NVIDIA open-sources NVKE and NVKR for GPU Kubernetes cluster management

NVIDIA has open-sourced NVIDIA Kubernetes Engine (NVKE) and NVIDIA Kubernetes Recipes (NVKR), two projects designed for building and operating managed Kubernetes services on GPU infrastructure. NVKR produces reproducible node images, container images, and packages from version-pinned sources, while NVKE provisions and manages Kubernetes clusters through a REST API and controllers running on a management cluster. Both projects are in public preview under the Apache 2.0 license and form part of the broader NVIDIA DSX AI Factory Platform.

The NVKE management plane is deployed via nvkectl, which bootstraps a temporary seed cluster, requests servers from NVIDIA Infra Controller (NICo) through CAPNICo, and forms a management cluster that creates isolated customer workload clusters with their own control planes and node pools. AI factories publish validated release sets combining machine images, cluster releases, and node pool releases, allowing customers to select Kubernetes versions without directly handling OS images or cloud-init. NVIDIA Cloud Partners Firmus and SK Telecom are integrating the stack, while Mirantis, Rafay, Spectro Cloud, OpenNebula Systems, vCluster, and Netris are incorporating components into their own offerings.

Architecture and deployment model

• NVKR build reproducibility: NVKR constructs node images, binaries, container images, and Helm charts from specific source versions with recorded build inputs, enabling validated software combinations across GPU generations and Kubernetes versions.
• NVKE management plane: The management cluster hosts a REST API and controllers that provision workload clusters; nvkectl bootstraps this from a temporary seed cluster on a bootstrap host equipped with containerd, nerdctl, kind, Helm, and kubectl.
• Infrastructure provisioning chain: CAPNICo, the Cluster API provider for NVIDIA Infra Controller (NICo), handles bare-metal server lifecycle management, feeding machines into management and workload clusters.
• Security attestation: Nodes must present an attested SPIFFE identity before receiving join credentials, integrating workload identity into the cluster bootstrap process.
• Observability stack: An OpenTelemetry-based telemetry pipeline collects logs, metrics, and traces via a Helm chart, exporting over OTLP to monitoring backends.

Why it matters

NVKE and NVKR give AI infrastructure operators an open, inspectable alternative to proprietary managed Kubernetes stacks for GPU workloads, with Apache 2.0 licensing enabling independent modification and contribution. The separation between NVKR's reproducible builds and NVKE's cluster management allows operators to adopt either component independently, reducing vendor lock-in compared to fully integrated proprietary platforms. For the Kubernetes ecosystem, NVKE's conformance certification on Kubernetes 1.36 and its integration with Cluster API patterns align it with existing cloud-native tooling, though the NICo and CAPNICo dependencies tie bare-metal provisioning to NVIDIA's infrastructure controller.