What was opened
NVIDIA has published the entire Run:ai platform, which manages AI workloads and GPU clusters on Kubernetes, as open source. The code is on GitHub under the Apache 2.0 license, and the platform is now positioned as part of NVIDIA's DSX AI Factory stack.
Run:ai does not replace Kubernetes; it runs on top of it. The control plane handles workload submission, resource management and monitoring, while services running in each cluster add GPU-aware scheduling. Users reach it through a web UI, a CLI, a REST API and an MCP server for AI assistants.
Four independent components
- KAI Scheduler: a Kubernetes-native scheduling engine with gang scheduling, time-based fair-share and topology awareness. It is Apache 2.0 and a CNCF Sandbox project.
- GPU Fractioning Operator: brings CUDA 13.4-based GPU fractions to Kubernetes, with the driver enforcing a hard memory limit per workload.
- KAI Resource Management: adds projects, departments, quotas and node pools.
- Workload Map: a Go library describing workload types such as PyTorchJob, RayCluster and KServe.
Trying it locally
A single script (local-dev.sh) builds about 50 Go services from source and creates a five-node Kind cluster pinned to Kubernetes v1.32.5. A fake GPU operator simulates 8 Tesla-A10 GPUs, so no real hardware is needed. macOS and Linux are supported, Windows is not; Linux hosts need 16 GB of RAM and 100 GB of free disk.
Limits
The community edition requires no paid subscription but excludes NVIDIA enterprise support. Commercial editions continue through NVIDIA AI Enterprise, NVIDIA Mission Control and select cloud marketplaces. AMD and Intel accelerators are not supported.
