SkyPilot

Teams use SkyPilot to run, manage, and scale AI workloads on whatever infrastructure they have: Kubernetes clusters, Slurm-based HPC systems, and more than 20 cloud providers including AWS, GCP, Azure, OCI, CoreWeave, Lambda Cloud, and RunPod. A job is declared once, in YAML or Python, with its resource requirements, setup commands, and run script, and SkyPilot provisions it anywhere without code changes, automatically choosing the cheapest available option and failing over to other regions or clouds when GPU capacity runs out. Operational features include gang scheduling for multi-node training, workload binpacking on shared clusters, autostop for idle resources, managed spot instances, and a single control plane spanning multiple clusters, plus interactive workflows like SSH access to pods and IDE integration for development. Originally from UC Berkeley's Sky Computing Lab and released as open source, the project is used in production by companies such as Shopify for training workloads, and appeals to ML platform teams trying to cut GPU costs and avoid single-vendor lock-in.