NVIDIA Open-Sources the Full Run:ai Platform

NVIDIA announced on October 13, 2026, that it has open-sourced the full Run:ai platform. The release is part of the company’s DSX AI Factory platform. Alongside the integrated platform, NVIDIA released four modular components that teams can adopt independently, plus two companion open-source projects.

Background: the acquisition

NVIDIA completed its acquisition of Run:ai on December 30, 2024. Run:ai built software for scheduling AI workloads and managing GPU resources on Kubernetes. The company helps teams pool GPUs, share them across projects, and allocate fractions of GPUs to different workloads.

Open-sourcing the technology is a shift from proprietary to shared infrastructure. Instead of selling the orchestration software as a closed product, NVIDIA is releasing it for the community to use, modify, and build on.

What’s released: the platform and four components

The release covers the complete integrated Run:ai platform. In addition, four modular components can be adopted on their own:

KAI Scheduler — a Kubernetes scheduler for AI workloads, released under the Apache 2.0 license and contributed to the Cloud Native Computing Foundation (CNCF) as a Sandbox project. It handles gang scheduling (placing groups of related workloads together), topology-aware placement (putting workloads where hardware connections are fastest), GPU sharing through NVIDIA’s MIG and MPS technologies, and fair-share quotas so teams get equitable access to GPU capacity.

GPU Fractioning Operator — provides native GPU fractioning on Kubernetes through CUDA 13.4, with driver-enforced memory limits. It allows multiple workloads to share a single GPU while keeping memory boundaries enforced at the driver level.

KAI Resource Management — maps GPU capacity to projects and departments, giving platform teams a way to manage who gets what across an organization.

Workload Map — a common contract for defining workload types, covering standard Kubernetes AI workload definitions such as PyTorchJob, RayCluster, and KServe InferenceService.

Of the four, the KAI Scheduler is the most independent-friendly piece. The Apache 2.0 license and the CNCF Sandbox status give it governance and licensing that do not depend on NVIDIA’s control. Teams can adopt it without buying into the rest of the stack.

Two companion projects

NVIDIA also released two companion projects with the announcement.

Model Streamer loads model weights concurrently to reduce inference cold starts. When a model is first loaded for serving, the delay before it can answer requests — the cold start — is a real operational problem. Streaming weights concurrently shortens that delay. Model Streamer has already been integrated upstream in vLLM and SGLang, two of the most widely used inference-serving frameworks.

Grove is a Kubernetes API for multi-component workloads, including agentic pipelines. As AI systems move from single models to pipelines of multiple components and agents, scheduling needs to understand the whole pipeline rather than individual jobs. Grove provides the API layer for that.

What Run:ai does

Run:ai is Kubernetes-based software for AI workload scheduling and GPU resource allocation. It handles GPU pooling and sharing, fractional allocation of GPUs to workloads, fair-share quotas across teams, and utilization monitoring. Teams use it to run AI workloads — both training and inference — on shared GPU clusters.

The practical problem it solves is GPU waste. GPUs are expensive, and many workloads do not fully use one. Without scheduling software, GPUs sit idle between jobs or get locked to teams that are not using them. Pooling and fractioning let platform teams pack more work onto the same hardware and measure who is using what.

Why it matters

NVIDIA is turning technology it acquired into open infrastructure. That is a deliberate move. By open-sourcing the full platform, NVIDIA makes the orchestration layer a shared standard rather than a competitive product. Platform teams at enterprises and AI cloud providers can adopt it whole or piecemeal, integrating individual components like the KAI Scheduler into their existing stacks.

The strategy also positions NVIDIA’s hardware as the natural substrate. Open scheduling software that works well with MIG, MPS, and CUDA integrations lowers the friction for running AI workloads on NVIDIA GPUs. Competitors can use the software too, but the deepest integration paths run through NVIDIA’s stack.

The CNCF Sandbox contribution for the KAI Scheduler is the clearest signal of intent. Sandbox status is the entry level of the CNCF process, but it puts the project under foundation governance and in front of the Kubernetes community. If the scheduler matures there, it could become the default way AI teams schedule GPU workloads on Kubernetes — similar to how Kubernetes itself became the default container orchestrator.

There is also a talent and ecosystem angle. Open-sourcing the platform lets NVIDIA’s orchestration work be studied, improved, and extended by the broader community. Contributions from AI cloud providers and enterprise platform teams could improve the software faster than NVIDIA could alone, and each improvement feeds back into the ecosystem around NVIDIA GPUs.

What to watch

Adoption will be the test. Open-sourcing a platform is only the start; the project needs maintainers, documentation, and production users. The KAI Scheduler’s CNCF path gives it the best chance of independent adoption. Model Streamer’s existing integration with vLLM and SGLang is a head start — it is already useful in real serving stacks.

For platform teams running shared GPU clusters, the release is worth evaluating now. The modular design means a team can try the KAI Scheduler or the GPU Fractioning Operator without committing to the whole platform. For AI cloud providers, the release adds a new open option for multi-tenant GPU infrastructure. The next year will show whether the community picks it up — and whether NVIDIA’s bet on open infrastructure pays off in ecosystem control.