NVIDIA Run:ai
lastmod 2026-11-11created 2026-11-111 edits
A Kubernetes-based platform for managing GPUs shared across multiple teams
About
- An open-source platform from NVIDIA that schedules AI development, training, and inference workloads and allocates resources on GPU infrastructure shared by multiple teams
- Lets multiple workloads share a single GPU and makes dynamic use of idle capacity to reduce unused GPUs
- Places workloads based on priority, fair-share allocation, resource reclamation, and GPU interconnect topology
- Sets GPU quotas and usage policies per department and project so that no single team monopolizes resources
- Manages workloads and resources across multiple Kubernetes clusters from one place; supports on-premises, cloud, and hybrid deployment
- Covers development workspaces, single-node and distributed training, and inference serving; integrates with Kubeflow, Ray, and KServe
- Provides a web UI, CLI, and REST API. An MCP server lets AI assistants run workloads
- Supports resource usage monitoring per GPU, node, and workload, along with role-based access control
- Installs directly on Kubernetes or OpenShift, with support for air-gapped deployment. The open-source distribution can be installed without NGC authentication
- Written in Go. Apache-2.0 license