NVIDIA Open-Sourced the Software That Runs AI Factories. Four Gaps It Admits Itself
Synergised Consulting Ltd
Last Updated: 8 November 2026
NVIDIA has published NVKE, the software layer that turns a building full of GPUs into a multi-tenant AI factory, as open source under the Apache 2.0 licence. Each customer of an AI factory running it gets their own Kubernetes cluster with a separate control plane, and machines must cryptographically prove their identity before they can even join. Just as useful for anyone buying from these operators: NVIDIA publishes the things the platform cannot yet do, in its own words, and that habit of admitting limits is the most transferable idea in the release.
What NVKE Actually Is
NVKE, short for NVIDIA Kubernetes Engine, is a management system for running Kubernetes clusters on NVIDIA GPU infrastructure. According to NVIDIA's own documentation [1], an AI factory operator installs it and then creates clusters for customers through an API: NVKE provisions the bare-metal nodes, forms each customer's cluster, and drives the ongoing work of scaling node pools, upgrading Kubernetes, replacing failed nodes and renewing certificates. Every customer gets a workload cluster with its own control plane, its own etcd database and its own node pools, which is what stops one tenant's workload from ever sitting on the same management layer as another's.
Two design details are worth a business reader's attention. First, node identity: before a machine receives its credentials to join a cluster, it attests through an identity framework called SPIFFE/SPIRE, using a TPM security chip on the motherboard itself. A machine that cannot prove what it is cannot join. Second, the build chain: a companion project called NVKR builds every software artefact from pinned, recorded sources, produces a software bill of materials, and signs its releases, so an operator can check where every running component came from before it touches a cluster [1].
NVIDIA is also explicit about what NVKE is not: not a Kubernetes distribution, not a re-badged product, but a self-hosting factory for clusters. The company announced the surrounding DSX platform, the wider AI factory playbook that DSX OS belongs to, on 31 May 2026, and claims that up to 40% more GPUs can run at peak efficiency within a fixed power budget [2].
Where It Sits in the Stack
NVKE is one layer of DSX OS, NVIDIA's open-source AI factory software stack announced at the end of May. The newsroom framing at launch was that the software, proven running NVIDIA's own operations, would be released so ecosystem partners could deliver AI services without months of custom development [2]. NVKE is the piece of that stack that manages the clusters everything else runs on.
Why an Owner Should Care About Someone Else's Data-Centre Software
The direct answer: because this release shows, in public, what genuine multi-tenant isolation looks like when a vendor documents it properly, and that is exactly the standard to hold any AI platform vendor to during diligence. If you are buying AI capacity, an AI-powered service, or agent tooling, the questions NVKE answers in public are the questions to ask of whoever runs your workloads.
In discovery workshops, owners rarely ask about isolation at all. The conversation is about capability and cost. But the pattern we see is that the questions that catch vendors out are structural, not feature questions: who holds admin credentials, what proves a machine is what it says it is, who signs off before software reaches the cluster. NVKE's design answers those three in its own documentation, which makes it a usable checklist even for a buyer who will never run a GPU.
The same logic runs through the governed agent work: separate identity per workload, credentials scoped to a role, an approval point before anything consequential runs. NVKE simply does this at data-centre scale, and the principle is identical at a five-person business.
The Four Gaps NVIDIA Admits
This is the part worth copying. NVKE is a preview, and NVIDIA says plainly it does not recommend it for production workloads. Its documentation then lists, in its own words, what the system cannot yet do [3]:
1. The management cluster does not manage itself. It cannot upgrade, scale or replace its own nodes. 2. There is no documented procedure to upgrade the management plane to a later NVKE release. 3. There is no backup or restore procedure, and the encryption key for the management database exists in exactly one off-node copy. 4. API authentication is off by default, the API serves unencrypted traffic on port 30080, and NVKE performs no authorisation at all: every caller the network admits can call every endpoint, including the ones that create and delete customer clusters. A handed-over admin credential also cannot be revoked.
Independent coverage reached the same verdict. According to heise online, the preview is aimed at GPU data-centre operators, the code is Apache 2.0, and NVIDIA has announced no date for a production-ready release [4].
Reading a Limitations Page Correctly
A list like this is not a reason to dismiss the platform. It is the evidence that the vendor distinguishes between what works and what is unfinished, which is precisely what a buyer cannot see from a marketing page. The vendors that worry about are the ones whose documentation has no gaps section at all. When you evaluate any platform, ask for the equivalent list. If the vendor cannot produce one, that is the finding.
What a Buyer Can Take From This
For a founder buying AI infrastructure or agent services, this release converts into three diligence questions you can ask any vendor this week. What proves a machine or workload is what it claims before it gets access? Who builds your software artefacts, and can you show where each one came from? And where is your list of what the platform cannot do yet, in your own words? A vendor with real answers, and a real gaps list, is the one whose isolation claims you can actually trust.
Sources
- [1] NVIDIA, "What is NVKE", NVKE documentation:
- [2] NVIDIA Newsroom, "NVIDIA DSX Gives Infrastructure Builders the Playbook for AI Factories", 31 May 2026:
- [3] NVIDIA, "Preview status and known limitations", NVKE documentation:
- [4] heise online, "NVKE: Nvidia previews its own Kubernetes engine", 7 November 2026: