NVIDIA Open-Sources the Full Run:ai Platform, Putting Its AI Factory Scheduler on GitHub Under Apache 2.0

NVIDIA has open-sourced its entire Run:ai platform, the GPU orchestration layer it acquired in 2024 and has since made the scheduling backbone of its AI factory deployments. The release positions Run:ai as an open foundation for the broader DSX AI Factory ecosystem, letting AI cloud providers and enterprises take the full platform or lift out individual components. The complete source now lives in NVIDIA's DSX AI Factory GitHub organization, with an integrated stack also published as a community product on NGC complete with Helm charts.

The centerpiece is the KAI scheduler, the Kubernetes-native engine that drives Run:ai. NVIDIA has released it under Apache 2.0 and contributed it to the Cloud Native Computing Foundation as a sandbox project, committing to keep all scheduler development in the community repository. KAI is built for GPU workloads at scale: gang scheduling with memory simulation so multi-pod jobs land on an all-or-nothing basis, fair sharing with time-based weighting and quota allocation, topology-aware placement, and support for nearly every GPU sharing mechanism in the field, including MIG, MPS, time-slicing, HaMi and the CUDA 13.4 MPS segmentation scheme, alongside Kubernetes dynamic resource allocation.

The second component brings CUDA-native GPU segmentation to Kubernetes. Run:ai pioneered hard-memory-isolated GPU sharing with a proprietary implementation, but that path is now deprecated in favor of native capability introduced in CUDA 13.4: the driver enforces a hard cap on each workload's memory, and a workload only ever sees its own slice. NVIDIA's new operator wires that capability into clusters and ships a metrics exporter so operators can watch per-pod segmentation usage. The remaining two components, KAI resource management and workload mapping, round out the modular set and can be adopted independently.

Ecosystem partners are already productizing the release. Rafay is turning Run:ai into a fully managed service that automates the stack from GPU infrastructure to control plane and tenant management; Saturn Cloud is using it for white-label inference services on its Token Factory platform with per-token metering and billing; Spectro Cloud has integrated it into PaletteAI for lifecycle management; and vCluster is offering pre-validated deployment templates that give every tenant a complete environment on a shared GPU pool.

Nothing changes for enterprises that buy rather than build. For NVIDIA Enterprise AI Factory deployments, Run:ai continues to ship through NVIDIA-supported commercial channels — as a standalone product, inside NVIDIA Mission Control and NVIDIA AI Enterprise, and in select cloud marketplaces — with availability and licensing terms varying by offering.

Strategically, the move reads as a land grab for the orchestration layer of the AI data center. By handing the scheduler to the CNCF and opening the segmentation stack, NVIDIA makes it easier for every AI cloud — including ones that compete with it — to standardize on its tooling, the same playbook that turned Kubernetes into infrastructure nobody questions. The bet is that the more the industry schedules GPUs through NVIDIA's code, the harder it becomes to build an AI factory around anything else.