NVIDIA announced on October 13, 2026, that it had released the complete Run:ai platform as open-source software, including its control plane and cluster services. Run:ai is a Kubernetes platform for orchestrating AI workloads and managing GPU resources. For infrastructure teams, the key distinction is that the release includes the full platform, while operators still supply and run the underlying infrastructure. NVIDIA’s announcement NVIDIA’s installation documentation
The project’s official repository identifies Apache License 2.0 for its software code. It also lists separate licenses for some nested third-party code and assets, so the repository’s license statement does not apply uniformly to every bundled item.
What each Run:ai component does
NVIDIA identifies four components that organizations can adopt independently. They address different parts of AI infrastructure rather than acting as interchangeable schedulers.
| Component | Function | Boundary or relationship |
| KAI Scheduler | Places GPU workloads on Kubernetes. NVIDIA lists gang scheduling, fair-share and quotas, topology-aware placement, GPU sharing, and Dynamic Resource Allocation support. | Can be deployed in an existing cluster without the complete Run:ai platform. |
| GPU Fractioning Operator | Integrates CUDA-native GPU fractions with Kubernetes and provides per-pod metrics. | Scheduler-agnostic; the CUDA driver enforces a memory limit for each workload’s fraction. |
| KAI Resource Management | Organizes projects and departments and manages GPU allocation through guarantees, sharing, fair-share, quotas, and node pools. | Can be combined with KAI Scheduler. |
| Workload Map | Provides a common way to describe and manage Kubernetes AI workload types, including their status, health, and policy-related details. | An offline-capable Go library; it does not itself schedule, queue, autoscale, or propagate workloads across clusters. |
These roles help identify which layer addresses a particular infrastructure need: scheduling and placement, fractional GPU allocation, organizational resource policies, or a shared description of workloads. Workload Map supplies that description and management contract; scheduling remains outside its own function. NVIDIA describes the components and their boundaries in its release announcement.
How a self-hosted Run:ai deployment is arranged
Every installation has a control plane and at least one Run:ai cluster. The control plane centralizes resource management, workload submission, monitoring, and analytics; cluster services extend scheduling and workload management. NVIDIA documents two arrangements: the control plane and first cluster can share a Kubernetes cluster, or the control plane can run on a separate cluster connected to one or more Run:ai clusters over a network. NVIDIA’s installation overview describes both topologies.
NVIDIA documents deployments on Kubernetes or OpenShift, in connected or air-gapped environments. Its listed installation paths are Guided Installation, Manual Helm, and Base Command Manager. NVIDIA lists Base Command Manager for the Support-Entitled Release only.
NVIDIA’s installation documentation also identifies GPU Operator and Prometheus among the prerequisites. The topology choice therefore affects how the control plane and cluster services are connected, while the deployment still depends on the operator’s environment and supporting services.
What access includes—and what operators provide
NVIDIA says its NGC Community Product requires no paid Run:ai subscription or support entitlement. Open-source platform images and Helm charts can be accessed from NGC without an API key, but downloading the Community Product requires acceptance of applicable NVIDIA software and product-specific terms. NVIDIA enterprise support is not included.
The Support-Entitled Release is the same underlying platform with support service-level agreements and other entitlements; NVIDIA documentation says it requires an NGC API key. NVIDIA also lists commercial routes through standalone Run:ai, NVIDIA Mission Control, NVIDIA AI Enterprise, and selected cloud marketplaces, with availability, licensing, and support terms varying by offering. The access terms and support distinction are described in NVIDIA’s announcement and installation documentation.
Self-hosting also leaves the underlying environment with the operator: hardware, operating system, Kubernetes distribution, networking, and storage. NVIDIA’s documentation assigns responsibility for infrastructure and supporting services to the operator, while identifying Run:ai platform images and Helm charts as NVIDIA’s responsibility.
