One GPU or several: how GPU count changes the cost

The bill counts GPU-hours. Four GPUs for six hours is twenty-four GPU-hours, the same as one GPU for a full day. If four GPUs finish the job in exactly a quarter of the time, the cost is unchanged and you have your model sooner.

They rarely do. Some time goes to moving data between the cards and keeping them in step. So the GPU-hours of a multi-GPU run are usually a little higher than those of the same run on one card.

Use more than one GPU when the job does not fit in the memory of one, or when the calendar matters more than the bill. A full fine-tune of a large model is the usual case for the first. A deadline is the usual case for the second.

For a LoRA run that fits on a single card, one GPU is nearly always the cheaper choice.

The newest datacenter GPUs are often sold as machines of eight. If your run needs two cards and the smallest rental is eight, you pay for six idle ones. On 12 Nov 2026, the lowest listed hourly total for an 8-GPU H200 machine was $27.92.

So ask for the smallest rental before you read the per-GPU price. The tables on this site print it beside each listing.

Time the same slice on one GPU and on the count you are considering. Multiply each time by its GPU count. If the multi-GPU figure is much higher, the extra cards are buying speed at a premium. Decide whether the earlier finish is worth that.

Runcrate lists each GPU with its smallest rental and its price per GPU-hour. Billing is by the second on most GPUs, so a two-card test costs what it ran. For a month or longer on many GPUs, Runcrate sells clusters by sealed bids.

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.