Replicate H100 $0.001525/s is $5.49/hour vs fal H100 $4.50 list

Replicate's pricing page (read 2026-10-02) lists an Nvidia H100 at $0.001525 per second, which is $5.49 for an hour. fal's pricing page (read 2026-10-02) lists H100 at $4.50 an hour list and $2.49 an hour in a discounted column. If you only ever call hosted models by the output, none of these hourly numbers is your bill.

Per second versus per hour

Replicate prints per-second prices, so the hourly figure is the rate times 3,600. fal prints per-hour prices with a list column and a discounted column; the page I read does not say what qualifies for the discount, so treat the lower column as conditional until you check.

The per-output price as a break-even

Replicate also lists per-output prices for some models, for example Wan 2.1 at $0.09 per video second at 480p and $0.25 at 720p. At $0.25 per output second, dividing by the H100 rate of $0.001525 per GPU-second gives about 164 GPU-seconds. If you ran the same model yourself and one second of 720p video took your H100 less than about 164 seconds, renting the GPU would be cheaper; if it took longer, the per-output price is the better deal. That ignores idle time, setup and your own engineering, which usually favors per-output prices for small volumes.

Where Sume sits

Sume does not rent GPUs or run your own weights. You call a catalog model and pay per job: the video docs say the cost is reserved on submit at provider list times 1.25, and usage.cost reports the billable amount. So the question "which hourly rate" does not arise; the question is the per-second rate of the model you call, available from GET /v1/videos/models.

If you need a model that is not in a hosted catalog, an hourly GPU is the right tool and the table above is the place to start. If you need standard models with an audit trail per job, per-output billing is simpler to reason about.