NVIDIA opens two free tools to run GPU-powered cloud computing for AI
NVIDIA released its cluster-management software for free so companies can let customers rent AI-ready computers on their own graphics cards.

What you could do with it
- A company that owns a fleet of NVIDIA GPU servers could launch a managed Kubernetes service for its customers without writing cluster-management software from scratch 1.
- An AI platform team could provision each customer with an independent Kubernetes cluster—its own control plane, etcd database, and worker nodes—through a REST API and kubectl 2.
- An operator could pin exact software versions to specific GPU generations in declarative recipes, then build signed container images, OS images, Helm charts, and SBOMs automatically 1.
What the tools do
For years, if you owned a rack of NVIDIA graphics servers and wanted to sell time on them to several AI teams — or to your own product group — you had to assemble that management layer yourself from pieces of open-source software, commercial products, and custom code, then keep it working through every new version of every component.
NVKE supplies the control surface for that job: an application programming interface and background controllers installed on a small management cluster that accept requests like 'create a customer cluster,' 'add ten worker nodes,' 'upgrade Kubernetes,' or 'replace a broken node.' Customers connect through the standard kubectl command-line tool and get ordinary clusters with their own isolated control planes 2.
NVKR is the companion build factory. It takes declarative lists — called Definitions — that pin exact software versions to exact GPU generations, compiles them into executable packages, container images, Linux images for bare-metal nodes, and Helm configuration files, and attaches cryptographic signatures, software bills of materials, and provenance records to every artifact 1.
Both are licensed under the Apache 2.0 open-source license and were announced as components of the broader NVIDIA DSX AI Factory Platform, a package aimed at organizations that operate GPU fleets as internal or external cloud providers 1.
How it is put together
An operator starts by bootstrapping a temporary single-node test cluster, then NVIDIA Infra Controller provisions three CPU servers for the management plane and boots them from an image built by NVKR 12. A third project, CAPNICo, bridges the infra controller to Kubernetes's own Cluster API framework so the management plane can talk to bare-metal servers .
When a customer asks for a cluster, NVKE sets aside its own control plane, etcd database, and set of worker nodes — all running upstream Kubernetes plus kubeadm — and installs OVN-Kubernetes for networking and NodeLocal DNSCache for fast name resolution 1. Each node must prove its identity using SPIFFE and a Trusted Platform Module before it receives join credentials; per-cluster encryption for the etcd data store is available 2. An OpenTelemetry template ships logs, metrics, and traces out over the OTLP protocol so operators can monitor everything from one place .
Inside the workload clusters, three companion projects fill the gaps that vanilla Kubernetes leaves for AI: AI Cluster Runtime handles GPU scheduling and isolation, NodeWright manages node-level packages, and NVSentinel provides break-fix automation 1.
Who is already using it
Four cloud partners — Firmus, SK Telecom, Mirantis, and Spectro Cloud, among others — are integrating NVKE or adding components to their managed Kubernetes products; Netris is enabling NVKE for network automation 1. The CNCF has certified that NVKE creates conformant clusters tested against Kubernetes version 1.36, earning both a Certified Kubernetes Installer badge and a Certified Kubernetes AI Platform badge .
A local evaluation runs on a single Linux host or Mac VM with just six CPUs, twelve gigabytes of memory, and one hundred gigabytes of disk — no GPU, no NVIDIA account required 12. Production installations need NICo-enrolled capacity, a KVM-enabled build host, and a container registry .
The catch
The NVKE documentation carries an explicit warning: the project is in public preview, APIs can change between releases, and NVIDIA does not recommend it for production workloads 23. Beyond that general notice, the MarkTechPost summary points out two concrete gaps — after installation the management API serves unencrypted HTTP on port 30080 with authentication disabled by default, and there is currently no documented upgrade path or backup procedure for the management plane itself 1. Both issues are consistent with a preview release, but anyone considering real deployments should plan to patch them before moving past testing.
This story reports on a trend for information only. It is not advice or encouragement to deploy any specific infrastructure.
How good is it
No throughput or provisioning-time benchmarks are published; the source says the quality signal is certification, not speed 1. NVKE is conformance-tested on Kubernetes 1.36 under the Certified Kubernetes AI Platform and Certified Kubernetes Installer programs . For comparison, Gardener (SAP) is self-described as production-ready and certifies up to v1.35, while NVKE itself is labeled preview and not recommended for production workloads because its default API runs over plain HTTP on port 30080 with authentication off and it lacks upgrade or backup procedures .
Key facts
- Two free NVIDIA tools give operators a ready-made management plane and build pipeline for GPU Kubernetes clusters
- Conformance tested against Kubernetes 1.36 with CNCF certification badges earned
- Preview release: authentication is off by default on port 30080 and no production upgrade path yet exists
- Local evaluation requires 6 CPUs, 12 GB RAM, 100 GB disk with no GPU needed
Sources
If this story concerns you or your company, or uses something that belongs to you, ask for a correction or for it to be taken down. How requests work
Keep reading
Apps and productsDyania's Synapsis AI reads 16 years of patient charts on hospital servers at Cleveland ClinicHerald of AI
Dyania's Synapsis AI reads 16 years of patient charts on hospital servers at Cleveland Clinic
An AI tool that can search through years of handwritten doctor notes inside hospital walls — doctors still decide everything.
Apps and productsNVIDIA Opens Source Run ai for Kubernetes GPU ManagementHerald of AI
NVIDIA Opens Source Run ai for Kubernetes GPU Management
The move gives developers free access to enterprise-grade GPU orchestration tools built for AI workloads.
Apps and productsHow Meta built TLX to help regular engineers squeeze more speed out of AI hardwareHerald of AIFigure: Guan et al., arXiv:2605.10905, CC BY 4.0
How Meta built TLX to help regular engineers squeeze more speed out of AI hardware
A new tool lets everyday AI developers write GPU code that matches specialist-level performance, powering the systems behind Facebook, Instagram and other services.