skypilot-org/skypilot

The AI Compute Platform for frontier teams. SkyPilot turns fragmented AI compute into one AI supercomputer, so frontier AI teams build custom intelligence faster.

What it solves

SkyPilot simplifies the process of running, managing, and scaling AI workloads across diverse infrastructure. It removes vendor lock-in by providing a unified interface for deploying jobs on various cloud providers, Kubernetes clusters, and Slurm clusters, while optimizing for cost and resource availability.

How it works

Users define their AI tasks using a unified YAML or Python API, specifying resource requirements (like GPUs/TPUs), data synchronization needs, and setup/run commands. SkyPilot then automates the heavy lifting: finding the cheapest available infrastructure, provisioning the necessary compute (VMs or pods), syncing the codebase, installing dependencies, and executing the job.

Who it’s for

It is designed for AI teams who need a simple, portable way to launch jobs and for infrastructure teams who require a unified control plane to manage AI compute, scheduling, and orchestration across multiple clouds and clusters.

Highlights

  • Multi-cloud and Multi-cluster Support: Works across 20+ clouds (AWS, GCP, Azure, etc.), Kubernetes, and Slurm.
  • GPU Fleet Optimization: Includes features like Autostop for idle resources, binpacking on shared clusters, and an intelligent scheduler for availability.
  • Kubernetes Enhancement: Adds AI-native capabilities to Kubernetes, such as gang scheduling, multi-node jobs, and simplified interactive development (SSH/IDE connection).
  • AI Agent Integration: Provides a "SkyPilot Skill" allowing AI agents (like Claude Code) to manage GPU access and jobs directly.
  • Sandboxes: Ability to run untrusted, LLM-generated code on existing Kubernetes clusters.