Providers and bare metal
Add a site from a GPU cloud account or from your own servers with the one-line join, and what each provider supports.
A site is one place your GPUs come from. There are two ways to add one: give denv credentials for your own account at a provider, or run the one-line join on servers you already have. You can mix both in one cluster, and one environment can serve from several sites.
denv does not sell or resell GPUs. Each provider bills you directly, at its own prices.
Provider accounts
| Provider | GPU types | How you connect | Status |
|---|---|---|---|
| Lambda | A10, A100 40GB, A100 80GB, H100, B200 | Your own Lambda API key | Available |
| Nebius AI Cloud | L40S, H100, H200, B200 | Your own Nebius service account key | Available |
| Amazon EC2 | A10G, L4, L40S, A100 40GB, H100 | Your own AWS credentials | Available |
| GPU aggregators | Depends on the aggregator | Your own API key | Coming soon |
To add one, choose the provider in the console, paste the credentials, pick a GPU type, and set the minimum and maximum number of machines. Credentials are sealed into your cluster's secret store. denv launches machines in your account up to the minimum, adds more as replicas need them, removes idle ones, and replaces a machine that repair cannot fix. It never goes above your maximum.
Which regions and GPU types you can launch depends on what your provider account offers at that moment. The providers page lists regions for each.
Your own servers
Hetzner dedicated GPU servers, on-prem servers, colo racks: any Linux machine with an NVIDIA GPU can join. Choose "my own machines" when you add a site, and the console shows a join token and a one-line command. Run it on each server:
curl -fsSL get.denv.ai | sh -s -- --token <token-from-the-console>
The command installs what the node needs, connects out to your cluster, and runs a GPU smoke test before the node is marked ready. It needs no inbound ports and no VPN.
Your own servers are a fixed pool: denv repairs them but cannot launch more or replace them. When repair cannot fix a machine, it stays out of service and you get the evidence to act on.
For GPU providers
If you run GPUs and want customers to deploy on them through denv, see how to get listed. Customers pay you directly. GPU clouds become providers on denv.