Self-host the control plane on Kubernetes you already run.

Self-hosted denv is the same software as the hosted product, installed in your account. It is not a second product with a second vocabulary. The install notes are in the self-hosted docs. This page is the short version for anyone searching “on-prem LLM inference” or “self-hosted GPU Kubernetes”.

Where it runs

One Helm chart installs the operator, the controllers, the gateway, the network coordinator and the console. You bring:

• a Kubernetes cluster: EKS, GKE, OpenShift or k3s
• PostgreSQL 15 or later
• S3-compatible object storage

The GPUs can still be Lambda, Nebius, EC2, Hetzner, a colo, or machines on your own floor. They join outbound. You open no inbound ports.

Air-gapped installs mirror the images to your registry. The join command pulls from there. Nothing calls home. Usage telemetry is opt-in.

What you still pay

The license is per GPU under management. The GPU bill, if there is one, stays with the provider or with your own power and hardware. denv does not sell GPUs and does not sell tokens. There is no free tier. Write to anup@denv.ai for a license.

What “like EKS on EC2, for inference” means here

If you already run EKS, you can put denv on it and still attach GPU nodes that are not EC2. The control plane is yours. The recipe, the OpenAI-compatible endpoint, MIG and DRA, health checks and failover are the same ones the hosted product uses.

You upgrade on your schedule by bumping the chart. Nodes stay compatible one minor version back and forward.

What denv will not operate for you

The management cluster, the database, and the network between your sites are yours. denv supports the chart, the controllers and the node software. If you want denv to run the control plane, use hosted instead and read pricing.

The product is early. The docs say so, and this page says so again.