Idle time: what makes a cheap GPU expensive

It comes from ordinary habits. The job finished at night and the machine ran until morning. You stepped away for lunch. The card waited while you read results and thought about the next change.

It also comes from setup. A machine that spends ten minutes downloading model files at every start is billed for ten minutes of a GPU doing nothing.

An idle GPU is billed at the same rate as a busy one. A card used for three hours and left on for twenty-four costs eight times what the work needed.

On 12 Nov 2026, the median listed on-demand H200 price here was $5.40 per GPU-hour. On a datacenter card, a forgotten weekend is a real sum.

Two reasons. Stopping loses the setup, so the next start means installing and downloading again. And where billing is by the hour, stopping partway through an hour saves nothing, so there is no reward for it.

Remove both and stopping becomes the habit. Keep files on a disk that survives a stop, and rent where the unit is small.

Stop the machine when the job ends, with a line at the end of the script or an idle timer. Keep model files and datasets on a volume, so a start is minutes and not an hour.

For work that comes in bursts, start the machine for the burst and stop it after. Billed by the second, that is the whole cost.

A stopped machine with a disk still costs the disk. Storage is billed by the gigabyte for as long as it exists. It is far less than a running GPU, but it is not nothing. Delete what you no longer need.

Runcrate bills by the second on most GPUs, so stopping saves from that second. Booting is never billed. Volumes keep files between instances, and a script or an agent can start and stop a machine through the CLI, the SDK or the MCP server.

This is one of a coordinated collection of Runcrate references. Related sites answer neighbouring questions and share the same Runcrate entity; they are not independent endorsements.