Run:ai Open Source: NVIDIA Frees Its $800M GPU Scheduler
NVIDIA released the complete Run:ai GPU orchestration platform as open source inside DSX AI Factory on October 13 — no paid subscription required, but no enterprise support either. Here's what the four components do, and what platform teams should adopt first.
On October 13, NVIDIA finally made good on a promise from its $800 million acquisition: the complete Run:ai platform is now open source. The Run:ai open source release lives inside NVIDIA's new DSX AI Factory GitHub organization, and any team can deploy it in production with no paid Run subscription. For platform teams staring at the most expensive line item in AI infrastructure — GPU time — this is the most consequential infrastructure drop of the year.
What the Run:ai open source release actually ships
Run:ai solves a concrete waste problem. Plain Kubernetes was never designed to share GPUs fairly between teams: one researcher's small job can reserve an entire GPU while another team's training run sits in a queue. The scheduler, the quotas, and fractional GPU sharing fix that — and until now, those were the paid parts of Run:ai.
Four components ship, and you can adopt them together or piecemeal:
- KAI Scheduler: the Kubernetes-native scheduling engine. Gang scheduling, fair-share and quota allocation, topology-aware bin-packing, and GPU sharing through MIG, MPS, time-slicing, and HAMi, plus Dynamic Resource Allocation support. It's Apache 2.0 and contributed to the CNCF as a Sandbox project — and it has been open since April 2025. NVIDIA says all scheduler development now happens in the community repo.
- GPU Fractioning Operator: retires Run:ai's old proprietary fractioning in favor of CUDA-native hard memory limits enforced by the CUDA 13.4 driver itself. The operator adds the Kubernetes integration layer and a per-pod metrics exporter, and it works with schedulers other than KAI.
- KAI Resource Management: the org layer — projects, departments, hierarchical quotas, node pools. This is how you guarantee capacity and share it across teams.
- Workload Map: a Go library that describes AI workload types (PyTorchJob, RayCluster, KServe and friends) as common CRDs, so tooling can schedule and observe them without hardcoding every framework.
Why platform teams should care
GPU time is the most expensive line on most AI infrastructure bills, and datacenter demand doubled this year — nobody is getting more capacity cheaply. Until this Run:ai open source release, the org layer (who gets what share of the cluster) and fractional GPUs were Run:ai's paid features. Now they are code anyone can read, fork, and run.
The economics are blunt: if your cluster sits at 30% utilization because teams hoard whole GPUs for small jobs, a fair-share scheduler plus fractioning can be worth more than buying another rack of cards. The AI data center power and cooling crunch already caps how fast you can add hardware — squeezing more out of the GPUs you have is the only lever left.
The catch: free code, no hand-holding
There is a Helm-installable Community Product build on NVIDIA NGC, and it downloads under NVIDIA's software license terms. But enterprise support is not included. NVIDIA's position is straightforward: you run it, you fix it. If you want support entitlements, the commercial offerings still exist.
That is not a dealbreaker — but it should shape your rollout. Run the community build in a non-production cluster first, learn the upgrade and failure modes, and make sure your on-call team can debug a scheduler instead of filing tickets. The NetBox and Grafana patterns GPU teams already use for fleet visibility matter more when you operate the scheduler yourself — the per-pod metrics from the Fractioning Operator are a good starting feed for your dashboards.
What to do this quarter
- Start with the GPU Fractioning Operator if your pain is underutilization. It is scheduler-agnostic, replaces proprietary fractioning with CUDA 13.4-native hard memory limits, and the per-pod metrics show exactly who is wasting what.
- Adopt KAI Scheduler when queue fairness is the problem — multiple teams fighting over one cluster, training jobs starved by inference pods. It has been open since April 2025, so the code already has community mileage.
- Watch the partner integrations before building your own. Crusoe, Mirantis, Nebius, Rafay, Spectro Cloud, and vCluster announced day-one integrations. If you already buy GPU cloud from one of them, let them do the operating for you.
The bigger play
NVIDIA did not spend an estimated $800 million to give away software out of generosity. An open scheduler that more teams can run means more GPUs consumed more efficiently — and more GPUs sold. DSX AI Factory is the wrapper: open foundations for AI workload orchestration that make NVIDIA hardware the default place to run it.
That is fine. Open-sourcing the layer that decides how GPUs get shared is the right move for everyone else too. The founders pledged this at acquisition time in December 2024, and NVIDIA kept the promise. Rare enough in this industry to say plainly: good.
