Serve a model on a B200 you launch yourself.

B200 shows up under Lambda and Nebius. Amazon EC2 in the current catalog does not list it. Your own server can join if it has the card.

denv does not rent B200s. There is no public denv benchmark on B200. The H100 numbers do not transfer.

What denv does with the card

Same as the other GPUs in the catalog:

• A recipe pins the engine, the precision and the arguments for B200, and it is tested on B200 before it serves.
• The node joins with the GPU driver installed. Repair can reset, reboot, reinstall the driver, or replace a cloud instance.
• MIG and Kubernetes DRA matter more on a larger card, not less. A small model should not occupy the whole GPU by default.
• A second site, H100 or otherwise, is how you fail over when the B200 provider is out of capacity. A failover target has to be a model build that was tested on that other GPU. The recipe system exists so those are not silent mismatches.

Write to anup@denv.ai if you have B200 capacity and want it listed, or if you want a model served on one. Listing is a conformance run, described on list your GPUs.