NVIDIA Run:ai Goes Open Source, and Support Is the Part It Kept

NVIDIA published the complete Run:ai platform as open source on October 13, so the scheduling layer under a shared GPU cluster is now installable without a subscription. The NVIDIA Run:ai open source release ships the integrated platform plus four components a team can adopt on its own. Enterprise support is the piece that stayed commercial.

Fractional GPU sharing moved down to the driver

The four parts map to real operator problems. KAI Scheduler places GPU workloads on Kubernetes with gang scheduling, fair-share quotas and topology-aware bin packing. The GPU Fractioning Operator lets several workloads share one card, and the memory ceiling is enforced by the CUDA driver rather than by the scheduler. KAI Resource Management adds the org layer, projects, departments and node pools. Workload Map describes a PyTorchJob, a RayCluster, a JobSet and a KServe object in one shape so other tools stop hardcoding each custom resource.

Fractional sharing is the part worth reading twice. Run:ai shipped it for years as a proprietary feature, and that implementation is now retired in favour of native fractioning in CUDA 13.4. The driver holds the isolation boundary, so a fractional workload only sees the memory it was given. Anyone sizing nodes around the old memory limits should recheck the new path.

The licence split is the practical catch. The community repository carries the source, and the NGC Community Product needs no paid subscription and no support entitlement. The commercial route continues through NVIDIA AI Enterprise, Mission Control and the cloud marketplaces, and terms vary by offering. Open code suits a platform team that can read it. A regulated cluster that needs a vendor on the hook is harder to sell.

Eight DSX partners went public the same day. Crusoe lists Managed Kubernetes as a compatible distribution. Nebius ships it as a one-click application, and Rafay, Spectro Cloud, Mirantis and vCluster published integrations. KAI Scheduler is already a CNCF Sandbox project, and scheduler development now happens in the community repository.

This follows NVIDIA’s earlier decision to hand out its rack and power designs. Three questions for your own cluster. Which team owns GPU quota today, a spreadsheet or a scheduler? Would fractional sharing let a notebook stop reserving a whole card? And when the code is open and the support is not, who do you page at 2am?