NVIDIA Kubernetes Engine runs managed Kubernetes on customers' own GPU hardware

A managed service on hardware you own

NVIDIA has published NVIDIA Kubernetes Engine (NVKE), a management plane that platform teams install to run a managed Kubernetes service on their own GPU infrastructure. The project reached Hacker News's front page on 6 November 2026 through a GitHub repository in the dsx-ai-factory organisation, which carries NVIDIA's copyright. According to the project's documentation, operators run the NVKE management plane themselves and use it to create and operate Kubernetes clusters for their customers, who then access those clusters through the standard Kubernetes API and kubectl.

Each workload cluster gets its own Kubernetes control plane, etcd instance and node pools. NVKE is deliberately not a Kubernetes distribution and not a Cluster API flavour, the documentation notes. Instead, a temporary seed cluster stands up a management cluster, and the management plane running on it then provisions and lifecycle-manages many isolated workload clusters on demand.

How it works

Clusters are requested through the NVKE API or the nvkectl command-line tool, and machine provisioning flows through NVIDIA Infra Controller (NICo), which is the only compatible infrastructure provider during the preview. The management plane handles the full cluster lifecycle: node pools can be created, scaled or removed, Kubernetes versions upgraded and individual nodes replaced.

Node identity gets explicit treatment. Before receiving credentials to join a cluster, nodes prove their identity through SPIFFE/SPIRE attestation backed by the machine's TPM, and etcd data can be encrypted on a per-cluster basis. A companion project, NVIDIA Kubernetes Recipes (NVKR), builds node images and NVKE components from pinned inputs; NVKE deploys the resulting artifacts, while NVKR is responsible for producing them and documenting their release.

Trying it and the preview caveats

For evaluation without GPU hardware, a local quickstart runs a development management plane and a workload cluster on simulated machines, with no NVIDIA account required. It needs a Linux host or VM, on amd64 or arm64, with at least 6 CPUs, 12 GiB of free memory and 100 GiB of free disk, plus Docker and virtual TPM support; on macOS it runs inside a Colima virtual machine. Real installations shift the burden to the operator, who supplies the machines, networking, node images and registries.

The preview has sharp edges. API authentication ships disabled and the API serves plain HTTP after installation, so both must be configured before admitting any callers. The management cluster cannot upgrade, scale or replace itself, and there is no management-plane upgrade or backup and restore procedure yet. Full-disk-encryption node images are for evaluation only and provide no confidentiality for data at rest. The project's maintainers do not recommend NVKE for production workloads while APIs, configuration and behaviour can change between releases.

The code is open source under Apache License 2.0, builds with Go 1.27.3 and GNU Make 4 or later, and produces nine binaries, including nvkectl, a management engine and separate operator and tenant APIs.

Why it matters

Managed Kubernetes has traditionally been a cloud-provider product. NVKE moves that operational model onto hardware the customer owns, which matters most for GPU fleets, where rented capacity is expensive and often constrained. Platform teams at organisations that have already invested in NVIDIA servers get a path to offering self-service clusters with per-tenant control planes and hardware-attested node identity, without outsourcing the infrastructure underneath.

The maturity gap is the catch: with authentication off by default, no management-plane upgrade path and no backup story, this is evaluation software today. Even so, the release extends NVIDIA's reach from silicon into the platform layer, and it gives the broader Kubernetes ecosystem a concrete blueprint for a self-hosted, multi-tenant cluster service on bare metal.