUs is priced at least 20% below the lowest published rate, by construction: the catalogue builder refuses to publish a price that breaks the rule.

They are the published hourly rates for one GPU, excluding storage, egress, taxes and any minimum commitment. Some providers bill per second, some per hour started, some per month; some charge for egress, some do not. The table tells you where to start, not what the invoice will say. Ours is billed per minute with no egress fees, and the price list is the whole bill.

They are sold as workstation or gaming parts, so they cost a fraction of a data-center GPU to buy, and marketplaces resell many of them from small operators. For models that fit in 24 or 32 GB, they are often the best price per unit of work. The trade-off is memory, the lack of NVLink between cards, and vendor licensing that keeps them out of most enterprise clouds.

Prices were recorded from the source linked on each row, with the date beside it, and the index is refreshed when we re-check it; our own prices are read from the live catalogue on every render. A wrong figure is corrected when someone sends the link.

Spot or interruptible capacity on a consumer card. Here that is $0.03 per GPU-hour on RTX A4000, billed per minute, and the cheapest data-center card on spot is the T4 at $0.04. The catch with spot is the interruption: read how it works here before you plan a deadline around it.