H200 vs H100 and older cards

Both cards belong to the Hopper generation. The H100 has 80 GB of memory and a memory bandwidth of 3.35 TB per second. The H200 has 141 GB and 4.8 TB per second. The gain is in memory: more of it, and faster.

So the H200 holds models the H100 cannot, and holds more requests at once for models both can run.

Older cards usually cost less per hour. On 12 Nov 2026, the median listed on-demand price per GPU-hour was $5.40 for the H200 and $3.90 for the H100. On 12 Nov 2026, 15 providers listed an H200 and 19 listed an H100.

A lower hourly price is not a lower cost. If a job needs two H100s where one H200 would do, compare the two H100s with the one H200.

Small models do not need 141 GB. If the model and its context fit in 80 GB with room to spare, an older card does the same work for less per hour. The same holds one step further down, for smaller models on cards with less memory again.

The test is fit. Load the model on the smaller card with your real context length and batch size. If it holds, you have saved money. If it runs out, you know why you are renting an H200.

In 2026 the H200 is the Hopper card that large models are sized for. New work tends to be planned around 141 GB or more. Older cards appear here as a comparison only. The tables on this site cover the H200 and the B200.

Runcrate offers 26 GPU models from $0.23/hr, including the H200 and the H100. Billing is by the second on most GPUs, and booting is never billed, so trying the smaller card first costs little. Runcrate can start an H200 now from $4.67 per GPU-hour.

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.