Skip to main content
Back to News
analysis/AI Infrastructure

NVIDIA Open-Sources the Missing Layer of the AI Factory

NVIDIA's open-source NVKE gives platform teams a management plane for GPU Kubernetes on bare metal, with per-tenant control planes and TPM attestation.

Stefan Trbojevic

Stefan Trbojevic

6 November 20266 min read
LinkedIn
Abstract render of layered glass server slabs joined by emerald light threads, evoking a multi-tier AI cluster factory

The takeaway

NVKE standardises the cluster factory layer above GPU hardware: per-tenant control planes, TPM-attested joins and signed node images. The preview is a reference architecture, not a resellable service, and the missing pieces (auth, backups, CA rotation) define the 2027 roadmap.

Why it matters for builders

If you rent GPU Kubernetes, hold providers to per-tenant control planes, hardware-attested node joins, per-cluster etcd encryption and signed node images. If you run your own metal, NVKE can be evaluated locally on a single Linux box with mock NICo before any GPU is bought. And the open gap on both NVKE and Google Substrate is the same: stateful, GPU-attached agent runtimes that survive a suspend, a node replacement and a tenant migration.

NVIDIA Open-Sources the Missing Layer of the AI Factory

NVIDIA has been quietly shipping releases of a project that answers a question most AI infrastructure teams have been asking for two years: if you own the GPUs, who runs the Kubernetes layer that turns them into a product? The answer, as of this week, is NVIDIA itself. The NVIDIA Kubernetes Engine (NVKE) is a preview-grade, Apache-2.0 management plane for GPU Kubernetes on bare metal, published under the company's dsx-ai-factory GitHub organization. The project shipped v0.23.1 on November 3, is already publishing release candidates for v0.24.0, and surfaced publicly on November 6 via AI Weekly's infrastructure tracker.

It is not a distribution. It is the layer above the machine.

What NVIDIA actually shipped

The repository's opening line is the whole pitch: "Run a managed Kubernetes service on your own NVIDIA GPU infrastructure." NVKE provides the management plane that a platform team installs to create and operate clusters for its own internal or external customers. Every workload cluster gets its own Kubernetes control plane, its own etcd, and its own node pools. Tenants reach those clusters through the ordinary Kubernetes API and kubectl, not through a proprietary console.

The bootstrap pattern is deliberate: a throwaway seed cluster stands up a management cluster, which then provisions and lifecycle-manages workload clusters on demand. NVIDIA is explicit that NVKE is "not a Kubernetes distribution and not a Cluster API flavour," but a self-hosting factory for clusters. Machines are provisioned through NVIDIA Infra Controller (NICo), the only first-class infrastructure provider in preview. Networking defaults to OVN-Kubernetes; telemetry runs through a pipeline called NKTP.

The identity model is the part worth reading twice. Nodes attest through SPIFFE/SPIRE using their TPM before they receive join credentials, and etcd encryption is available per cluster. That is hardware-rooted admission, not a bootstrap token with a ten minute window.

Abstract render of layered glass server slabs connected by glowing fiber lines, representing a three-tier cluster factory

Build evidence is now part of the product

The second half of the design is a companion project, NVKR (NVIDIA Kubernetes Recipes), which builds node OS images and NVKE components from pinned inputs. The division of responsibility is stated plainly: NVKR owns build and release evidence, NVKE owns deployment of those artifacts.

This is supply chain discipline applied to node images. Anyone who has operated bare metal knows the failure mode: an image is built on someone's laptop, the checksum is lost, and six months later nobody can explain what is inside the kernel on node forty-one. Making the build system a separate, signed, pinned-input project turns the operating system of the cluster into a versioned artifact rather than a snowflake.

Practically, the operator owns the whole stack, which is both the appeal and the liability.

Why this is a control-plane fight, not a Kubernetes fight

Nobody needed another way to install Kubernetes. The scarce thing in 2026 is the ability to hand an isolated, attested, billable slice of GPU capacity to a tenant without rebuilding the operational model each time.

That is the same battle being fought one layer up. Google moved its open-source agent stack, Agent Substrate and gVisor, under CNCF governance, as we covered on November 5, and is shipping bare-metal machine series pitched explicitly at agentic sandboxes. Neoclouds sell managed GPU Kubernetes as their entire business. NVKE is NVIDIA's move to make the cluster factory a commodity that runs on hardware NVIDIA profits from, rather than a differentiator owned by a cloud or a neocloud.

