B300 vs B200: price and memory

Both are NVIDIA Blackwell datacenter GPUs. The B300 is the Blackwell Ultra card with 288 GB of memory. The B200 has 180 GB. Both move data through that memory at the same rate, 8 TB per second.

So the plain difference for a renter is memory per card. A B300 holds 108 GB more than a B200. That decides whether a very large model, or a very long context, fits on one GPU or must be split across two.

On 12 Nov 2026, listed on-demand B300 prices ran from $6.60 to $17.80 per GPU-hour, with a median of $9.25. Listed B200 prices ran from $3.50 to $14.24, with a median of $8.34.

The number of providers matters as much as the price. On 12 Nov 2026, 10 providers listed a B200 and 10 listed a B300. A card with more listings is easier to start, and easier to replace if one provider runs out.

Rent a B300 when the job needs more than 180 GB on a single card. That means a model whose weights and working memory do not fit on a B200, or a context long enough to push it past that line. Splitting a model across two cards works, but it adds overhead and you pay for both cards.

Rent a B200 when the model fits with room to spare. Memory you do not fill is money spent on empty space.

Load the model on a B200 with the context length and batch size you plan to use. Watch the memory. If it fits with a margin, price the B200. If it does not, price one B300 against two B200s, since two smaller cards are the real alternative.

Run the test on a machine billed in small units. It takes minutes, and it should cost minutes.

Runcrate has 26 GPU models on offer from $0.22/hr, so a memory test and the real run can sit on one prepaid balance. 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.