TL;DR: On October 13, 2026, NVIDIA released the full Run:ai platform as open source software. Run:ai is the GPU and AI workload orchestration layer NVIDIA bought in late 2024 for an estimated $800 million. The release includes the complete platform plus four parts you can use on their own: KAI Scheduler, GPU Fractioning Operator, KAI Resource Management and Workload Map. Eight cloud and infrastructure partners say they are already building it into their products. For teams running AI on Kubernetes, it means production-grade GPU sharing and scheduling without a paid license, though NVIDIA enterprise support is not included.
What NVIDIA actually released
Run:ai started in 2018 as an Israeli startup building software that decides how AI jobs share expensive graphics processors. Instead of letting teams fight over GPUs, or leaving a chip half idle because one notebook grabbed the whole thing, Run:ai allocates capacity dynamically, including slicing a single GPU between several smaller workloads. NVIDIA completed its acquisition in late 2024 after an eight-month regulatory review in the US and Europe, according to Calcalist’s CTech, which put the price at roughly $800 million.
At the time of that deal, founders Omri Geller and Ronen Dar said they planned to open source the software. This week NVIDIA followed through. In its technical blog announcement, the company said the complete platform now lives in the NVIDIA DSX AI Factory GitHub organization, and is also offered on NVIDIA NGC as a “Community Product” with a Helm chart for installing the whole stack.
NVIDIA frames the move as part of its DSX AI Factory Platform, its umbrella for the hardware and software that power large AI data centers. In the company’s own words, open sourcing Run:ai lets “every NVIDIA DSX ecosystem partner and AI cloud” increase throughput and GPU utilization.
The four building blocks
The interesting part of the release is that it is modular. You can adopt the integrated platform, or pick only the pieces that fill a gap in a stack you already run.
| Component | What it does | Good fit for |
|---|---|---|
| KAI Scheduler | Kubernetes-native scheduler for GPU workloads: gang scheduling, fair-share and quotas, topology-aware placement, bin-packing and GPU sharing | Clusters shared by many teams running training and inference side by side |
| GPU Fractioning Operator | Lets workloads use CUDA-native GPU fractions on Kubernetes, with hard memory limits and per-pod metrics | Notebooks, dev sessions and small models that don’t need a full GPU |
| KAI Resource Management | The organizational layer: projects, departments, guaranteed shares and node pools | Platform teams currently tracking GPU quotas in spreadsheets |
| Workload Map | A Go library that describes AI workload types (PyTorchJob, RayCluster, KServe and others) so tools can handle them consistently | Builders of dashboards, schedulers and ML platforms |
KAI Scheduler was already open
KAI Scheduler itself is not new to open source. NVIDIA released it under the Apache 2.0 license in April 2025 and contributed it to the Cloud Native Computing Foundation as a Sandbox project, as Nebius notes in its launch post. NVIDIA says all scheduler development still happens in the community repository, so the commercial product and the open version share one codebase. What is new is everything around it: the control plane, user management, organizational quotas and workload handling.
GPU fractions move into CUDA
One detail matters a lot for anyone who has used Run:ai’s GPU slicing. NVIDIA says its old proprietary fractioning implementation has been retired in favor of a native capability in CUDA 13.4, where the driver itself enforces a hard memory limit for each workload. The new operator is the Kubernetes layer that exposes those fractions, and NVIDIA describes it as scheduler-agnostic, so it can work with other schedulers as well as KAI.
Who is building on it
NVIDIA lists eight ecosystem partners integrating the open platform. These descriptions are the companies’ own, as summarized by NVIDIA:
- Crusoe is pairing it with Crusoe Managed Kubernetes.
- Mirantis is extending its k0rdent AI stack with it.
- Nebius offers it as a one-click application on Nebius AI Cloud.
- OpenNebula Systems is integrating it with its AI cloud platform for open and sovereign clouds.
- Rafay plans to deliver it as a fully managed service.
- Saturn Cloud is using it for white-label inference services with per-token billing.
- Spectro Cloud is adding it to PaletteAI.
- vCluster provides pre-validated templates that give each tenant its own environment.
What’s free and what isn’t
This is not the end of paid Run:ai. NVIDIA says the commercial product continues through supported offerings, including as a standalone product, inside NVIDIA Mission Control and NVIDIA AI Enterprise, and on some cloud marketplaces.
The open release comes with a few conditions worth reading carefully:
- The source code is governed by its own open source licenses.
- Downloading the Community Product from NGC requires accepting NVIDIA’s software license agreement and product-specific terms.
- No paid subscription is needed for the Community Product, but NVIDIA enterprise support is not included.
In practice, a startup or research lab can now run the same orchestration layer that large enterprises pay for, as long as it is comfortable supporting it in-house or through a partner.
Related projects: Model Streamer and Grove
NVIDIA also pointed to two companion open source projects. Model Streamer speeds up the cold start of inference servers by streaming model weights straight into GPU memory, and it is already integrated upstream in vLLM and SGLang through the --load-format runai_streamer flag. Grove is a Kubernetes API for multi-part inference setups, such as disaggregated prefill and decode, where components need to scale together.
Why it matters
GPUs are the most expensive line item in most AI budgets, and idle capacity is wasted money. Scheduling and sharing software has quietly become as strategic as the chips themselves. By giving the platform away, NVIDIA lowers the barrier for smaller AI clouds to offer serious multi-tenant GPU services, and it makes NVIDIA’s own software the default way to run that hardware on Kubernetes.
There is a competitive angle too. Open source orchestration tools such as Kueue and Volcano already exist in the Kubernetes world, and NVIDIA’s blog notes that Grove supports Volcano as a scheduler backend. Releasing Run:ai in full gives platform teams one more mature, battle-tested option, and makes it harder for rivals to win on orchestration features alone.
The bottom line
NVIDIA has kept a promise made when it bought Run:ai: the full platform is now open source, modular and available on GitHub and NGC. If you run AI workloads on a shared Kubernetes cluster, KAI Scheduler and the GPU Fractioning Operator are worth testing first, since they can drop into an existing setup. Just plan for support yourself, because the free version doesn’t come with NVIDIA’s help desk.
The Teck Hustlers Editorial Team covers technology, business, finance and gaming news for readers who want the story behind the headline. Every article is researched, written and edited by our in-house team, checked against company announcements, filings and reputable reporting, and updated when new facts emerge. We focus on what each development means for founders, professionals, investors and everyday users. Have a tip or a correction? Reach us through our contact page.
