What the calculator adds that price lists don't

Billing step. Vultr and Together AI bill by the hour, so a 20-minute job costs a full hour. DigitalOcean bills per second with a 5-minute minimum, and keeps billing a stopped Droplet (DigitalOcean).
Whole nodes. CoreWeave's H100, H200 and B200 instances and Azure's ND H100 v5 are 8-GPU machines. Ask for one GPU and the calculator bills eight.
Storage while idle. RunPod charges $0.10 per GB-month for a volume on a running Pod and $0.20 when it's stopped (RunPod). Jarvislabs charges storage while an instance is paused.
Egress. Hyperstack, CoreWeave and Lambda don't charge for egress. Modal charges $0.04 per GiB after the first TiB a month (Modal).

How to use the result

The cheapest row is a starting point. Before you commit, check the provider still has capacity in the region you need, and whether your framework needs CPU and RAM that's billed separately (Modal, TensorDock).

When a fee isn't published, the table says "not stated" rather than guessing. That row's total covers only what's known.

Example

Ten hours on one H100 with 200 GB kept for a week and 50 GB downloaded: on-demand compute ranges from about $25 (Hyperstack PCIe) to about $40 (several clouds at $3.99), while a provider that only rents 8-GPU nodes bills eight GPUs. Change the inputs above to see your own numbers.