NVIDIA Releases Full Run:ai GPU Orchestration Stack as Open Source

NVIDIA published the full source code for its Run:ai GPU orchestration platform under the Apache 2.0 license, according to MarkTechPost. This move allows platform teams to inspect, fork, and self-host the stack on Kubernetes or OpenShift clusters without maintaining a paid subscription. The release integrates Run:ai into the NVIDIA DSX AI Factory platform, giving developers direct control over production GPU scheduling, quotas, and workload management.

What components are included in this release?

The release comprises a monorepo containing the control plane, cluster services, installer, CLI, and an MCP server. It features approximately 50 Go services. Key modules include the KAI Scheduler, a CNCF Sandbox project that handles gang scheduling and topology awareness. Also included is the GPU Fractioning Operator, which uses CUDA 13.4 to enforce hard memory limits per workload. KAI Resource Management adds quota and project controls, while Workload Map describes job types like PyTorchJob and RayCluster.

How does this affect operational support and resources?

Teams adopting the open-source version lose access to NVIDIA enterprise support, according to MarkTechPost. Local development builds require 16 GB of RAM and 100 GB of disk space on Linux systems. The local development environment simulates eight Tesla-A100 GPUs, allowing builds on macOS or Linux without physical hardware. NVIDIA has not published platform-level benchmarks for this specific release, though related Model Streamer tests showed faster loading on Azure H100 VMs.

What is the technical scope of the platform?

Run:ai operates on top of Kubernetes and does not replace it. The control plane manages workload submission and monitoring, while cluster services provide GPU-aware scheduling. Users interact via web UI, CLI, REST API, or an MCP server for AI assistants. The GPU Fractioning Operator is scheduler-agnostic and exports per-pod metrics. NVIDIA retired its older proprietary fractioning method in favor of this CUDA-native approach.

What should operations teams consider next?

Organizations should evaluate whether self-hosting this stack aligns with their internal capacity. The absence of enterprise support means teams must handle troubleshooting and updates independently. Security operations must review the code for vulnerabilities before deployment. Teams should verify that their Kubernetes clusters meet the resource requirements for local builds and production scheduling.

Questions to ask your suppliers

Verify whether your current GPU management tools can integrate with open-source schedulers. Ask vendors if they provide support for community-maintained forks of orchestration platforms. Determine who inside your organization will maintain the codebase and handle security patches. Check if your licensing agreements cover modifications to upstream projects. Ensure your risk team understands the implications of removing vendor support for critical infrastructure components. Confirm that your development teams have the skills to build and debug Go-based services locally.

What to do and how to stay safe: NVIDIA

• Review the Apache 2.0 license terms to ensure compliance with your organization's software usage policies.
• Audit the provided source code for security vulnerabilities before deploying to production environments.
• Verify that your Kubernetes clusters have sufficient resources to run the additional control plane services.
• Monitor community channels for updates and security advisories regarding the open-source components.

Step-by-step guide: Browser Security Settings for Small Business Teams

General security guidance from the TechSec Feed newsroom. It is not confirmed advice from the organisations named in this story.

Frequently asked questions

Is NVIDIA Run:ai free to use?

Yes, the source code is released under the Apache 2.0 license, allowing free use and modification.

Does this release include enterprise support?

No, the community edition has no NVIDIA enterprise support, according to MarkTechPost.

What hardware is needed for local development?

Local builds require 16 GB RAM and 100 GB disk on Linux, with simulated GPUs.