Nvidia previews NVKE, a bare-metal GPU-Kubernetes factory

NVIDIA has pushed preview code for a system that lets a platform team spin up Kubernetes clusters on its own GPU hardware, each with its own control plane and etcd. The project is called NVKE, short for NVIDIA Kubernetes Engine, and it ships under Apache 2.0 from the `dsx-ai-factory` GitHub organization.

The README's opening line is "Run a managed Kubernetes service on your own NVIDIA GPU infrastructure." What it provides, in the project's own wording, is "the management plane that platform teams install to create and operate Kubernetes clusters for their customers." Each workload cluster, the docs say, has "its own Kubernetes control plane, etcd and node pools." Customers reach those clusters through the Kubernetes API and `kubectl`.

The identity piece is TPM-anchored. "Nodes attest through SPIFFE/SPIRE using their TPM before receiving join credentials," the README states, and "encryption of etcd data is available per cluster."

NVIDIA is explicit about what the project is not. "NVKE is not a Kubernetes distribution and not a Cluster API flavour," the README reads; it calls itself "a self-hosting factory for Kubernetes clusters" where a temporary seed cluster bootstraps a management cluster, which then provisions isolated workload clusters on demand. A companion project, NVKR (NVIDIA Kubernetes Recipes), "builds node images and NVKE components from pinned inputs" — "NVKE deploys those artifacts; NVKR owns their build and release evidence."

The project is preview. "NVKE is available as a preview," NVIDIA writes. "APIs, configuration and behavior can change between releases, and we do not recommend NVKE for production workloads." It is the 219th NVIDIA item our tracker has logged in the last 90 days.