GPUs
Run inference on an H200 without renting it from denv.
Nebius lists H200 in the denv catalog. denv serves the model, checks the GPU, and can fail over. No published H200 tok/s yet.
H200 is in the catalog under Nebius. You can also join a machine you already have. denv does not rent H200s and does not quote a price per hour.
There is no denv tokens-per-second number for H200 on this site. The public figures are H100 only. Treating an H100 result as an H200 result would be a different, worse benchmark.
What changes relative to H100
You still deploy a recipe: one model, one GPU type, vLLM or SGLang, pinned and tested on H200 before it serves. Health checks, GPU driver management, MIG and DRA, and failover are the same mechanisms. The weights cache is per site, so a cold Nebius region does not borrow a warm cache from somewhere else.
Recipe pages that name H200 say which model you are deploying and that speed is not published. Llama 3.1, Qwen3 and DeepSeek are the ones those pages cover.
If your account cannot launch H200 in eu-north1, eu-west1 or us-central1, Nebius’s capacity wins, not this page. The providers page lists those regions as the ones denv has checked.
The product is early. Write to anup@denv.ai before you plan a migration around a GPU type you have not seen come up healthy.