NVIDIA has released the complete Run

platform as open source, the company said in a post on its developer blog on 13 October. Run
is the software that decides which AI jobs get which chips in a shared cluster, and how one chip can be split between several jobs. The release sits inside NVIDIA’s DSX AI Factory Platform, its umbrella for data center software.

One part of this is not new. The KAI Scheduler, the piece that places jobs on hardware, was already open source under the Apache 2.0 licence and already a CNCF Sandbox project. What opened this week is everything around it. NVIDIA says the platform comes as four components that can be adopted separately: the scheduler, a GPU Fractioning Operator, KAI Resource Management, which handles who gets how much capacity, and Workload Map (a library that describes the structure of AI jobs to the cluster).

The licence is where a prospective user should slow down. NVIDIA’s post never names a single licence for the platform. It says the source code is “governed by its applicable open source licenses,” which leaves a reader to check the repository component by component. The packaged version, which NVIDIA calls the Community Product and distributes through its NGC catalog, is a different matter: downloading it requires accepting NVIDIA’s Software License Agreement and product-specific terms. No paid subscription is needed, but the post states plainly that NVIDIA enterprise support is not included.

For a company weighing adoption, that is a materially different offer from a permissively licensed project with a community around it. You can read and build from the code. Getting a version NVIDIA stands behind still means going through NVIDIA’s commercial products, which the post says remain on sale through Mission Control, AI Enterprise and select cloud marketplaces, with terms that “vary by offering.”

The quieter news is a retirement. Run

’s fractional GPU sharing, which let several workloads use one chip with a hard memory limit each, was proprietary and, by NVIDIA’s own description, one of the platform’s most widely used features. That implementation has been dropped. The new operator relies on fractioning built into CUDA 13.4, where the driver enforces the memory cap, and it works with schedulers other than NVIDIA’s own.

That is a company retiring one of its better-liked differentiators in favour of a feature of the underlying platform. It suggests NVIDIA sees more value in making CUDA the place where sharing happens than in keeping a scheduling layer that only works its way. It also lowers the cost of leaving: a scheduler-agnostic operator is easier to carry to a competing orchestration tool.

Eight partners from NVIDIA’s DSX ecosystem are building it in. Crusoe is integrating it with its managed Kubernetes, and Nebius offers it as a one-click deployment. Rafay plans a fully managed service, and Saturn Cloud is using it for white-label inference services. Mirantis, OpenNebula Systems, Spectro Cloud and vCluster round out the list. Several are GPU clouds and platform vendors, the kind of business that gain most from a ready-made way to pack customers onto expensive chips.

The post also covers related projects, including Model Streamer, which loads model weights into GPU memory faster at startup. Its numbers come from the parties that ran the tests, not NVIDIA. WEKA reports 3.5 seconds, VAST Data reports 1.7 seconds, and DDN says it streamed a 65.5 GB model across eight GPUs in under four seconds, each on its own storage. None is an independent result.

The practical question for any team running its own GPUs is whether the code’s licences and NVIDIA’s download terms fit its legal review. That answer, not the word “open,” decides how much of Run

it can actually use.

Based on NVIDIA’s developer blog post by Ekin Karabulut and Sam Heywood, 13 October 2026.