NVIDIA open-sources Run workload and GPU orchestration components

NVIDIA’s full Kubernetes-based Run platform is now open source, with modular GPU scheduling, sharing and resource management. Community availability does not include NVIDIA enterprise support.

NVIDIA open-sources Run workload and GPU orchestration components in a release dated 13 October 2026, making the full Kubernetes-based platform available through its DSX AI Factory initiative. The code includes scheduling, fractional GPU sharing, organisational resource allocation and workload management. Operators can deploy the integrated platform or adopt individual components without purchasing a commercial Run subscription.

In its official developer announcement, NVIDIA identified four independently adoptable modules: KAI Scheduler, GPU Fractioning Operator, KAI Resource Management and Workload Map. The release covers software for allocating GPU capacity between teams and coordinating AI jobs, not new GPU hardware.

KAI Scheduler allocates jobs and GPU capacity

KAI Scheduler handles GPU-aware placement in Kubernetes, including fair-share scheduling, quotas, topology awareness and gang scheduling for jobs whose components need to run together. The scheduler was already open source under Apache 2.0 and part of the Cloud Native Computing Foundation’s Sandbox programme; the October announcement opens the wider Run platform.

GPU Fractioning Operator lets multiple suitable workloads use portions of a GPU through CUDA 13.4-native memory isolation. It integrates with KAI but can also work with other schedulers. Fractional allocation is relevant for notebooks and smaller inference jobs that do not require an entire accelerator.

Organisational controls sit above scheduling

KAI Resource Management defines departments, projects, quotas and node pools, enabling platform teams to distribute available GPUs according to organisational rules. Workload Map gives infrastructure tools a common description of jobs such as Ray clusters, training tasks and inference services.

Workload Map is a library for recognising workload structure and lifecycle state. It does not independently schedule or autoscale jobs, a distinction important for teams adopting it without the full platform.

Open-source access does not include enterprise support

NVIDIA says the full Run source is published in its DSX AI Factory GitHub repository. An integrated Community Product is also distributed through NVIDIA NGC with Helm charts for installation. Although no paid subscription is required for that community offering, its downloads have applicable NVIDIA terms, and enterprise support is not included.

Supported commercial Run products remain available separately, including through NVIDIA AI Enterprise and partner channels. NVIDIA named eight DSX ecosystem partners, including Crusoe, Nebius and Mirantis, integrating the software into their offerings; that does not establish identical rollout status for every partner.

Related open-source projects Model Streamer and Grove address model loading and multicomponent inference respectively. They complement this release rather than being the four core Run modules.