GPUs
Bring an H100, H200 or B200. denv serves the model.
denv does not rent GPUs. It runs inference on H100, H200, B200 and A100 capacity you already have, with repair and failover.
People search “rent H100” and land on a price table. denv is not that table. denv does not rent or resell GPUs. If you already have an H100, or you are about to get one from a cloud that bills you, denv is the layer that turns it into an endpoint.
GPUs in the catalog
| GPU | Where denv can place it today | Page |
|---|---|---|
| H100 | Lambda, Nebius, Amazon EC2, or any NVIDIA server you join | H100 |
| H200 | Nebius, or a server you join | H200 |
| B200 | Lambda, Nebius, or a server you join | B200 |
| A100 | Lambda (40GB and 80GB), Amazon EC2 (40GB), or a server you join | A100 |
Regions and stock move. The providers page is the list denv has checked. Your account decides what you can actually launch.
What you get on the card
A recipe, vLLM or SGLang, health checks, GPU driver management, MIG and DRA when you want a slice, and failover if you run more than one site. One OpenAI-compatible endpoint.
The only public speed figures are on one H100: 3rd of 23 on Qwen, and 50.8 to 213 tok/s. Nothing on this site claims those numbers for H200, B200 or A100.