NVIDIA Open-Sources NVKE and NVKR for Managed Kubernetes on GPU Infrastructure

NVIDIA has released two open-source projects, NVIDIA Kubernetes Engine (NVKE) and NVIDIA Kubernetes Recipes (NVKR), to support managed Kubernetes on GPU infrastructure.…

By Vane November 17, 2026 4 min read

NVIDIA has released two open-source projects, NVIDIA Kubernetes Engine (NVKE) and NVIDIA Kubernetes Recipes (NVKR), to support managed Kubernetes on GPU infrastructure. Both tools are part of the NVIDIA DSX AI Factory Platform and are currently in public preview under the Apache 2.0 licence.

Key specifications and limitations

The architecture requires a management cluster on three CPU servers, plus a separate control plane for each workload cluster on three CPU servers, alongside a pool of GPU worker nodes. A local evaluation environment runs on six CPUs, 12 GiB RAM, and 100 GiB disk space without needing a GPU.

Systems run on NVIDIA GPU bare-metal servers provisioned through the NVIDIA Infra Controller (NICo). Node images are available for amd64 and arm64 architectures. NVIDIA has not published benchmarks for throughput or provisioning time; the quality signal relies on certification rather than speed metrics.

The projects are certified as a Kubernetes AI Platform and a Kubernetes Installer, with conformance testing on Kubernetes 1.36. The primary drawback is that the NVKE API serves plain HTTP on port 30080 with authentication disabled after installation. There is also no defined procedure for upgrading or backing up the management plane.

What is NVIDIA Kubernetes Engine?

NVKE is software that allows an AI factory to operate a managed Kubernetes service on its own NVIDIA GPU hardware. The repository describes it as a self-hosting factory for Kubernetes clusters rather than a distribution. It exposes a REST API and controllers on a management cluster. Customers receive standard clusters and interact with them using kubectl. The operator manages these clusters via the NVKE API or the nvkectl command-line tool.

The software uses upstream Kubernetes, kubeadm, and Cluster API. It is written in Go and does not have a runtime dependency on NVKR.

What does NVIDIA Kubernetes Recipes do?

NVKR is a build system that uses declarative definitions to produce executables, container images, Helm charts, and node images for amd64 and arm64. Recipes define validated combinations of software versions and GPU generations. Hosted workflows attach signatures, Software Bills of Materials (SBOMs), and provenance data to digest-pinned OCI references. The tool can also build artifacts for other Kubernetes management systems.

How does the NVKE architecture work?

The NVIDIA technical blog outlines five stages in the process:

  1. Bootstrap: The command nvkectl bootstrap site starts a temporary single-node kind cluster on a bootstrap host.
  2. Management cluster: NVKE requests three servers from NICo, which boots them from the node image. The temporary cluster is then deleted.
  3. Machine provisioning: NICo handles the bare-metal lifecycle. CAPNICo, a Cluster API provider, connects NICo to NVKE.
  4. Release sets: Operators publish a machine image, a cluster release, and a node pool release. Customers select a release by name.
  5. Workload clusters: Each customer cluster has its own control plane, etcd, and node pools. NVKE installs OVN-Kubernetes and NodeLocal DNSCache.

Before a node receives join credentials, it must present an attested SPIFFE identity. The README states that SPIRE uses TPM for this purpose. Per-cluster etcd encryption is available. An OpenTelemetry-based Helm chart exports logs, metrics, and traces over OTLP.

Inside workload clusters, three companion projects integrate: AI Cluster Runtime (AICR) for the GPU runtime, NodeWright for node packages, and NVSentinel for break-fix automation.

Comparison with Gardener and Kamaji

The following table compares NVKE and NVKR with Gardener and Kamaji.

NVKE + NVKR (NVIDIA)Gardener (SAP)Kamaji (Clastix)
What it isManaged Kubernetes management plane plus artifact builderKubernetes clusters as a serviceHosted control plane manager
Control plane modelDedicated per cluster: own control plane, etcd, node poolsRuns as pods in seed clustersRuns as pods; shared Datastore possible
Target infrastructureNVIDIA GPU bare metal via NICoAWS, Azure, GCP, OpenStack, Alicloud (conformance table); extensibleWorks alongside Cluster API providers
LicenseApache 2.0Apache 2.0Apache 2.0
ConformanceCertified Kubernetes AI Platform and Installer; tested on 1.36Certified up to v1.35Described as CNCF-compliant
Node identitySPIFFE/SPIRE attestation with TPMNot disclosedNot disclosed
Signed build pipelineYes (NVKR: signatures, SBOMs, provenance)Not disclosedNot disclosed
MaturityPublic preview, not for productionSelf-described production-readyNot disclosed

Who is adopting NVKE and NVKR?

NVIDIA Cloud Partners Firmus and SK Telecom are integrating both projects into AI factories. Mirantis, Rafay, Spectro Cloud, OpenNebula Systems, and vCluster are adding components to managed Kubernetes offerings. Netris is enabling NVKE for network automation.

How do you try NVKE?

A local quickstart runs on one Linux host or VM, supporting amd64 or arm64. It requires at least six CPUs, 12 GiB free memory, 100 GiB free disk, Docker, and virtual TPM support. No GPU or NVIDIA account is needed. A real installation requires a NICo deployment with enrolled capacity, a build host with KVM, an HTTP server, and a registry.

NVKR has no qualified release yet.

What it means

Developers can now build and sign their own cluster images using NVKR, then deploy them via NVKE on NVIDIA bare-metal hardware. This removes the need to rely on NVIDIA’s cloud-managed services if they prefer to host everything themselves. However, the public preview status and the lack of management-plane backups mean this setup is not suitable for production environments at this stage.

Scroll to Top