NVIDIA Open Sources Run:ai Under Apache 2.0 for Kubernetes GPU Orchestration
MarkTechPost Asif Razzaq
NVIDIA has open sourced Run:ai on GitHub under Apache 2.0. That means teams can inspect, fork and self-host its GPU scheduling stack without a paid subscription.
Based on reporting by MarkTechPost, Asif Razzaq — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
NVIDIA has put the full Run:ai platform out as open source software, with the code now on GitHub under Apache 2.0. The release folds Run:ai into the NVIDIA DSX AI Factory platform, but it also changes who gets to tinker with it: platform teams can inspect the stack, fork it and run it themselves without paying for a subscription.
This is not just a scheduler. Run:ai covers workload submission, resource management and monitoring on top of Kubernetes, and it reaches users through a web UI, CLI, REST API and an MCP server for AI assistants. NVIDIA says the platform is meant for development, training and inference across shared GPU clusters, while Kubernetes stays in place underneath.
The open release is split into four modular pieces. KAI Scheduler handles Kubernetes-native scheduling and includes gang scheduling, fair-share controls, topology awareness, bin-packing and DRA support. GPU Fractioning Operator adds CUDA 13.4-native GPU fractions with hard memory limits enforced by the CUDA driver. KAI Resource Management deals with projects, departments, quotas and node pools. Workload Map is a Go library for describing workload types such as PyTorchJob, RayCluster and KServe services.
NVIDIA also documented how the stack runs locally. The dev flow is built around a single script, local-dev.sh, which compiles about 50 Go services in containers and brings up a 5-node Kind cluster pinned to Kubernetes v1.32.5. A fake GPU operator simulates 8 Tesla-A10 GPUs, so no actual GPU hardware is required for local work. Linux needs 4 CPUs, 16 GB of RAM and 100 GB of free disk; macOS on Podman needs at least 12 GB of RAM.
For production, teams can deploy from source or use the NGC Community Product Helm chart. That edition does not require a paid subscription, but it also does not include NVIDIA enterprise support. NVIDIA points to eight DSX ecosystem partners, plus commercial routes through NVIDIA AI Enterprise, NVIDIA Mission Control and selected cloud marketplaces. The company also published adjacent projects such as Model Streamer and Grove, while noting that it did not publish a platform-level benchmark for Run:ai itself.
My take — AI-written commentary, not fact-checked reporting
This is NVIDIA doing the classic open-source move: give away the plumbing, keep the polished extras and the support contracts. That’s not cynical, just business. The interesting bit is that GPU orchestration is becoming important enough that even the vendor wants outsiders building on top of it instead of around it.
Read more about this at: MarkTechPost
Related stories
Microsoft Open-Sources TauGrid: A Kubernetes-Native Stack for GPU AI Workloads
MarkTechPost · 1 month ago ·
32
NVIDIA Open-Sources OSMO: One YAML Orchestrates Physical AI Training, Simulation, and Robot Testing
MarkTechPost · 1 month ago ·
32
NVIDIA and Local AI Community Fuel Open Source Models and Intelligent Agents
NVIDIA · 2 months ago ·
25