NVIDIA Introduces NVKE, a Managed Kubernetes Service for On‑Prem GPU Clusters

Introducing NVIDIA Kubernetes Engine (NVKE)

NVKE is a preview‑stage platform that lets organizations run a managed Kubernetes service on their own NVIDIA GPU hardware.

The engine supplies a management plane that platform teams install to create and operate isolated Kubernetes workload clusters for their customers.

Each workload cluster receives its own Kubernetes control plane, etcd datastore, and dedicated node pools, separating it from other clusters.

Customers interact with their clusters using the standard Kubernetes API and the kubectl command‑line tool.

Key Features and Operational Model

NVKE provisions machines through NVIDIA Infra Controller (NICo), the only supported infrastructure provider during the preview.

Cluster creation is triggered via the NVKE API or the nvkectl CLI, which then directs NICo to provision the required GPU nodes.

Node identity is established through SPIFFE/SPIRE attestation using a TPM, ensuring that only verified machines join the cluster.

Etcd data can be encrypted on a per‑cluster basis, adding a layer of confidentiality for cluster state.

The companion project NVIDIA Kubernetes Recipes (NVKR) builds the node images and NVKE components from pinned inputs, providing reproducible artifacts for deployment.

Getting Started and Preview Limitations

Developers can evaluate NVKE without physical GPU hardware by using a local quickstart that runs a simulated management plane and workload cluster on a Linux host or VM.

The quickstart requires at least six CPU cores, 12 GiB of RAM, 100 GiB of free disk space, Docker, virtual TPM support, and other tools listed in the guide.

On macOS, the evaluation runs inside a Colima virtual machine via the local site guide.

For production‑like installations, users must supply their own machines, networking, node images, and container registries, and configure access to NICo.

Because the preview is not intended for production, the API authentication is disabled by default and communicates over plain HTTP after installation.

Operators are expected to configure authentication and protected access before exposing the API to callers.

The management cluster itself does not support self‑upgrade, scaling, or backup/restore procedures during the preview.

Full‑disk‑encryption (FDE) node images are provided only for evaluation and do not guarantee data‑at‑rest confidentiality.

All APIs, configuration options, and behavior are subject to change between releases, and known limitations are documented for prospective users.

Documentation, contribution guidelines, and help channels are available through the project’s GitHub repository.

In summary, NVKE offers a self‑hosting factory model that uses a temporary seed cluster to stand up a management cluster, which then orchestrates multiple isolated workload clusters on demand.

Why This Matters

NVKE gives enterprises the ability to harness on‑prem NVIDIA GPUs for scalable Kubernetes workloads while retaining full control over the underlying infrastructure, a capability that is currently limited to cloud providers.