What was released

Run:ai solves a familiar waste problem. On a shared cluster, one team's small job can reserve a whole GPU while another team's training run waits. The platform schedules work across the cluster, lets several workloads share one GPU in fractions with hard memory limits, and applies fair-share quotas so teams can borrow idle capacity and give it back when its owner needs it.

Nvidia has now published that whole stack. A team can take the integrated platform, or take only the piece it lacks. KAI Scheduler is the engine: gang scheduling for multi-pod jobs, topology-aware placement, GPU sharing through MIG and MPS, and support for Dynamic Resource Allocation. The GPU Fractioning Operator is a Kubernetes integration for CUDA 13.4-native GPU fractioning, replacing Run:ai's earlier proprietary fractioning mechanism. Nvidia says the CUDA driver enforces memory limits, while the operator exposes GPU fractions to Kubernetes and provides per-pod metrics. KAI Resource Management maps capacity to projects and departments. Workload Map gives tools a common way to handle workload types such as PyTorchJob, RayCluster and KServe InferenceService without hard-coding each one. It does not itself schedule, queue or autoscale workloads.

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

Where the paid line sits

The announcement is explicit about the split, and it is worth reading literally. The code is available to inspect, modify and build on. The Community Product on NGC needs no paid subscription. It also includes no enterprise support, and downloading it requires accepting Nvidia's software licence agreement and product-specific terms for AI products, on top of the open-source licences that govern the code itself.

That leaves three distinct offers where there used to be one. Self-supported open source for platform teams with their own expertise. Partner-integrated or managed versions, which arrived immediately: Nvidia named eight DSX ecosystem partners integrating Run:ai, including Mirantis and vCluster Labs (per-tenant managed clusters), and Lambda separately announced Run:ai as a first-class option on Lambda Kubernetes Service. And Nvidia's own commercial Run:ai, still sold with support through its enterprise channels.

This also completes a promise made at the acquisition. When Nvidia closed the Run:ai deal in December 2024, founders Omri Geller and Ronen Dar said open-sourcing the software could extend its reach beyond Nvidia hardware. Almost two years later, the reach argument is the strategy: a scheduler that partners embed in their own clouds is harder to displace than one sold only as an Nvidia add-on.

What to watch

Adoption is the test that matters now. KAI Scheduler development will happen in the community repository, according to Nvidia, so its commit history and CNCF progress are public signals rather than press-release claims. The other signal is whether clouds that are not in Nvidia's partner list adopt the components, especially the scheduler and the fractioning operator, which are useful without the full platform.

For most teams the practical question is narrower. If idle GPU capacity and queue fights are costing real money, the open release removes the licence excuse for not testing Run:ai or KAI Scheduler. If the cluster carries production workloads with an uptime commitment, the price of the supported version was never mainly the code.

Sources

  1. [1] Nvidia developer blog, "NVIDIA Run:ai Is Now Open Source as Part of NVIDIA DSX AI Factory Platform", 13 October 2026 (primary source for release scope, components, NGC terms and commercial availability)Read source
  2. [2] Reuters, "Nvidia completes Run:ai acquisition", December 2024 (reported approximately $700 million price)
  3. [3] CTech (Calcalist), "Nvidia opens up Israeli-founded Run:ai software two years after $800 million acquisition", 13 October 2026 (acquisition price estimate of $800 million, history, founders' open-source commitment)Read source
  4. [4] Business Wire / vCluster Labs, "vCluster Labs Turns Open-Source NVIDIA Run:ai Into a Managed Cluster Product", 13 October 2026Read source
  5. [5] Lambda blog, "Run:ai open source comes to Lambda Kubernetes Service", 13 October 2026Read source
  6. [6] Business Wire / Mirantis, "Mirantis Integrates Open-Source NVIDIA Run:ai with k0rdent AI", 13 October 2026Read source