Deploy Llama 3.1 70B on H200.

Llama 3.1 70B on H200 is a recipe denv will test on H200, not a resized copy of an H100 recipe. Weights at 16-bit are about 140 GB before KV cache. At 8-bit, about 70 GB. H200 has more memory than H100, which is why people ask for it, and it is still not a reason for this page to invent a fit or a tok/s.

Nebius is the catalog provider that lists H200. Your own server can join. Lambda and EC2 rows do not list H200. You pay Nebius or you pay for the server. denv does not sell the GPU.

The replica stays inside one site. If the recipe needs more than one GPU, they are in that site, on that network. A second site is another replica, used for capacity or failover.

Engine, precision and arguments are whatever the tested recipe pins: vLLM or SGLang. Repair includes driver reinstall. No public speed number exists for this pair. The H100 Qwen benchmark is a different model on a different GPU.

The product is early. Write to anup@denv.ai to ask for the recipe or to bring your own weights.