Self-host DeepSeek V3 on H200.

H200 is listed under Nebius. DeepSeek V3 on that GPU is a recipe to test, not a number to advertise. denv has not published tokens per second or a required GPU count for this pair. The H100 page explains why a count would be a guess. The same reason applies here, with more memory per GPU and still no measurement.

You connect a Nebius project or join your own server. denv launches machines only in an account you control, up to the maximum you set. Nebius bills you. A joined server is a fixed pool: repair can reset and reinstall the driver, and it cannot buy another H200.

The endpoint is OpenAI-compatible once the environment is ready. llm-d is available for cache-aware routing and for a prefill/decode split if the recipe’s profile is large. Failover needs another site that is already warm. A cold H200 region is not a failover target until the weights are cached there.

Write to anup@denv.ai. Bring the checkpoint revision you actually want served. “DeepSeek V3” names a family. The recipe names a hash.