NVIDIA has published two new open-source projects in public preview to support building and operating managed Kubernetes services on GPU infrastructure: NVIDIA Kubernetes Engine (NVKE) and NVIDIA Kubernetes Recipes (NVKR). Both are part of the NVIDIA DSX AI Factory Platform and are available on GitHub under the Apache 2.0 license.
Purpose and significance
AI Factory customers often want to consume GPU capacity through existing Kubernetes APIs, tools and ecosystem while the AI Factory operates their clusters. Delivering such a service requires API-driven provisioning, tenant isolation, upgrades, recovery and service commitments. It also requires qualifying GPU hardware, networking, drivers and runtime software across GPU generations and Kubernetes versions.
NVKR and NVKE split this work: NVKR focuses on reproducible software and image builds, while NVKE provides the management plane and APIs to provision and operate managed Kubernetes clusters for customers.
What NVKR does
NVKR defines NVIDIA-validated Kubernetes software configurations and provides a reproducible build pipeline. From version-pinned sources NVKR produces node images, binaries, container images and Helm charts, and records every build input so teams can identify and rebuild artifacts later.
The walkthrough in NVIDIA’s announcement shows teams cloning the nvkr repository and running quickstart scripts to build nvkectl, the NVKE controller image and a node image (named node-image-01 in the example). The build produces a QCOW2 disk image that includes compiled NVKE node agents, OVN-Kubernetes from source, checksum-pinned Kubernetes 1.36 packages and the SPIRE agent.
What NVKE does
NVKE runs the management plane for a managed Kubernetes service and exposes APIs to provision and manage clusters. Using nvkectl, operators start a temporary seed cluster on a bootstrap host that forms the management cluster; the NVKE REST API and controllers run on that management cluster. The management plane creates workload clusters for AI Factory customers, each with its own Kubernetes control plane, etcd and node pools.
Bare-metal lifecycle is managed by NVIDIA Infra Controller (NICo) via CAPNICo, the Cluster API infrastructure provider for NICo. Nodes must present an attested SPIFFE identity before receiving join credentials. NVKE’s telemetry pipeline, delivered as an OpenTelemetry-based Helm chart, collects logs, metrics and traces and exports them over OTLP to a monitoring backend.
NVKE is conformance-tested and holds Certified Kubernetes AI Platform and Certified Kubernetes Installer recognition on Kubernetes 1.36.
How the projects work together and independently
NVKR builds the software artifacts; NVKE consumes those artifacts to provision and operate clusters. Teams can adopt NVKE and NVKR together or independently: NVKE has no runtime dependency on NVKR, and NVKR can produce artifacts for other Kubernetes management systems.
Example deployment workflow
The public walkthrough covers these main steps and prerequisites:
-
Prerequisites include: three CPU servers for the management cluster and a separate Linux bootstrap host with containerd, nerdctl, kind, Helm and kubectl; three CPU servers for the workload cluster control plane; a pool of GPU servers for worker nodes; a NICo installation with enrolled capacity; a Linux build host with Docker and KVM; an HTTP server for the node image and a container registry for the controller image; and internet access to download open-source images and charts.
-
Build the node image with NVKR: clone the nvkr repository, run the quickstart to build nvkectl and the controller image, then run the OS image script to create node-image-01. The process compiles agents and stages Kubernetes and networking packages to produce a QCOW2 image.
-
Deploy the NVKE management plane: on the bootstrap host run nvkectl bootstrap site -f site-a.yaml --nvke-controller-image registry.example.com/nvke/nvke-controller:<tag>. The bootstrap command starts a temporary kind-based cluster on the bootstrap host, the NVKE controller requests three servers from NICo, NICo boots them from the node image, and once the nodes have joined the temporary cluster is removed to leave the management cluster.
-
Publish a release set: AI Factories publish named releases for customers; a release set consists of a machine image (iPXE script), a cluster release (machine image + cloud-init and the recorded Kubernetes version) and a node pool release. These are created with nvkectl create machineimage/clusterrelease/nodepoolrelease commands.
-
Create a customer cluster: a cluster request includes NICo inputs and the Kubernetes version (the article uses 1.36.3 as an example). Each control plane and node pool references release names (example: k8s-1.36.3-amd64). Submit the request to the NVKE API with nvkectl create cluster -f customer-a.yaml. The cluster remains in Provisioning until the first control-plane node runs kubeadm init and NVKE installs OVN-Kubernetes and NodeLocal DNSCache; the status then becomes Provisioned.
-
Give the customer access: the customer uses a kubeconfig to access the cluster API while the AI Factory manages the cluster through the NVKE API. Operators retrieve the kubeconfig with nvkectl get kubeconfig customer-a --output-file customer-a.kubeconfig.
Partners and integrations
NVIDIA Cloud Partners Firmus and SK Telecom are integrating NVKE and NVKR into their AI Factory offerings. Other partners—Mirantis, Rafay, Spectro Cloud, OpenNebula Systems and vCluster—are integrating NVKE/NVKR components into their managed Kubernetes products to help AI factory customers run GPU AI workloads on Kubernetes. Netris is enabling NVKE in its network automation solution.
Availability and limitations
NVKE and NVKR are available today in public preview but are not yet recommended for production workloads. The projects are evolving rapidly and APIs, configuration and behavior may change between releases. NVIDIA provides public roadmaps and documentation and invites feedback and contributions from users.
Getting started
Source code and documentation are hosted in the NVIDIA DSX AI Factory GitHub organization, along with CAPNICo. Teams can start with the NVKE documentation or try the local quickstart on a single machine (no GPU or NVIDIA account required). NVIDIA also hosts a Slack workspace channel (#nvke) for the community and will present additional demos at KubeCon (NVIDIA booth #170).
