Blog
NVIDIA open-sources the full Run:ai platform for sharing GPUs across AI workloads
On 13 October 2026 NVIDIA released the complete Run:ai platform as open source, along with four parts that can be used on their own, including the KAI Scheduler and a GPU Fractioning Operator. Eight partners are building it into their clouds. The free version comes without NVIDIA enterprise support.
On 13 October 2026 NVIDIA released the full Run:ai platform as open source software. Run:ai is software that decides how AI jobs share a company's GPUs, the expensive chips used to train and run AI models. Until now, the complete platform was a paid NVIDIA product.
NVIDIA says the release is part of its DSX AI Factory platform, its blueprint for large AI data centres, and is meant to let cloud providers and partners build their own products on top of it.
What Run:ai does
Run:ai was founded in 2018 as an Israeli start-up. NVIDIA bought it in 2024. Its software runs on Kubernetes, the widely used system for managing applications across many servers, and hands out GPU capacity to AI work such as data processing, training and inference.
Instead of letting teams and applications fight over GPUs, the platform can assign capacity on the fly. It can also split a single GPU between several smaller jobs that do not need a whole chip. Calcalist's CTech reports that the founders, Omri Geller and Ronen Dar, said at the time of the acquisition that they planned to open-source the software.
What is in the release
NVIDIA has published the complete platform plus four parts that can also be used on their own:
- KAI Scheduler: the engine that decides which jobs run on which GPUs. It was already open source under the Apache 2.0 licence and is a CNCF Sandbox project, an early-stage project at the foundation that hosts Kubernetes. It can place multi-part jobs all at once or not at all, share capacity fairly between teams, and take server layout into account.
- GPU Fractioning Operator: lets several workloads share one GPU, each with a hard memory limit. NVIDIA says it has retired its older proprietary method in favour of fractioning built into CUDA 13.4, its GPU software layer, and this operator makes that feature usable on Kubernetes.
- KAI Resource Management: maps GPU capacity to projects, departments and quotas. NVIDIA gives the example of guaranteeing 50% of a cluster to a research team while stopping any other team from taking all the rest.
- Workload Map: a common description of AI workload types, so tools can schedule and monitor them without custom code for each one.
The code is on GitHub, and the full platform is also offered on NVIDIA's NGC catalogue as a "Community Product" with an installer. NVIDIA says no paid Run:ai subscription is needed for that version, but enterprise support is not included. Downloading it from NGC still requires accepting NVIDIA's software licence terms.
Partners and paid options
NVIDIA names eight partners building the open-source platform into their own offerings: Crusoe, Mirantis, Nebius, OpenNebula Systems, Rafay, Saturn Cloud, Spectro Cloud and vCluster. Nebius, for example, offers it as a one-click install in its AI cloud.
Run:ai is still sold as a supported commercial product, including through NVIDIA Mission Control and NVIDIA AI Enterprise. NVIDIA says availability, licensing and support terms vary by offering.
NVIDIA also pointed to two related open-source projects: Model Streamer, which speeds up loading model weights onto GPUs, and Grove, for running inference systems made of several connected parts.
Why it matters
GPUs are expensive, and idle chips waste money. Software that keeps them busy and shares them fairly is becoming a key part of AI infrastructure. By open-sourcing Run:ai, NVIDIA makes that layer free for any cloud or company to inspect, adapt and run.
The move could also extend NVIDIA's reach beyond chips, as more AI clouds build on its orchestration software and wider DSX platform. Claims about better GPU use and throughput are NVIDIA's own and will depend on how each operator deploys the software.