NVIDIA Open-Sources NVKE and NVKR: Running Managed K8s on Bare-Metal GPUs

NVIDIA open-sources NVKE and NVKR under Apache 2.0, enabling managed K8s on your own bare-metal GPUs.
NVIDIA open-sources NVKE and NVKR under Apache 2.0, both in public preview. The combined solution lets AI factories run a managed Kubernetes service on their own bare-metal GPU servers, moving away from reliance on public clouds. NVKE provides the engine layer for cluster lifecycle and control plane operations, while NVKR offers reusable recipes for assembling GPU components. Production adoption still depends on how well the surrounding ecosystem keeps pace.
NVIDIA has open-sourced NVIDIA Kubernetes Engine (NVKE) and NVIDIA Kubernetes Recipes (NVKR) under the Apache 2.0 license. Together, the two projects let so-called "AI factories" run a managed Kubernetes service on their own bare-metal GPU servers. Both are currently in public preview.

What this actually changes
Managed Kubernetes has long been the cloud providers' business. Running K8s without the operational burden usually means handing your workloads to a public cloud and accepting its control plane, its billing model, and its networking model. The NVKE and NVKR combination takes a different approach: the control plane comes from NVIDIA's software stack, while the nodes remain your own bare-metal GPU machines.
This is directly appealing to teams with substantial GPU assets. Large-scale GPU clusters are expensive to procure, and idling them is pure waste, while renting GPU instances from a public cloud often carries a premium. If you can get a near-managed Kubernetes experience on your own hardware, the burden of scheduling, upgrades, and operations drops noticeably.
The control plane and data plane of managed Kubernetes: A Kubernetes cluster is typically divided into a control plane and a data plane. The control plane includes the API Server, scheduler, controller manager, and etcd, and is responsible for receiving requests, scheduling Pods, and maintaining cluster state; the data plane consists of the nodes running the actual workloads. Cloud providers' managed services (such as EKS, GKE, and AKS) essentially operate the control plane as a service, leaving users responsible only for nodes. NVKE aims to replace precisely this part—turning the control plane into software delivered for users to deploy themselves, while retaining bare-metal GPU nodes as the data plane. Understanding this layering is what makes it clear that a "managed experience" and "your own hardware" are not contradictory: what is managed is operational responsibility, not physical location.
Bare metal versus virtualization trade-offs: A bare-metal server is a physical machine that runs the operating system directly, without a hypervisor virtualization layer. For GPU workloads, bare metal avoids the performance overhead of virtualization and makes GPU passthrough, RDMA networking, and specific driver stack configurations more direct and controllable—which is why AI factories tend to prefer bare metal. The cost is losing the on-demand elasticity and rapid provisioning of the cloud, and having to manage node lifecycles yourself—which is exactly the part NVKE attempts to fill in.
What NVKE and NVKR each handle
The division of labor is evident from the names. NVKE targets the engine layer—the managed Kubernetes service itself—distilling cluster lifecycle management and control plane operations, normally handled by cloud providers, into deployable software. NVKR stands for "Recipes," offering reusable recipes that help users assemble GPU-related components, drivers, and runtimes in the recommended way.
The source article does not disclose finer module breakdowns or architectural details, but the fact that the two are open-sourced together says something about their positioning: one is the foundation, and the other is the operating manual that turns that foundation into a usable GPU platform. For teams that do not want to assemble a GPU scheduling stack from scratch, the value of Recipes may be more immediate than the engine itself.
Typical components of a GPU scheduling stack: Running GPU workloads on Kubernetes is not just a matter of installing drivers. It usually also requires NVIDIA GPU Operator to centrally manage drivers, GPU access for the container runtime (such as containerd), the Device Plugin (which exposes GPU resources to kubelet), DCGM metrics collection, and GPU Feature Discovery to label nodes with model and capability information. In addition, multi-GPU training involves NCCL communication libraries, GPUDirect Storage/RDMA, and other components. These components have complex version compatibility relationships and a specific assembly order—the value of NVKR's "recipes" lies precisely in codifying these recommended combinations into repeatable processes, reducing the time teams spend trial-and-error on the version matrix.
The signal sent by Apache 2.0
