NVIDIA has made the full Run:ai platform open source, according to a post on its technical blog published October 15. The release is part of the company’s DSX AI Factory Platform and is meant to give cloud providers, platform teams and software vendors an open foundation for scheduling AI workloads on shared GPU clusters.

What is included

NVIDIA says the release includes the complete Run:ai platform plus four modular components that can be adopted independently:

  • KAI Scheduler, a GPU-aware scheduler for Kubernetes.
  • GPU Fractioning Operator, which enables CUDA-native sharing of a single GPU between workloads.
  • KAI Resource Management, which handles organizational quotas and hierarchies.
  • Workload Map, described as a consistent workload-management contract.

Two related open-source projects round out the offer: Model Streamer, which speeds up loading models into GPU memory, and Grove, which orchestrates inference workloads made of multiple components. NVIDIA invites teams to adopt the platform as a whole, build their own products on it, or pick individual pieces.

Why GPU orchestration matters

GPUs are the most expensive resource in an AI cluster, and they are often poorly used. A training job may need many GPUs at once for hours, while a small inference service might need only a fraction of one card. Without a scheduler that understands GPUs, teams tend to reserve whole devices “just in case,” leaving them idle. Run:ai-style orchestration tackles this in a few ways: it queues and prioritizes jobs by team quotas, it can pack workloads together to keep cards busy, and fractioning lets several light workloads share one GPU. NVIDIA’s pitch is that this raises throughput and utilization for shared AI infrastructure.

The company’s post says Run:ai has been used for several years to run AI workloads on Kubernetes at scale, allocating GPU resources dynamically across the lifecycle from data processing through training to inference. Opening it up means anyone can inspect, modify and combine the code instead of buying it only as a commercial product.

Ecosystem adoption

NVIDIA lists eight DSX ecosystem partners integrating the open-source platform into their offerings: Crusoe, Mirantis, Nebius, OpenNebula Systems, Rafay, Saturn Cloud, Spectro Cloud and vCluster. These are mostly cloud and Kubernetes platform vendors, which fits the target: companies that resell GPU capacity and want a differentiated scheduling layer without writing one from scratch. The post does not give licensing terms beyond calling the release open source, so teams should read the repository licenses before building a commercial product on it.

Strategic angle

NVIDIA’s software ecosystem, CUDA above all, is a core reason its hardware is hard to replace, a theme we cover in why NVIDIA dominates AI. Releasing orchestration software openly can deepen that ecosystem: the more clouds build on NVIDIA-originated scheduling and sharing tools, the more natural it is for them to keep standardizing on NVIDIA GPUs. It also answers competitors and open projects in Kubernetes scheduling by making NVIDIA’s approach the common baseline. This is an interpretation, not something NVIDIA states.

For practitioners, the practical question is memory and sharing. Fractional GPUs only help when workloads fit in the slice they receive, so sizing still matters; see our GPU memory sizing guide for how to estimate it, and how LLM inference serving works for the serving side.

What to watch next

  • Which license terms apply to each repository, and how contributions from outside NVIDIA are handled.
  • Independent measurements of utilization gains from GPU fractioning in production.
  • Whether cloud partners ship managed products built on the open platform.
  • How the open KAI Scheduler compares with other Kubernetes batch schedulers.
  • Whether other chipmakers release comparable open orchestration layers.

Key takeaways

  • NVIDIA open-sourced the full Run:ai platform as part of its DSX AI Factory Platform.
  • Four components can be used independently: KAI Scheduler, GPU Fractioning Operator, KAI Resource Management and Workload Map.
  • Eight cloud and platform partners are integrating it.
  • The goal is higher GPU utilization on shared Kubernetes clusters.
  • Utilization gains are NVIDIA’s claim until measured independently.

Frequently asked questions

What is Run:ai?

It is software for scheduling and sharing GPUs among AI workloads on Kubernetes, now released by NVIDIA as open source.

Do I need NVIDIA GPUs to use it?

The components are built around NVIDIA GPUs, including CUDA-native fractioning, so it is aimed at NVIDIA-based clusters.

Can I use only one part, such as the scheduler?

Yes. NVIDIA says the four modular components can be adopted independently of the full platform.