SkyPilot
SkyPilot's mission is to accelerate the world's most ambitious AI teams by turning fragmented AI compute into one AI supercomputer. It provides an open, provider-agnostic layer and a managed platform for running the AI lifecycle across clouds, Kubernetes, Slurm, VMs, and on-premises infrastructure.
At a Glance
- Frontier AI labs and AI-native startups
- Fortune 500 and enterprise AI teams
- ML platform and infrastructure teams
- Research organizations and universities
- +2 more
AI Tools by SkyPilot
(1)SkyPilot
AI Compute Platform Across Clouds
Discussions
No discussions yet
Be the first to start a discussion about SkyPilot
Latest News
Grace Blackwell needs a new scheduler: SkyPilot reports 27% more usable GPUs from the same racks.
Post-training a 4B model with reinforcement learning to fly the Tesla Roadster 21% better than GPT-6 Astra.
RL is Everything, Everywhere, All at Once: infrastructure requirements for modern RL post-training.
SkyPilot Agent Sessions: mission control for coding agents on customer infrastructure.
Products & Services
Open-source system and unified YAML/Python interface for running, managing, and scaling AI workloads on Kubernetes, Slurm, VMs, and 20+ cloud providers.
Fully managed AI compute platform for frontier teams, bringing heterogeneous GPU clusters and clouds under one control plane for training, inference, RL, evaluations, and production serving.
Real-time dashboard for browsing GPU SKUs, providers, regions, on-demand and spot pricing, and launching GPU clusters across 20+ clouds.
One-YAML, multi-cluster inference deployment and management layer that places replicas across available clusters and survives cluster failures.
Market Position
SkyPilot positions itself as an open-source, provider-agnostic AI compute control plane that unifies clouds, neoclouds, Kubernetes, Slurm, and on-premises resources, rather than locking teams to a single cloud or cluster scheduler. Its managed Platform adds fleet governance, reliability, utilization optimization, and support for frontier-scale workloads; SkyPilot's sandbox materials explicitly compare its bring-your-own-cluster model with hosted alternatives such as Modal.
Leadership
Founders
Zongheng Yang
UC Berkeley computer-science PhD (2022), advised by Ion Stoica, and postdoctoral researcher in Berkeley's Sky Computing Lab; he is SkyPilot's co-founder and CEO.
Zhanghao Wu
UC Berkeley AI/cloud-systems researcher and co-founder; he co-authored SkyPilot and related cloud-scheduling research, including the 2024 NSDI Outstanding Paper Award work on spot-instance savings.
Romil Bhardwaj
UC Berkeley researcher and co-founder focused on AI systems and infrastructure; he co-authored the SkyPilot intercloud-broker research and is the company's Chief Product Officer.
Ion Stoica
Databricks co-founder and UC Berkeley professor; a long-time distributed-systems and cloud-computing researcher and SkyPilot co-founder.
Scott Shenker
UC Berkeley professor and networking/cloud-computing pioneer; he co-founded SkyPilot with the Berkeley research team.
Executive Team
Zongheng Yang
Co-Founder & CEO
Berkeley computer-science PhD and former postdoctoral researcher in the Sky Computing Lab; creator of the original SkyPilot research direction.
Zhanghao Wu
Co-Founder & CTO
UC Berkeley cloud and AI-systems researcher and co-author of SkyPilot's intercloud and scheduling research.
Board of Directors
Founding Story
SkyPilot began as an open-source project in UC Berkeley's Sky Computing Lab. The researchers had compute credits from several providers, but lab users avoided non-AWS resources because each provider required a different setup and migrating workloads and state was too difficult; they set out to unify all compute into a single, provider-agnostic pool (a 'sky' of cloud compute).
Business Model
Revenue Model
Open-source software is available for self-managed use; SkyPilot Platform is a commercial, fully managed enterprise offering for customers bringing their own compute, with demo-based access and managed support. The company also offers select on-premises deployments.
Pricing Tiers
Self-managed open-source software; customers pay their own underlying cloud or infrastructure costs.
Managed enterprise platform with team controls, quotas, SSO/RBAC, HA, support, and production fleet capabilities; public page directs teams to request a demo.
Target Markets
- Frontier AI labs and AI-native startups
- Fortune 500 and enterprise AI teams
- ML platform and infrastructure teams
- Research organizations and universities
- Healthcare, banking, robotics, and other data-intensive industries
- Organizations operating multi-cloud, Kubernetes, Slurm, or on-premises GPU fleets
- Large-scale model pretraining and distributed training
- Online reinforcement learning and agent training
- LLM inference and production multi-cluster serving
- Batch inference, model evaluation, and hyperparameter sweeps
- GPU development environments and coding-agent workloads
- Running untrusted LLM-generated code in isolated sandboxes
- Abridge
- Applied Compute
- H Company
- Nubank