Best GPU provider for training

Training runs are long and steady. The GPU is busy for days, so the hourly rate is nearly the whole bill and the billing unit hardly matters. What matters is that the machine stays yours until the run ends.

That rules out spot capacity for any run that cannot resume from a checkpoint. It also puts weight on storage, because checkpoints must outlive the machine.

Small models train on one card. Larger ones need a node, most often eight GPUs in one machine. The largest need several nodes linked as a cluster.

Know which you need before you compare. A per-GPU figure for a node means paying for every card in it. On 12 Nov 2026, the lowest listed hourly total for an 8-GPU B200 machine was $28.00.

If the run lasts days, rent on demand. If it lasts a month or more, price a term. Multiply the term rate by every hour in the term and compare that total with the on-demand cost of the hours you will really use.

A term pays off only when the GPU is busy for most of it. Count the idle days between runs before you commit.

Ask whether the card can be started today or needs a sales call. Ask what happens to your disk when the machine stops. Ask whether CPU cores, memory and local storage are in the GPU price.

On 12 Nov 2026, 10 providers listed an on-demand B200 price, with a median of $8.34 per GPU-hour.

Ask also how a term is paid for, and whether you can start the same card on demand before any commitment. An hour on the real hardware answers more than a quote.

Runcrate rents GPUs on demand and lets you prepay a machine for 1 to 12 months at a discount. Clusters go by sealed bids for a month or longer. Volumes keep files between instances, so a checkpoint survives the machine.

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.