SkyPilot is an open-source platform for running AI workloads on any cloud, Kubernetes cluster or on-prem GPU pool through one unified interface. Describe a job once in YAML, and SkyPilot finds capacity, provisions machines, syncs your code and data, and runs it โ across more than twenty infrastructure providers. It has about 10,700 stars on GitHub and is in daily active development in 2026.
Frontier AI teams use SkyPilot to escape cloud fragmentation: training runs, batch inference, LLM serving and development environments all launch with the same commands regardless of whether the GPUs live on AWS, GCP, Azure, Lambda, RunPod, Nebius or a local Slurm cluster. Its optimizer hunts for the cheapest available accelerators, including spot and preemptible instances, with automatic recovery when instances are reclaimed. Managed jobs queue and retry unattended, volumes and file mounts sync data between clouds, and a dashboard helps compare GPU pricing across providers. The project grew out of UC Berkeley’s Sky Computing Lab and underpins production training and serving at many AI companies. SkyPilot is vendor-neutral by design, with a live global catalogue tracking capacity and prices across providers, and teams also run it as an internal platform layer that gives researchers self-serve access to pooled GPUs with quotas and shared storage. It originated in UC Berkeley’s Sky Computing Lab, whose earlier work on cloud optimization shapes the project’s cost-first philosophy throughout.
SkyPilot is free and open source under the Apache 2.0 license. You pay only your cloud providers for the infrastructure it provisions.
SkyPilot is for ML engineers and platform teams who train or serve models across multiple clouds and want lower GPU bills without rewriting jobs for every provider.