The hidden lines of a compute bill

A disk that keeps your data while the machine is stopped is billed by the gigabyte for as long as it exists. A stopped machine with a large disk still costs money every day.

Model files and datasets are large. Count the storage you will keep between runs, and for how long, before you compare GPU prices.

Egress is data leaving a provider’s network. Some providers bill it by the gigabyte and some include it. It matters when you move datasets or checkpoints out, or serve large outputs to users elsewhere.

The rate is seldom on the GPU pricing page. Look for it in the network or storage pricing.

A machine left running after the job ended is billed at the full rate. This is the largest avoidable cost on most bills, and the billing unit decides how much stopping saves.

Billed by the hour, stopping ten minutes into an hour saves nothing. Billed by the second, it saves fifty minutes.

A per-GPU figure may or may not include the CPU cores, memory and local disk that come with the card. Two prices that differ by a few percent can describe quite different machines.

Read the machine’s full description, not only the GPU name and the number.

Run one hour of real work on the smallest machine. Keep a little data on a disk afterwards. Move a file out. Then read the invoice line by line.

Every line you find goes into your brief, with its unit. Multiply each by what you expect to use in a month. The total is the price you are comparing, not the GPU-hour alone.

On Runcrate one prepaid balance pays for GPUs, models and storage. Volumes keep files between instances. Billing is by the second on most GPUs, and booting is never billed.

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.