Manage Kubernetes Clusters for GPU Infrastructure with NVIDIA Kubernetes Engine

NVIDIA announced two open-source projects in public preview: NVIDIA Kubernetes Engine (NVKE), which runs a managed Kubernetes service on an AI factory's own hardware, and NVIDIA Kubernetes Recipes (NVKR), which builds reproducible node images, container images and packages from version-pinned sources.

Customers of GPU clouds increasingly want capacity through the Kubernetes tools they already use, with the provider running the clusters. Building that service, with provisioning, tenant isolation, upgrades and recovery, is hard. NVIDIA is offering the blueprint under Apache 2.0.

We see this as a real gift to the long tail of GPU clouds. The hyperscalers have spent years building managed Kubernetes. A regional AI factory doesn't have that time. Giving them validated software combinations across GPU generations, plus a management plane that turns bare metal into customer clusters, lowers the barrier to offering a credible service.

The reproducibility piece is underrated. NVKR records every build input, so an operator can say exactly what's in a node image and rebuild it later. For anyone who's chased a driver mismatch at two in the morning, that matters, and so does the requirement that a node present an attested identity before it gets join credentials.

The strategic read is less neutral. NVKE leans on NVIDIA Infra Controller for bare-metal lifecycle and slots into the DSX AI Factory Platform. NVIDIA says the projects can be used independently and that NVKR can build for other systems, which is good. But the default path runs through NVIDIA's stack, and every layer of operator tooling it supplies is one more reason not to bring in other accelerators.

And it's a preview. NVIDIA says plainly that APIs and behavior may change and the projects aren't recommended for production yet. Partners are integrating anyway: cloud partners Firmus and SK Telecom, plus Mirantis, Rafay, Spectro Cloud, OpenNebula Systems, vCluster and Netris. We'd watch how fast that turns into customer-facing services.

In NVIDIA's walkthrough, a bootstrap tool starts a temporary cluster, requests servers through NVIDIA Infra Controller, forms a management cluster, and then provisions customer workload clusters with their own control planes and node pools. Operators publish named release sets so customers pick a validated Kubernetes version without handling OS images. NVKE is conformance tested on Kubernetes 1.36.