NVIDIA Run:ai Is Now Open Source as Part of NVIDIA DSX AI Factory Platform

GPU clusters are still famously underutilized. Run:ai made its name helping enterprises schedule training and inference on Kubernetes without wasting silicon. Today NVIDIA is releasing the entire platform as open source under the DSX AI Factory banner, with four components teams can adopt together or à la carte: KAI Scheduler, GPU Fractioning Operator, KAI Resource Management, and Workload Map.

That is not a token dump of abandoned code. KAI Scheduler is already an Apache 2.0 CNCF Sandbox project. Eight DSX partners — Crusoe, Mirantis, Nebius, OpenNebula Systems, Rafay, Saturn Cloud, Spectro Cloud, and vCluster — are integrating the open stack into managed Kubernetes, white-label inference, and tenant-isolated fleets. Commercial Run:ai via Mission Control and NVIDIA AI Enterprise remains for customers who want NVIDIA support.

We're watching NVIDIA do something it usually avoids: put the production orchestration layer in the community's hands. The strategic read is straightforward. AI factories only scale if the ecosystem ships differentiated products on NVIDIA's rails. Open Run:ai lowers the cost of building those products — quota hierarchies, fractional GPUs, gang scheduling — so partners stop reinventing the control plane and start competing on ops, billing, and sovereign clouds.

The modular split is the smart part. Plenty of platform teams already have a scheduler they like and only need CUDA-native fractioning or a workload contract that understands PyTorchJob and KServe without hardcoding every CRD. Workload Map as a Go library that works offline on unstructured objects is the kind of boring infrastructure that actually gets adopted. GPU Fractioning Operator retiring the proprietary path in favor of CUDA 13.4-native hard memory limits is also a tell: NVIDIA is aligning the open stack with the driver, not protecting an old product moat.

I'm not naive about the downside. Open source without enterprise support still leaves regulated buyers on the paid SKU, and "Community Product on NGC" still means accepting NVIDIA's license terms to download the packaged stack. The likelier trajectory is a two-track market: hyperscalers and AI clouds take the open components and differentiate, while Fortune 500 AI factories keep writing checks for Mission Control. That is fine — it is how Kubernetes won — but it only works if NVIDIA keeps landing contributions upstream instead of forking privately for the commercial line.

Related pieces matter here too. Model Streamer cutting cold-start weight loads, and Grove orchestrating disaggregated prefill/decode graphs, are the inference-side companions. Together they sketch NVIDIA's bet: open the scheduling and lifecycle layer so the factory stays NVIDIA-shaped even when the operator logo on the rack is someone else's.

Run:ai was acquired into NVIDIA's stack after years as an independent GPU orchestration vendor. KAI Scheduler was previously open-sourced and donated to CNCF; today's release extends that openness to the rest of the platform and folds it explicitly into DSX AI Factory positioning for partners building metal-to-model clouds.