GPU cloud prices compared by what the job really costs
Compare GPU cloud prices from 12 providers' own pages, dated. Price a whole job with storage and egress, size VRAM for your model, and decide spot or reserved.
The short version
Renting the same GPU can cost three times as much depending on where you rent it. On 26 October 2026, an on-demand H100 cost from $2.50 an hour (PCIe, Hyperstack) to $12.29 (per GPU in Azure's 8-GPU ND H100 v5 VM), read on each provider's own price page or API.
The hourly rate is only part of the bill. What else changes the total:
| Cost driver | Why it matters |
|---|---|
| Billing step | Per second, per minute or per hour. A 10-minute test billed by the hour costs 6 times as much. |
| Minimum unit | Some clouds rent H100s only as 8-GPU nodes. |
| Idle storage | Volumes can cost more while the GPU is stopped. |
| Egress | Downloading checkpoints can be free or charged per GB. |
| Commitment | Reserved rates need months or years of commitment. |
How we rank
By estimated total cost for your job. No provider pays us or chooses its position. See the methodology.
Start here
- Price a job across every provider with a price for your GPU.
- Find which GPU fits your model, from 8B to 405B.
- Browse the dated price table and download it as CSV.
- H100 price per hour, A100 vs H100 and H100 vs H200.