NVIDIA Open Sources Run:ai GPU Orchestration Platform Under Apache 2.0 License
NVIDIA releases its full Run:ai GPU orchestration stack as open source software, enabling teams to self-host Kubernetes scheduling without enterprise support.
Key points
- NVIDIA released the complete Run:ai platform under the Apache 2.0 license on GitHub.
- The release includes the KAI Scheduler, GPU Fractioning Operator, and resource management tools.
- Platform teams can now inspect, fork, and self-host the stack without a paid subscription.
NVIDIA has released the full Run:ai GPU orchestration platform as open source software under the Apache 2.0 license. According to MarkTechPost, the source code is now available on GitHub, allowing platform teams to inspect, fork, and self-host the stack without requiring a paid enterprise subscription.
The release integrates Run:ai into the NVIDIA DSX AI Factory platform. It puts production GPU scheduling, quota management, and workload control directly into the hands of developers. The monorepo contains the control plane, cluster services, an installer, a command-line interface, and an MCP server. Local development builds approximately 50 Go services.
Technical Components and Architecture
NVIDIA split the release into four distinct components that teams can adopt independently. The KAI Scheduler is a Kubernetes-native scheduling engine that is also a CNCF Sandbox project. It supports gang scheduling, time-based fair-share, topology awareness, bin-packing, and Device Resource Allocation (DRA) support.
The GPU Fractioning Operator brings CUDA 13.4-native GPU fractions to Kubernetes. This operator uses the CUDA driver to enforce hard memory limits per workload. NVIDIA retired its older proprietary fractioning method in favor of this approach. The operator is scheduler-agnostic and exports metrics for each pod.
KAI Resource Management adds features for projects, departments, quotas, and node pools. NVIDIA provided an example configuration that guarantees 50 percent of a cluster to a specific research team. Workload Map is a Go library that describes workload types such as PyTorchJob, RayCluster, and KServe services. It works offline and does not handle scheduling, queuing, or autoscaling.
Deployment and Performance Context
The platform runs on Kubernetes or OpenShift clusters equipped with NVIDIA GPUs. It operates on top of Kubernetes and does not replace the underlying orchestration system. The control plane handles workload submission, resource management, and monitoring. Cluster services run on each Kubernetes cluster to add GPU-aware scheduling capabilities.
Users can interact with the system through a web UI, CLI, REST API, and an MCP server for AI assistants. NVIDIA published no platform-level benchmark for this specific release. However, the related Model Streamer tool loaded models three to six times faster on eight NVIDIA H100 virtual machines on Azure.
Who Is Affected
Platform teams and developers are the primary beneficiaries of this release. They can now self-host the stack without an enterprise support contract. The community edition lacks NVIDIA enterprise support, which may impact organizations relying on vendor assistance. Local builds require 16 GB of RAM and 100 GB of disk space on Linux systems.
Local development runs on macOS or Linux using a single script named local-dev.sh. This environment simulates eight Tesla-A10 GPUs, meaning no physical GPU hardware is needed for local testing. This lowers the barrier to entry for developers who want to inspect or modify the code before deploying to production clusters.
What should you do?
Teams interested in self-hosting GPU orchestration should review the Apache 2.0 license terms. Verify that your internal security policies allow forking and modifying open-source scheduling engines. Check if your current Kubernetes version supports the required Device Resource Allocation features.
What to do and how to stay safe: NVIDIA
- Audit your internal Kubernetes clusters to ensure they meet the hardware requirements for local development or production deployment.
- Review the Apache 2.0 license obligations before integrating the code into your proprietary software supply chain.
- Verify that your CUDA driver versions are compatible with the new GPU Fractioning Operator features.
- Monitor the CNCF Sandbox project status for updates on the KAI Scheduler roadmap and community contributions.
Step-by-step guide: Mobile Device Security: How to Protect Data Beyond Passwords
General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
Can I use Run:ai without paying for an enterprise subscription?
Yes, the full platform is released under the Apache 2.0 license, allowing teams to self-host without a paid subscription.
Does this release include NVIDIA enterprise support?
No, the community edition has no NVIDIA enterprise support attached to the open-source release.
Do I need a physical GPU to test the local development environment?
No, the local dev environment simulates eight Tesla-A10 GPUs and runs on macOS or Linux without hardware GPUs.
