Per-Second GPU Providers

How does per-second GPU billing work?

With per-second billing the meter counts each second a GPU is yours, from start to stop. The rate per hour stays the same. What changes is how much you pay for short runs.

Compare GPU rates by vendor

rent.dev bills its own sandboxes by the second. Its GPU pods run on your own provider key, and that provider bills you.

From start to stop

The meter starts when the machine or container is created, not when your code begins. Boot time, pulling the image and loading the model are all on it. It stops when you terminate the machine.

Minimums per start

Many providers add a minimum charge each time a machine starts. With a minimum of a few minutes, a job that ends early still pays for the whole minimum, so batch short jobs onto one machine.

Stopped is not ended

A stopped GPU machine can keep billing for its disk, its address, or even the whole machine while it holds the hardware. Terminate it when the work is done, and delete volumes you no longer need.

When it saves money

Per-second billing pays off for work that is short or stops early: tests, evaluations, bursts of inference and runs that crash. A long training run sees almost no change, because rounding up the last hour is small next to the whole run.

Per second at rent.dev

rent.dev meters its own sandboxes by the second from one prepaid balance. GPU pods are different: they run on your own key at the GPU provider rent.dev connects to, and that provider bills you on its own terms.

Questions

Which GPU providers bill per second?

Check the provider's own pricing page for a sentence that says so, and for any minimum per start. The table on this site's home page quotes that sentence for each provider on rent.dev's price board, or says the page has none.

Is per-second GPU billing cheaper than hourly?

The rate is the same. You save the unused part of the last hour, so the saving is large for short runs and small for long ones.

Does the meter run while a GPU machine boots?

On most platforms it does. Boot, image pulls and model loading all count, so a lean image and cached weights cut the cost of every start.

Does rent.dev bill GPUs by the second?

rent.dev bills its own sandboxes by the second. Its GPU pods run on your own key at the GPU provider rent.dev connects to, and that provider bills you for them.

Run a per-second sandbox

Agents can start at rent.dev/llms.txt, which lists every route.