Fine-tune cost on an H200

The H200 is a Hopper datacenter GPU with 141 GB of memory, more and faster than the older card it replaced. For fine-tuning, that memory holds larger models, longer sequences or larger batches than a 96 GB card.

On 12 Nov 2026, on-demand H200 prices listed here ran from $3.49 to $10.60 per GPU-hour. The median was $5.40. Runcrate can start an H200 now from $4.67 per GPU-hour.

The spread is wide because the same card is sold singly and inside 8-GPU machines. The lowest listed hourly total for an 8-GPU H200 machine was $27.92.

Time a slice on an H200 and scale it. Multiply the hours by the number of H200s, then by the price per GPU-hour. Two H200s for twelve hours are twenty-four GPU-hours, and the cost is twenty-four times the hourly figure.

A full fine-tune writes whole-model checkpoints. Add their storage for as long as you keep them.

Choose the H200 when the job needs between 96 GB and 141 GB on one card, or when a timed slice shows it finishes your run for less in total than a smaller GPU. Choose several when a full fine-tune of a large model does not fit on one.

If the job fits in 96 GB, price the RTX Pro 6000 first.

Check the smallest rental: one card or a whole machine. Check that the price is on demand and not spot or a term. Check the billing unit. And check that a new account can start it today without a quota request or a sales call.

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.