NVIDIA has open-sourced the full Run:ai platform, its software for orchestrating AI workloads and sharing GPUs across Kubernetes clusters. The complete codebase — control plane, cluster services, installer, CLI and an MCP server — is now on GitHub under the Apache 2.0 license.
Run:ai schedules training, development and inference jobs across shared GPU fleets, with quotas and GPU-aware placement that plain Kubernetes was not designed to provide. Until now, the layer that decided which team got which slice of GPU time was the paid part of the product.
The release goes beyond the already-open KAI Scheduler to include four independently adoptable components: the KAI Scheduler itself, a GPU Fractioning Operator that lets multiple workloads share GPU capacity with CUDA-native memory limits, KAI Resource Management for hierarchical quotas across teams, and a Workload Map that describes training and inference jobs in one format.
Teams can now inspect, fork and self-host the entire stack without a paid subscription. A Helm-installable Community Product build is also available through NVIDIA’s NGC collection, and developers can test-drive it locally on a five-node Kind cluster with eight simulated Tesla-A10 GPUs — no real GPU needed.
The move follows NVIDIA’s acquisition of Run:ai, which cleared European regulatory review in December 2024, and folds the platform into the company’s DSX AI Factory lineup. The community edition carries no NVIDIA enterprise support, so adopters take on upgrades and incident response themselves.
For teams running open-weight models such as Meta AI or DeepSeek on shared clusters, the economics of inference just got friendlier — the real question is whether the savings outweigh the operational load.
Source: MarkTechPost
