GPU rental rates per hour: A100, H100, L40S, 4090 and L4 across 13 providers

Renting a GPU by the hour looks simple until the thirteenth tab is open: one provider quotes per whole node, the next per GPU, the marketplace quotes a floor no listing honours, and the euro prices hide in an API. This page puts the thirteen rental providers we track on one board for the five GPUs renters actually search for, every cell generated from the catalogue behind it, every row stamped with the capture date and linked to the source page.

The board: on-demand $/hr by provider

Cells are on-demand hourly rates per GPU; Vast.ai cells are marketplace floor prices with the median beside them in the catalogue notes, RunPod cells are Community Cloud with Secure Cloud beside them in the notes, and Scaleway cells are euros. A dash is an unpublished price, not a zero price: Novita publishes its instance rates through a public market API, and the catalogue records what the API returned on the capture date. On the capture date Novita's market API showed rentable stock only for the RTX 4090 ($0.33/hr) and the L40S ($0.55/hr); the RTX 5090 ($0.56/hr), the RTX 6000 Ada ($0.77/hr) and the H100 SXM ($3.39/hr) listed prices but zero available GPUs; no A100 SXM SKU listed at all. FluidStack publishes no per-GPU prices, and Scaleway’s L40S page carries no price, so those cells say nothing instead of repeating an old figure.

How to read a cell

On-demand is the walk-up rate with no commitment; the catalogue behind this page also tracks spot hourly and reserved monthly figures where the provider publishes them, and the notes name the exact configuration each price buys, like an 8-GPU node divided per GPU or a two-GPU minimum. Marketplace floors move intraday, so treat the Vast.ai column as a starting bid and check the live offers before budgeting. The capture window on the table is the oldest and newest provider check behind its rows.

How we checked, and what would change our answer

On 2026-10-27 we re-read all thirteen provider pricing pages: static tables and docs directly, JavaScript pages through a rendered reader, OVH’s embedded price list, Scaleway’s public instance pricing API across eight zones, and Novita’s public market API, whose per-SKU availability flags decide which of its rates are rentable and which are listed but out of stock. Prices a page no longer states were cut from the row rather than carried forward: DeepInfra’s dedicated GPUs repriced upward across the board, three Jarvislabs rates rose, Vast.ai’s marketplace moved on every tracked GPU, and OVH’s L4 flavor nearly doubled. A drift test fails the site build when the catalogue moves, so this page is regenerated instead of rotting: a repriced GPU, a new fee, or a newly listed Novita SKU breaks the build until the row is rewritten.