Deploy DeepSeek R1 on H200.

Same model family as the H100 R1 page, different GPU. H200 in the catalog means Nebius, or a server you join. There is no published denv speed, no published GPU count, and no license to reuse the Qwen H100 figures.

R1’s long traces make KV cache the constraint as much as weight size. More memory on H200 is why people ask. It is also why a recipe tested only on H100 cannot be waved through. Cache, precision and engine arguments move with the GPU.

Nebius bills you. denv does not. Repair can replace a Nebius machine up to your maximum. It cannot replace a server you own. Failover to H100 is a second tested build, and a request that already started streaming is not replayed.

If you are deciding between hosting R1 at a token API and self-hosting it, the comparisons are the honest version: denv is the control plane on GPUs you bring. Write to anup@denv.ai. The product is early.