NVIDIA open-sources full Run:ai platform for sharing and scheduling GPUs

Teams managing shared AI computing can now inspect, modify and build on the full software platform NVIDIA uses to schedule workloads and allocate GPUs. On October 13, 2026, NVIDIA released the complete Run:ai platform as open source within its DSX AI Factory Platform, offering both an integrated stack and components teams can adopt separately.

Run:ai manages AI workloads on Kubernetes across data processing, model training and inference—the work of generating model responses. It dynamically allocates GPU resources rather than leaving every workload with a fixed assignment. NVIDIA says opening the platform will help customers and partners improve GPU utilization and workload throughput; those are the company’s stated benefits, not measured results from this release.

Four pieces, without an all-or-nothing adoption

The release combines additional platform code with four independently usable components. A team can install the complete system, build a product around it, or fill a specific gap in its existing infrastructure. The components divide responsibility between placing jobs, sharing GPU memory, assigning capacity to teams and understanding different workload structures.

• KAI Scheduler decides where GPU jobs run. It balances resource shares and quotas, considers hardware layout, and places related groups of tasks together on an all-or-nothing basis. Already open source under Apache 2.0, it remains Run:ai’s scheduling engine, with all scheduler development taking place in the community repository. It can run without the full platform.
• GPU Fractioning Operator lets suitable workloads share a GPU through CUDA 13.4’s native fractioning capability. NVIDIA has retired Run:ai’s proprietary implementation in favor of this approach. The CUDA driver enforces each workload’s GPU-memory limit, while the operator adds Kubernetes integration and per-task usage metrics. It works with KAI Scheduler and other schedulers.
• KAI Resource Management governs who gets capacity. It maps resources to projects and departments, supports guaranteed allocations and sharing, and groups different hardware into node pools. That lets platform teams manage a shared cluster around organizational commitments rather than scheduling decisions alone. The component can be combined with KAI Scheduler.
• Workload Map gives tools a consistent description of different Kubernetes workloads, including their components, roles and status. Instead of each tool building its own integration for every workload type, that structure is defined once. It ships as a Go library that works without a cluster installation; it does not schedule, queue or automatically scale workloads.

Code access and download terms are distinct

The complete source is available in the Run:ai repository in NVIDIA’s DSX AI Factory GitHub organization. NVIDIA also distributes the integrated platform through its NGC catalog as a Community Product, with a Helm chart for installation. Teams can inspect the code, contribute changes or build on it directly.

The source code is governed by its applicable open-source licenses. Downloading the NGC package separately requires accepting NVIDIA’s applicable software agreement and AI-product terms. Commercially supported Run:ai offerings remain available, including the standalone product, NVIDIA Mission Control, NVIDIA AI Enterprise and select cloud marketplaces. Availability, licensing and support vary by offering.

Partners keep the infrastructure underneath

NVIDIA names eight partners integrating Run:ai into their offerings: Crusoe, Mirantis, Nebius, OpenNebula Systems, Rafay, Saturn Cloud, Spectro Cloud and vCluster. Their deployment approaches differ. Nebius offers a one-click application deployment, Rafay is expanding its integration into a fully managed service, and vCluster supplies pre-validated templates for separate tenant environments.

Crusoe’s October 13 announcement describes an existing integration, not a newly created deployment path. Run:ai controls scheduling and GPU sharing while Crusoe Managed Kubernetes operates the underlying cluster and GPU infrastructure. Crusoe says it and NVIDIA validated cluster registration, distributed multi-node PyTorch training, fractional GPU sharing and inference serving. Customers can continue using that integration while gaining the ability to inspect, extend and contribute to Run:ai.