NVIDIA says the Run:ai platform is now open source

On 13 October 2026, NVIDIA published NVIDIA Run:ai Is Now Open Source as Part of NVIDIA DSX AI Factory Platform. The page records a publish time of 14:19 UTC and a modification the same day at 21:46 UTC. The byline on the article is Ekin Karabulut and Sam Heywood. NVIDIA says it is making the full NVIDIA Run:ai platform available as open source software, as part of the NVIDIA DSX AI Factory Platform, for AI workload and GPU orchestration on Kubernetes.

The day before, at the Open Compute Project summit, NVIDIA described a Vera CPU rack for agent work beside GPU inference. That is a hardware design, covered in the 12 October account. This release is the orchestration software.

What is in the release

NVIDIA names the complete platform and four components that can also be adopted on their own: KAI Scheduler, GPU Fractioning Operator, KAI Resource Management, and Workload Map. The platform repository is dsx-ai-factory/run-ai.

KAI Scheduler is the Kubernetes scheduler NVIDIA says powers the platform. The blog says NVIDIA released it under the Apache 2.0 license and contributed it to the CNCF as a Sandbox project, and that scheduler development happens in that community repository. The features NVIDIA lists there include gang scheduling, fair-share and quota, topology-aware placement, GPU sharing (MIG, MPS, time sharing, HaMi, and CUDA 13.4 MPS fractioning), and Dynamic Resource Allocation. NVIDIA says it can run in an existing cluster without the rest of the platform.

GPU Fractioning Operator is the Kubernetes layer for CUDA-native GPU fractions. NVIDIA says its older proprietary fractioning implementation has been retired in favor of the memory limit in CUDA 13.4, where the driver enforces a hard GPU-memory limit and each workload sees only its fraction. The operator enables that on a cluster and exports per-pod metrics. NVIDIA says it is scheduler-agnostic and also integrates with KAI Scheduler. The blog links kai-scheduler/kai-gpu-fractioning, which on 14 October redirected to kai-scheduler/gpu-fractioning.

KAI Resource Management is the organizational layer: projects and departments, quota and fair-share across that hierarchy, and node pools for heterogeneous hardware. NVIDIA’s example is a platform team guaranteeing 50% of a cluster to research. That 50% is an illustration in the post, not a measured deployment.

Workload Map is the contract NVIDIA says the platform uses so schedulers and other tools can treat PyTorchJob, RayCluster, JobSet, KServe InferenceService, Dynamo graph deployments, and custom resources through one description. NVIDIA says it ships as a Go library that works offline on unstructured objects, and that it does not itself schedule, queue, autoscale, or propagate across clusters. The blog links dsx-ai-factory/karta, which on 14 October redirected to dsx-ai-factory/workload-map.

Licenses on the repositories

The blog itself names Apache 2.0 for KAI Scheduler. On 14 October the LICENSE file in each of these repositories was the Apache License, Version 2.0: run-ai, KAI-Scheduler, kai-gpu-fractioning (the URL the blog uses), kai-resource-management, and workload-map. The run-ai, gpu-fractioning, and resource-management files also carry a 2026 NVIDIA copyright line above the Apache text.

What NVIDIA says is still commercial

The post draws a line between the source code and supported products. It says the source is governed by its open source licenses. It also says the complete platform is on NVIDIA NGC as a Community Product, with a Helm chart, and that downloading that Community Product requires accepting the NVIDIA Software License Agreement and Product-Specific Terms for NVIDIA AI Products. No paid Run:ai subscription is required for the Community Product, and NVIDIA enterprise support is not included.

For NVIDIA Enterprise AI Factory deployments, the post says Run:ai continues through NVIDIA-supported commercial offerings: as a standalone product, through NVIDIA Mission Control and NVIDIA AI Enterprise, and on select cloud marketplaces. Availability, licensing, and support terms vary by offering. That split is NVIDIA’s statement.

NVIDIA also names eight companies it says are integrating the open source platform: Crusoe, Mirantis, Nebius, OpenNebula Systems, Rafay, Saturn Cloud, Spectro Cloud, and vCluster. Those integrations were not checked with each company. The post separately points at Model Streamer and Grove as related open source projects. Their licenses were not part of this read.

The practical split is the one in the post: the repositories above are public under Apache 2.0, and a supported NVIDIA deployment is still a commercial offer with its own terms.