Follow the money and the pattern is consistent: whoever owns the control plane owns the margin, the trust boundary and the migration cost. NVIDIA would rather that control plane be open and installed on its own silicon.

Abstract render of dark server blocks emerging from a glowing gate on an assembly line

Where the preview genuinely falls short

NVIDIA's own limitations page is refreshingly blunt, and builders should read it before planning anything:

  • The API serves plain HTTP on port 30080 with authentication off by default, and NVKE does no authorization at all: any admitted caller can create and delete workload clusters.
  • The management cluster does not manage itself: no self-upgrade or node replacement, no management-plane upgrade procedure, no backup or restore procedure.
  • Nodes only join if you supply your own SPIRE certificate authority at bootstrap. Without it, the self-signed CA means no node can attest and nothing joins.
  • A workload cluster's admin kubeconfig cannot be revoked, because NVKE does not rotate the cluster certificate authority.
  • Node labels and taints are fixed at creation. nvkectl upgrades the control plane version, but node pool Kubernetes versions require a raw API call.
  • The full-disk-encryption node image generates one volume key per build and leaves a publicly documented bootstrap passphrase in the GRUB config, so it provides no data-at-rest confidentiality.
  • The default telemetry backend keeps data in memory and gives anonymous users admin access to Grafana.

Read that list as a maturity statement. NVKE today is a runnable reference architecture for how a GPU cluster factory should be shaped, not a service anyone should resell on Monday. The security posture in particular assumes an air-gapped evaluation network and a trusted operator, which is the opposite of multi-tenancy.

Abstract render of a central dark chip with emerald light threads radiating to smaller chips, representing hardware-attested node identity

What this means if you run agents on GPUs

Three practical consequences for teams building agent platforms.

First, if you rent GPU Kubernetes, NVKE hands you a concrete checklist to hold your provider to: per-tenant control planes, hardware-attested node joins, per-cluster etcd encryption, signed node images with pinned build inputs. If a provider cannot answer those four questions, you are not buying isolation, you are buying a namespace.

Second, if you run your own metal, you can now evaluate the shape without buying anything. The local quickstart stands up a management plane and a workload cluster on simulated machines with no NVIDIA account. Six CPUs, twelve gigabytes of RAM, a hundred gigabytes of disk, Docker and a virtual TPM are enough to test whether your deployment tooling survives a real cluster boundary.

Third, the gaps that remain are exactly where agent workloads live. Google's Substrate documentation is candid that gVisor cannot snapshot live CUDA contexts and that open connections, including sessions to MCP servers, are not preserved when an agent suspends. NVKE's preview has no story for GPU device passthrough into tenant workloads at all beyond the node. Between the two, the hard problem of 2027 is now visible: stateful, GPU-attached, long-lived agent runtimes that survive a suspend, a node replacement and a tenant migration.

What to watch

The v0.24 line is where the answers will appear. Authentication on by default, backup and restore, and certificate authority rotation are the three features that would move NVKE from interesting to adoptable. After that, watch whether any cloud or neocloud adopts the NVKE shape instead of building its own control plane behind a managed service.

The important shift is not that NVIDIA open-sourced one more tool. It is that the company that sells the accelerators has decided the management layer above them should be free, installable, and shaped by NVIDIA. That is a statement about where the value is expected to sit in the next phase of the buildout.

Share𝕏

The Automation Brief

Read 5 AI stories instead of 50.

The essential moves in AI agents, models, automation and infrastructure — filtered for builders and operators, with the part that actually matters.

No noise. Unsubscribe anytime.

Editorial notes

Reported by

Stefan Trbojevic

Edited by

n8n Lab Editorial

Published

6 November 2026

Updated

6 November 2026

AI disclosure: AI assisted with research and drafting. Factual claims are reviewed by an editor.

n8n Lab is an independent service provider. We are not affiliated with, endorsed by, or sponsored by n8n GmbH. “n8n” is a trademark of n8n GmbH and is used here only to describe the platform-specific implementation and automation services we provide.

    NVIDIA Open-Sources the Missing Layer of the AI Factory | n8n Lab News