Per-second billing for cloud compute

Per-second billing means the meter counts seconds from start to stop instead of rounding each run up to the hour. It matters most for short and bursty work: a job that runs for ninety seconds on an hourly meter pays for sixty minutes. The label hides details worth reading in each provider's terms, such as a minimum charge per start, which machine states still bill, and whether boot time is counted. rent.dev's own sandboxes bill per second from one prepaid balance; GPU pods run on the agent's own key at the GPU provider rent.dev connects to, on that provider's billing terms.

rent.dev: one API for the sandboxes and GPU pods an agent needs.

Task prices, from the shortest task up

Each price is one task on a 2 vCPU, 4 GiB sandbox, from rent.dev's price board, generated 2026-11-14. See the whole board. Browse the full catalog.

Questions

Which cloud compute providers bill per second?

Large clouds commonly bill virtual machines per second or per minute, often after a one-minute minimum, though some licensed operating system images still bill by the hour. Serverless containers and functions are usually metered per second or in smaller slices of request time. GPU marketplaces and specialised GPU clouds mostly meter by the second or the minute while an instance runs. Check the increment and the minimum in the provider's own pricing terms, because they can differ between two products from the same provider. rent.dev's own sandboxes are metered by the second; on its price board, generated 2026-11-14, a 30 min agent loop task on 2 vCPU and 4 GiB costs $0.0452 on them.

How does per-second billing work for serverless GPU workloads?

A serverless GPU platform meters the time a container holds a GPU, from when it starts to when it scales down, usually per second. The parts that surprise people are cold starts, which are often billed while the model loads, and the idle window a container is kept warm before it scales to zero. A shorter idle window cuts cost and adds cold starts; a longer one does the reverse.

How does pay-by-second pricing change GPU costs?

It removes the rounding waste on short runs, so many small jobs cost what they used. It does not change the rate: a GPU-hour costs the same either way, so a long training run sees little difference. The saving is largest for evaluations, inference bursts and experiments that stop early.

Does per-second billing have a minimum?

Often, yes. A one-minute minimum per start is common for virtual machines, and some platforms round each request or container start up to a fixed slice. Many short starts can then cost more than one longer run doing the same work, so batch small jobs onto one machine where you can.