NVIDIA Run:ai licensing, pricing and requirements: Kubernetes, GPUs, self-hosted or SaaS

How NVIDIA Run:ai is licensed in November 2026

NVIDIA Run:ai is a GPU orchestration platform for Kubernetes: it schedules AI workloads, allocates GPUs in fractions and manages quotas for departments and projects. NVIDIA's product page of 13 October 2026 describes it as open source, and the source repository is under the Apache License 2.0. NVIDIA also offers it as supported software through NVIDIA AI Enterprise, which is licensed per GPU, through NVIDIA Mission Control and NVIDIA DSX OS, and as standalone software. It publishes no price and no licence metric for Run:ai alone.

The current version, 2.26 of 12 July 2026, supports Kubernetes 1.34 to 1.36 and, on servers, any NVIDIA GPU that the NVIDIA GPU Operator supports. This guide follows NVIDIA's product page, its licensing guide and the Run:ai documentation as read on 12 November 2026. How a scheduler divides GPUs between teams, with a worked queue design, is in our comparison of KAI Scheduler, Kueue and Slurm; this page covers the licence and the requirements.

Three NVIDIA statements, by date

NVIDIA's pages describe the licence of Run:ai in three ways. In the order of their dates:

• The get-started page of NVIDIA AI Enterprise, dated 2 March 2026, lists Run:ai under "exclusive software that isn't open source, such as NVIDIA Run:ai, NVIDIA NIM, and NVIDIA vGPU™ for Compute."
• The licensing guide of 2 September 2026 describes "NVIDIA Run:AI, included with NVIDIA AI Enterprise", and the Quick Start Guide of the same date states: "The NVIDIA AI Enterprise license includes both NVIDIA Run:ai self-hosted and NVIDIA Run:ai SaaS deployments".
• The product page of 13 October 2026 opens with the sentence "NVIDIA Run:ai is the open source AI workload and GPU orchestration platform for Kubernetes."

The product page is the newest of the three. It also says that Run:ai "is available as supported software through NVIDIA DSX OS™, NVIDIA Mission Control™, and NVIDIA AI Enterprise". Which other NVIDIA components are open source or free of charge is listed in CUDA, TensorRT, NIM: free or NVIDIA AI Enterprise.

NVIDIA pages as named in the third column, read 12 November 2026.

Mission Control is cluster software for DGX B200 and DGX B300 clusters and for NVL72 rack systems; what it contains is in our article NVIDIA Mission Control: what DGX buyers get.

Run:ai pricing and what NVIDIA publishes

None of these pages prints a price for Run:ai, and none defines a licence metric for Run:ai as a product of its own. A sentence such as "Run:ai costs a fixed amount per GPU" has no NVIDIA source today. The published figure belongs to the suite: NVIDIA's pricing page for NVIDIA AI Enterprise, last updated on 2 September 2026, lists the subscription with Business Standard support at $4,500 per GPU for one year. That is NVIDIA's list price for the whole suite. It is neither a price for Run:ai nor our price. The terms are in our guide to NVIDIA AI Enterprise licensing.

Two details matter for some clusters. NVIDIA's licensing guide states that each H200 NVL includes a five-year NVIDIA AI Enterprise subscription, which has to be activated and may be installed only on the NVIDIA-Certified System in which the cards are installed. Whether that subscription covers Run:ai on a given cluster is a point to have confirmed in the quote. The second detail concerns evaluation: NVIDIA's note on the 90-day trial of NVIDIA AI Enterprise says the trial licence "does not include NVIDIA Run:ai".

We supply NVIDIA AI Enterprise licences with the GPUs and servers of a cluster and quote them in writing; send us the number of GPUs, the card models and the term you have in mind and the quote reaches you within one business day.

What Run:ai adds to a Kubernetes cluster

The documentation describes two parts. The NVIDIA Run:ai cluster "Provides scheduling and workload management, extending Kubernetes native capabilities." The control plane handles resource management and workload submission and "can be deployed on the cloud or on-premise". Underneath runs the KAI Scheduler, which NVIDIA calls "an open source Kubernetes scheduler designed for AI workloads and shared GPU infrastructure". According to the README of the repository, Run:ai includes the KAI Scheduler and adds, among other things, fair-share scheduling, preemption, fractional GPUs and quota hierarchies across departments and projects.

Two announced changes concern GPU sharing, and the documentation of 2.26 still announces both: "CUDA-based fractions will replace the legacy engine", and "GPU Memory SWAP will be deprecated and removed from future releases." A design that relies on swapping GPU memory between workloads should plan without it. What fractions, MIG and time-slicing each isolate is explained in GPU sharing in Kubernetes.

Self-hosted or SaaS control plane

NVIDIA names two deployment forms, "self-hosted (customer-managed control plane) or SaaS (NVIDIA-managed cloud service control plane)", each with its own collection on NGC. The difference is who operates the control plane. For the self-hosted form the documentation states these minimums:

• 10 CPU cores, 12 GB of memory and 110 GB of disk for the control plane;
• Kubernetes 1.34 to 1.36 or OpenShift 4.18 to 4.22, with Helm 3.14 or later and a default storage class;
• a fully qualified domain name, since an IP address is not accepted, and TLS;
• optionally an external PostgreSQL database, and then "only version 16".

An air-gapped installation is documented, which suits sites without a route to the internet; the rest of an offline setup is in our article on the air-gapped LLM server. Installation runs through Helm or through NVIDIA Base Command Manager, for which the documentation has configuration pages for DGX systems, the DGX B300 among them. These figures are stated minimums: the pages we read give no sizing for a highly available control plane.