Choosing Apache 2.0 rather than a more conservative license means commercial use, modification, and redistribution carry no additional restrictions. For infrastructure software, this is not a casual choice.
On one hand, NVIDIA needs ecosystem partners—systems integrators, managed service providers, private cloud vendors—willing to embed NVKE in their own products, and a permissive license is a prerequisite. On the other hand, it lowers the compliance review cost for enterprises adopting it internally. GPU infrastructure decisions already have long cycles, so the less license friction the better.
The difference between Apache 2.0 and copyleft licenses: Apache 2.0 is a permissive license that allows users to freely use, modify, and distribute code, including embedding it in closed-source commercial products, as long as copyright and license notices are retained, with patent grant provisions attached. By contrast, copyleft licenses such as GPL require derivative works to be open-sourced under the same license. For infrastructure software, a permissive license lets systems integrators and private cloud vendors confidently package NVKE into their own commercial distributions without having to disclose their own product source code—a key prerequisite for building an ecosystem of distribution.
What the public preview stage means
Both projects are labeled public preview. This usually means interfaces and components may still change, and business-critical production workloads require careful evaluation. The source article does not provide a timeline for general availability.
For teams currently planning their own GPU clusters, a reasonable approach at this stage is to treat it as a technology validation target: first run through the NVKR recipe workflow on a non-critical cluster, observe how the control plane performs in your own network and storage environment, and then decide whether to expand. Feedback during the public preview often shapes the design of later versions as well.
Practical implications for AI infrastructure builders
Over the past few years, competition in GPU compute has focused on chip supply, but what really determines cluster utilization is often the software layer above. A mature managed Kubernetes lets the same hardware deliver higher effective throughput and lets teams shift their energy from operations to models and business.
The open-sourcing of NVKE and NVKR turns capabilities at this layer from proprietary cloud provider products into open-source components that can be inspected and modified. It may not immediately change the market landscape, but for organizations whose GPU assets have grown large enough that scheduling efficiency must be taken seriously, it is a new path.
It is worth noting that the completeness of the managed experience ultimately depends on the surrounding ecosystem—monitoring, storage, network plugins, and multi-tenant isolation. Whether these keep pace still needs to be tested in actual deployments. Open source is only the starting point.
Reference source: MarkTechPost report "NVIDIA Open-Sources NVKE and NVKR for Managed Kubernetes on GPU Infrastructure."
One-line Summary
NVIDIA open-sources NVKE and NVKR under Apache 2.0, enabling managed K8s on your own bare-metal GPUs.
Paragraph Summary
NVIDIA has open-sourced NVIDIA Kubernetes Engine (NVKE) and NVIDIA Kubernetes Recipes (NVKR) under the Apache 2.0 license, both in public preview. The combined solution lets "AI factories" run a managed Kubernetes service on their own bare-metal GPU servers, changing the previous paradigm where managed K8s had to rely on public clouds. NVKE provides the engine layer, turning cluster lifecycle and control plane operations into software; NVKR provides reusable recipes that assemble GPU drivers, runtimes, and other components in the recommended way. The permissive license lowers the integration barrier for ecosystem partners and the internal compliance threshold for enterprises. Public preview means interfaces may still change, making it suitable for technology validation at this stage; production adoption will depend on whether the supporting ecosystem—monitoring, storage, networking—keeps pace.
Related articles

Gemini 4 Neon Tested in Blender: Cheap but Not the King
Bilibili creator tests Gemini 4 Neon in Blender across five rounds vs Claude Fable, Opus, and GPT. Neon's detail nears Fable but animation still lags; at $44 vs Fable's $52, cheaper Argon under $14 is the value surprise.

Gemini 4 Neon Hands-On: Can It Challenge Opus 5.5?
Google released 4 Gemini 4 models in 49 days. We test whether Neon can challenge Opus 5.5 across benchmarks, real video tasks, and subscription costs.

Former Inmate Reveals Psychopath Ratio in Prison: About 200 Out of Every 880
Former inmate Larry Lawton shares on Lex Fridman's podcast that about 200 of 880 life-sentence inmates are psychopaths—a 23% ratio. He explores the blurred line between good and evil, social dynamics in prison, and wisdom from dark experiences.