Run inference on an H100 you already pay for.

An H100 is the GPU those searches start from. denv will not rent you one. Lambda, Nebius and Amazon EC2 are the clouds in the catalog that list H100, and any NVIDIA server you join can bring one too. You pay them. denv runs the model.

What was measured

Two different tests, both one request at a time, both on one H100:

• 3rd of 23 serving Qwen3.8-27B. 2.6–3.9× a typical provider in that test. The H100 in that race sat in Europe. Median output speed on the published board was 241 tok/s.
• 50.8 to 213 tok/s on one H100 in a Nebius account, by moving from 16-bit weights to FP8 plus a draft model. No new hardware.

Do not average those into one number. Read the caveats on each page. They are not a promise for your prompts, your batch size, or your model.

What to deploy

Recipe pages for Llama 3.1 8B and 70B, Qwen3, DeepSeek V3 and DeepSeek R1 on H100. An 8B model does not need the whole card. MIG and Kubernetes DRA are how a recipe takes a slice. A 70B model at 16-bit is on the order of 140 GB of weights before the KV cache, so the recipe, not a blog post, decides whether one H100 is enough. That memory rule of thumb is in the inference guide.

The engine on the measured runs was vLLM. Driver install is part of joining the node. Auto-repair can reinstall the driver, reset the GPU, or replace the instance if the provider account allows it.

If you are comparing rental prices

Do that on the provider’s site. denv does not publish a dollar per hour, because denv does not sell the hour. When you want the endpoint to survive a bad H100, add a second site and read failover.