NVIDIA Open-Sources Run:ai, the AI Workload Orchestration Platform for Kubernetes (KAI Scheduler and GPU Fractioning)

AIDeveloper44 Team

Overview of the Open Source Release

NVIDIA has announced the transition of its Run:ai platform to an open source model, incorporating it into the broader NVIDIA DSX AI Factory platform. This release provides developers, enterprise platform teams, and ecosystem partners with direct access to production-grade tools designed for AI workload and GPU orchestration. By moving these capabilities to an open source framework, NVIDIA aims to foster greater integration across diverse AI infrastructure environments.

The platform is specifically built for Kubernetes, serving as an orchestration layer that helps manage shared AI resources. According to NVIDIA, the project is designed to help organizations maximize throughput and GPU efficiency, particularly in environments where multiple teams share underlying computing clusters.

Key Modular Components

The open source release encompasses the full stack of the Run:ai platform but is structured modularly. This allows infrastructure teams to adopt specific components based on their existing stack requirements. The four primary modular components include:

• KAI Scheduler: A Kubernetes-native scheduler that provides GPU-aware capabilities, enabling more sophisticated workload placement and bin-packing.
• GPU Fractioning Operator: A tool designed to enable CUDA-native GPU sharing, allowing multiple workloads to run on a single physical GPU through fractional allocation.
• KAI Resource Management: Functionality for managing organizational quotas, hierarchical resource allocation, and team-based infrastructure policies.
• Workload Map: A framework that establishes a consistent contract for managing various AI workload types, from development environments to training and inference tasks.

In addition to these core components, the repository includes associated platform services, CLI tools, and the MCP server, providing a complete environment for managing shared AI infrastructure.

Platform Integration and Adoption

NVIDIA has structured the project to support multiple adoption paths. Organizations can deploy the source code directly from the official repository or install the platform using Helm charts available on NVIDIA NGC as a community-supported product. This approach provides flexibility for developers who wish to inspect the code, contribute to the project, or build custom solutions on top of the established orchestration foundation.

Several ecosystem partners are already integrating these open source capabilities into their own offerings. Companies including Crusoe, Mirantis, Nebius, OpenNebula Systems, Rafay, Saturn Cloud, Spectro Cloud, and vCluster are leveraging the platform to enhance their managed Kubernetes and AI infrastructure services. These integrations typically focus on automating lifecycle management and simplifying the deployment of AI-ready clusters.

Compatibility and Support

The open source release functions as a scheduling and governance layer on top of existing Kubernetes environments. It does not replace the container runtime or core Kubernetes scheduling, but rather extends them with specialized logic for GPU-bound workloads. This compatibility makes it suitable for on-premises, cloud, or hybrid infrastructure setups.

While the software is now available as open source, NVIDIA continues to offer support-entitled versions for organizations that require enterprise-grade SLAs, commercial support, and formal product maintenance. This dual-track approach ensures that organizations can access the technology via the community-driven model while retaining the option for vendor support when mission-critical requirements dictate.

For those interested in development, the project is hosted as a monorepo, which includes the control plane services, Kubernetes operators, and shared client libraries. Interested parties are encouraged to explore the official documentation and the GitHub repository to begin testing the platform components within their own Kubernetes clusters.

Enjoyed this?

Get more posts like this delivered to your inbox.

Subscribe