NVIDIA Hands Run:ai to the Open Source Community

NVIDIA today released the full Run:ai platform as open-source software. The GPU orchestration platform, which NVIDIA acquired at the end of 2024, now serves as an open foundation for the company’s DSX AI Factory platform.

Run:ai helps organizations run AI workloads on Kubernetes. It allocates GPU resources on the fly and manages workloads across their full lifecycle, from data processing and training to inference. Until now, most of that capability was available only through a commercial license.

The timing makes sense. GPUs are expensive and still hard to get. Yet many of them sit idle. Cast AI’s 2026 State of Kubernetes Optimization Report, which analyzed operational data from tens of thousands of clusters, found average GPU utilization hovering around 5%. And a VentureBeat Research survey released in July found that 86% of enterprises running their own GPUs say those chips operate at 50% utilization or less. Better scheduling won’t solve all of that. But it’s a big part of the fix.

What’s In the Release

NVIDIA released the complete platform along with four components that teams can adopt on their own.

KAI Scheduler. This is the Kubernetes-native scheduler at the core of Run:ai. NVIDIA had already released it under the Apache 2.0 license and contributed it to the Cloud Native Computing Foundation (CNCF) as a Sandbox project. It handles gang scheduling, so it places multipod jobs all at once or not at all. It also manages fair-share and quota allocation across teams, topology-aware placement, bin-packing, GPU sharing, and Dynamic Resource Allocation.

GPU Fractioning Operator. Fractional GPUs have long been one of Run:ai’s most popular features. They let several workloads share a single GPU with hard memory isolation. NVIDIA has retired its proprietary version in favor of native fractioning in CUDA 13.4. The operator is the Kubernetes layer that makes those fractions usable, plus a metrics exporter for per-pod visibility. It also works with other schedulers.

KAI Resource Management. This is the organizational layer. It maps GPU capacity to projects, departments, and node pools. NVIDIA’s example: a platform team could guarantee half a cluster to research while keeping any single team from grabbing the rest. If you’re tracking GPU quotas in a spreadsheet today, this is what replaces it.

Workload Map. AI workloads on Kubernetes rarely show up as a simple Deployment or Job. They’re PyTorchJobs, RayClusters, KServe InferenceServices or custom resources. Workload Map defines each type once, so schedulers, dashboards, and cost tools don’t have to hardcode every custom resource definition. It ships as a Go library and adds nothing to the cluster.

The code lives in the DSX AI Factory GitHub organization. NVIDIA also offers a Community Product on NGC with a Helm chart for installing the full stack. It doesn’t require a paid subscription. But it doesn’t include enterprise support, either. Supported versions remain available as a standalone product, through NVIDIA Mission Control and NVIDIA AI Enterprise, and on select cloud marketplaces.

Partners Are Already Building On It

Eight DSX ecosystem partners are integrating the open source platform. Crusoe is adding it to Crusoe Managed Kubernetes. Mirantis is extending k0rdent AI with it. Nebius offers it as a one-click deployment. OpenNebula Systems is building it into its platform for open and sovereign AI clouds. Rafay will deliver it as a fully managed service. Saturn Cloud uses it to help operators launch white-label inference services with per-token billing. Spectro Cloud is bringing it into PaletteAI. And vCluster provides per-tenant templates for stronger isolation on a shared GPU fleet.

That list says a lot about who this is for. AI clouds and platform vendors can now build on a proven orchestration layer instead of writing their own.

NVIDIA also pointed to two related open-source projects. Model Streamer loads model weights concurrently and streams them straight into GPU memory to cut cold-start times. It’s already integrated in vLLM and SGLang. Microsoft’s published benchmarks on eight NVIDIA H100 virtual machines showed loading three to six times faster than the default loader. Grove, part of NVIDIA Dynamo, orchestrates multicomponent inference workloads such as disaggregated prefill and decode.

What It Means for Platform Teams

Open source matters here for three practical reasons. Teams can inspect the code that decides where their workloads run. They can adopt only the piece they need. And they’re less tied to a single commercial distribution. That last point counts. In a survey by ClearML, an AI infrastructure software company, 44% of respondents rated flexibility and avoiding vendor lock-in as “very important” when selecting infrastructure.

But there’s a catch. When NVIDIA closed the deal, Run:ai said its software, which at the time worked only with NVIDIA products, would be open sourced so rivals like AMD and Intel could adapt it for their hardware. Today’s announcement doesn’t address support for other accelerators.

“Open source doesn’t equate to hardware-neutral, and buyers should judge this release on that line,” said Mitch Ashley, vice president and practice lead for CIO & Technology Buyers, and Software Lifecycle Engineering at The Futurum Group. “The fractioning layer runs on CUDA, and nothing announced today changes that. Teams that adopt it are choosing an NVIDIA-dependent control plane for their GPU estate, a commitment that outlasts the next procurement cycle.”

That doesn’t make it a bad choice. Most enterprises running AI at scale already run NVIDIA hardware.

“For enterprises already running NVIDIA fleets, that commitment is reasonable,” Ashley said. “Make it deliberately, with lock-in tolerance as the test. The real check comes when a second accelerator vendor enters the data center.”

For most teams, start small. If GPU sharing is the pain point, try the fractioning operator. If teams keep fighting over capacity, look at KAI Resource Management. And if you need production support, plan for a commercial offering or a partner.

The tools are now open. Whether utilization actually improves depends on how well teams use them.