Rent a GPU for your LLM

The order is the price; the recommendation is the arithmetic. Every cloud table sorts by the hourly price the provider quoted, among machines that hold the whole model, and you can re-sort by cost per token or by speed. Where one machine is marked recommended, the rule is written next to it and is the same everywhere: the lowest cost per million tokens among machines that answer at 20 tokens a second or more. Neither the sort nor the rule reads which provider pays us. A provider that pays nothing ranks above one that pays whenever the figures say so — OpenRouter, which pays us nothing, is listed first whenever its per-token price is lowest.

Some links earn a referral. The buttons to the providers below pay this site a share of what a new customer spends. It costs the customer nothing — the price is the same, and a new RunPod account gets a one-time $5 credit on its first $10 — and it is how the calculator stays free. Clicks are counted by provider and by page, never by person.

Store links, too. Where a page names a card the calculation chose — a larger one that holds the model, or the card a device page is about — it may link an Amazon or eBay search for it, and a purchase through one may pay this site a small commission. The link is attached after the card is chosen; no store pays for a position, and no store price is shown because a search page has none this site can vouch for.

Considered and left out, each for a reason that is checked again by 2026-12-22:

• DigitalOcean (10% for 12 months, cash (Awin)): Its GPU prices are rarely the cheapest for a card (H100 at $4.41/h on 2026-09-22), so an honest ranking would almost never send anyone there. Not integrated.
• Massed Compute (up to 10% on approval, 20% discount code for readers): No public price API, and the approved terms (duration, payout) are unpublished. Add once both are in writing.
• Vultr ($35 or $100 per active customer): Almost its whole GPU catalogue was out of stock on 2026-09-22 and it has no public price feed.
• TensorDock, Lium, Verda, Jarvislabs, Hyperstack (none found): No referral programme and no unauthenticated price API to read; an unsourced price is not shown.
• Thunder Compute, Hyperbolic, Spheron, Clore.ai (credits only or token amounts): Nothing a reader gains from, and nothing withdrawable.
• Hetzner, Cudo Compute (closed): Hetzner ended its referral programme in June 2026; Cudo's affiliate page returns 404.

The programme terms were read from each provider on 2026-09-22. Who runs this site.