Run:ai 2.26 requirements for Kubernetes, operators and nodes

NVIDIA Run:ai documentation (self-hosted), system requirements and support matrix, read 12 November 2026. The requirements page speaks of "the latest release", which is 2.26.

Run:ai 2.26 ends at Kubernetes 1.36, while the upstream project ended support for 1.34 on 27 October 2026 and maintains 1.35, 1.36 and 1.37 (as of 5 November 2026). Plan a new Run:ai cluster on 1.35 or 1.36. For the GPU Operator, Run:ai accepts 25.10 to 26.3, while NVIDIA's GPU Operator documentation of 14 October 2026 is at 26.7.x. Pin the versions from Run:ai's matrix. For the Network Operator the two Run:ai pages disagree, so confirm the upper bound before you install a version above 26.1.

The requirements page lists vanilla Kubernetes, OpenShift, RKE2 and the managed services EKS, GKE and OKE, among others, on x86 and ARM. The partner-validated platforms include VMware VKS, Mirantis k0rdent, Rafay and Spectro Cloud. Related guides of ours cover OpenShift AI on GPU servers and Kubernetes DRA for NVIDIA GPUs.

NVIDIA Run:ai documentation, product version life cycle and support matrix, read 12 November 2026.

NVIDIA's support policy for Run:ai, last updated on 22 March 2026, gives a major version nine months of full support and three further months of extended support; support ends twelve months after the release. For 2.26 the documentation gives 12 April 2027 as the end of full support and 12 July 2027 as the end of extended support. A cluster therefore needs at least one upgrade a year to stay supported. NVIDIA states no release rhythm; by our estimate, counted from the release dates, the four versions from 2.23 of 28 September 2025 to 2.26 are 112, 105 and 70 days apart. The infrastructure releases 8.2 and 7.8 of NVIDIA AI Enterprise both list Run:ai 2.26.

GPUs and servers Run:ai accepts

The hardware rule is one sentence in the requirements: "The NVIDIA Run:ai cluster supports x86 and ARM CPUs, and any NVIDIA GPUs supported by the NVIDIA GPU Operator." Of the cards we supply, the GPU Operator's platform list, release 26.7.x of 14 October 2026, names the RTX PRO 6000 Blackwell Server Edition with driver 575.57.08 or later, the RTX PRO 4500 Blackwell Server Edition, the L40S, the L4 and the H200 NVL. It does not name the workstation RTX PRO cards or the RTX 5000 Ada.

Workstations are not supported, so the worker nodes are servers. The same requirements page states: "vGPU is not supported. NVIDIA Run:ai currently supports GPU passthrough only." Worker nodes are therefore bare-metal servers or virtual machines with whole GPUs passed through. Under VMware that means DirectPath I/O, whose licence side is covered in NVIDIA AI Enterprise on vSphere.

Of the cards on that list, the NVIDIA RTX PRO 4500 Blackwell Server Edition is the passive single-slot card with 32 GB at 165 W (buy page, specifications); the RTX PRO 6000 Blackwell Server Edition and the H200 NVL have product pages of their own. We build GPU servers to order with these cards and supply Supermicro, Gigabyte, ASUS, Dell and HPE servers per project. Where Run:ai is to run under an NVIDIA AI Enterprise subscription, the server model should also be on the NVIDIA-Certified Systems list, because NVIDIA requires such a system for support of the suite; see NVIDIA-Certified Systems: what the label covers.

When a small cluster does without Run:ai

Run:ai is a platform built on a scheduler, and the scheduler can be used alone. The KAI Scheduler is an open-source project of its own, and Kueue adds quotas on top of the standard Kubernetes scheduler. Our article on GPU scheduling for ML teams designs queues for three teams on 16 GPUs with either of them. Our recommendation by situation:

• Two to four GPU servers and a few teams whose quotas one platform team keeps as manifests: the scheduler alone is enough, and nothing has to be licensed from NVIDIA.
• Several departments with projects of their own, self-service submission of workloads and quota hierarchies: this is what the Run:ai control plane adds, open source or supported.
• A requirement for vendor support on the orchestration layer: NVIDIA AI Enterprise per GPU. NVIDIA's licensing page for Base Command Manager of 7 October 2026 describes the suite as including support for the GPU Operator and the Network Operator as well.
• Nodes that are workstations, or virtual machines with vGPU: Run:ai's requirements exclude both.

A pilot on one server for one team gets by without an orchestration platform; when a shared cluster begins is the subject of pilot to production: scaling private AI hardware. Quotas with chargeback per department are covered in GPU chargeback and quotas for a shared GPU platform.

What to send us for a Run:ai quote

With support, Run:ai is quoted as NVIDIA AI Enterprise on the GPUs of the cluster, or through one of the other routes in the table above. For the quote we need:

1. the number of GPU nodes, the card model and the number of cards per node;
2. whether the cards already carry an NVIDIA AI Enterprise subscription, as each H200 NVL does;
3. the Kubernetes distribution and version, checked against the range 1.34 to 1.36;
4. self-hosted or SaaS control plane, and whether the cluster is air-gapped;
5. the subscription term you plan.

Our software page lists Run:ai, together with NVIDIA AI Enterprise and vGPU, among the licences we supply. Write to us with these five points, or by e-mail to request@eurokommerz.at; the written quote follows within one business day.

FAQ