Nvidia has published a project called NVIDIA Kubernetes Engine — NVKE — at github.com/dsx-ai-factory/nvke. It is Apache 2.0, written in Go, and described in its own first line as "a management plane for GPU Kubernetes on bare metal". It is the most interesting thing the company shipped this week that nobody announced as a product. It is also the most honest, because the README says outright that you should not run it in production.
What it does
NVKE is not a Kubernetes distribution and not a Cluster API flavour. It is, in its own documentation's phrase, a "self-hosting factory for Kubernetes clusters". An operator installs the management plane. That plane then creates independent workload clusters on demand — through an API or the nvkectl command-line tool — each with its own control plane, its own etcd and its own node pools. The customer gets a kubeconfig and uses kubectl; the customer never sees the management plane.
Beneath that sits a second layer that gives the game away. Machines are provisioned through NVIDIA Infra Controller, or NICo, which is the only supported infrastructure provider at preview — a polite way of saying Nvidia's own hardware, for now. Nodes attest through SPIFFE/SPIRE using their TPM before they receive join credentials; etcd encryption is available per cluster. A companion project, NVKR, builds and signs node images and components from pinned inputs so that what gets deployed is traceable. Building from source yields nine binaries, among them agent, controller, nvkectl, management-engine, tenant-api and operator-api.
To try it, you need a Linux host with six CPUs, 12 GiB of free memory and 100 GiB of disk, plus Docker and a virtual TPM — on macOS, inside a Colima VM. You can stand up a simulated management plane and workload cluster without owning a single GPU.
Where it becomes honest
The limitations section is unusually forthright, and worth reading as a document in itself. After installation, API authentication is off and the API speaks plain HTTP; an operator who exposes it before configuring auth has an open control surface. The management cluster does not upgrade, scale or replace itself. There is no management-plane upgrade procedure and no backup or restore procedure. The full-disk-encryption node images are for evaluation and provide no confidentiality for data at rest. APIs and behaviour can change between releases.
That is not marketing copy. That is a build guide with honest warning labels, and it says something about the maturity: working, but without the unglamorous work that turns a demonstration into a product.
Why Nvidia is doing this
Nvidia sells the machines. But running a GPU cloud requires twenty further decisions: how nodes are taken possession of, how they are partitioned, how a tenant gets a cluster rather than a card, how usage is metered, how a node that dies at 3am gets replaced. Until now each operator answered those itself, or bought the answer from VMware, Red Hat, or one of the many Kubernetes management vendors. NVKE is Nvidia's bid to supply that answer itself — and thereby to be the default shape in which Nvidia's hardware gets divided.
That is not a side quest. AI infrastructure is being built in enormous, hastily constructed halls, often by operators who were doing something else two years ago. Hand them a working control tool and you set the operational shape for years. Nvidia's acquisition of Run:ai, now fully open source under Apache 2.0, points the same way: the software layer above the GPUs should be free, open, and shaped by Nvidia.
The catch
Nvidia's generosity is not the same as openness. The only supported infrastructure provider is NICo — Nvidia's own controller. An operator who adopts NVKE is automating their fleet with Nvidia's tool, which works beautifully as long as the fleet is Nvidia's. That is a standards play dressed as a convenience, and it is precisely the kind of play that has worked in this industry for years.
And it is a preview. No auth, no upgrade path, no backup: those are not gaps you close on a roadmap, they are gaps you close when customers start paying. Anyone building on NVKE today is building on a statement of intent with a working compiler.
What to watch
Whether infrastructure providers beyond NICo are supported — that decides whether NVKE becomes a product or a platform; whether the management plane gains upgrade and backup procedures before someone puts it in production; whether authentication becomes the default rather than homework; and whether the large GPU renters adopt it or keep building their own tooling. GitHub stars are the least interesting indicator here. The number that matters is how many clusters are actually running on it in twelve months.

