Nvidia Open-Sources Run:ai, Its $800M GPU-Orchestration Platform

Nvidia is giving away one of the most expensive pieces of software it ever bought. The chip giant announced on October 13 that it has released the complete Run:ai platform as open source — the GPU-orchestration software it picked up when it acquired the Israeli startup for roughly $800 million, a deal that closed on December 30, 2024.

The release makes production-grade AI workload scheduling available to customers, developers, and ecosystem partners without a paid Run:ai subscription or Nvidia support entitlement, according to a post on the Nvidia Developer Blog. It slots into Nvidia's DSX AI Factory platform as an open foundation for the wider ecosystem.

Nvidia published the whole stack. The complete Run:ai platform — including its control plane — is open, and four components can also be adopted independently: the KAI Scheduler, the GPU Fractioning Operator, KAI Resource Management, and Workload Map.

Each covers a different part of the waste problem the platform was built to solve. The KAI Scheduler handles gang scheduling for multi-pod jobs, topology-aware placement, and GPU sharing through MIG and MPS so small jobs stop hogging entire chips.

The GPU Fractioning Operator is the Kubernetes integration for CUDA-native GPU fractioning, with the CUDA driver enforcing memory limits and per-pod metrics exposed.

KAI Resource Management maps capacity to projects and departments with fair-share quotas, and Workload Map gives tools a common way to handle workload types such as PyTorchJob, RayCluster, and KServe InferenceService without hard-coding each one.

The code carries open-source licenses — the project repository identifies Apache License 2.0 for its software code. But free does not mean fully supported: enterprise support is not included, so teams using the community version must run their own infrastructure, handle their own upgrades, and manage their own operations.

Nvidia will continue to sell commercial versions through its enterprise AI products and supported cloud offerings.

The release brings two related projects with it. Model Streamer loads model weights concurrently into GPU memory to cut inference cold-start time, and it is already integrated upstream in vLLM and SGLang. Grove provides a Kubernetes API for multi-component workloads, including agentic pipelines and disaggregated serving.

Nvidia says eight partners are already integrating the open-source Run:ai release. The pitch to the ecosystem is straightforward: adopt the whole platform, take a single component, or build on it for their own products — and bring differentiated AI-cloud solutions to market faster.

The strategy is about standards, not software revenue. Founded in 2018, Run:ai built a software layer that manages how AI workloads use GPUs — the processors central to training and running modern AI models.

Making that layer open could extend its reach beyond Nvidia's own GPUs and into the broader AI ecosystem, embedding Nvidia's orchestration model wherever AI clouds get built.

In a sense, Nvidia is trading license revenue for gravitational pull: the more infrastructure runs on Run:ai, the stickier its hardware ecosystem becomes.

Whether customers run the free community product or pay for the supported tier, they are building on Nvidia's foundation — which, two years after an $800 million acquisition, may be worth more as a free standard than as a paid product.