SkyPilot
7.4/10Open-source and enterprise system for running, scheduling and scaling AI training, inference and batch jobs across Kubernetes, Slurm, on-prem and 20+ clouds.
Data last updated November 2026 · Our methodology
Overview
SkyPilot started as a UC Berkeley research project and is now sold by SkyPilot Inc. of San Francisco. It is an orchestration layer for AI compute, not a hosting provider. Teams describe GPU, TPU or CPU jobs in YAML, and SkyPilot provisions the resources in their own accounts. The documentation lists 25 supported backends, including Kubernetes, Slurm, AWS, GCP, Azure, OCI, CoreWeave, Lambda and RunPod. The scheduler chooses available or cheaper capacity and fails over to other regions or clouds when GPUs run out. Managed jobs recover from spot preemptions and crashes, and Kubernetes clusters get gang scheduling, queueing and multi-cluster support. Idle resources can be stopped automatically. Developers can SSH into machines or connect an IDE. SkyServe adds model endpoints with autoscaling. There are two editions. The open-source project is free under Apache 2.0, and you run its API server yourself. The commercial SkyPilot Platform adds SSO, RBAC, workspaces, quotas, cost reporting, GPU health monitoring with automatic remediation, sandboxes, airgapped deployment and SOC 2 Type II compliance. The vendor publishes no prices for the platform, so buyers must request a demo. You still pay your cloud providers for the compute you use. It best suits ML and platform teams at startups and enterprises that want one interface across several GPU clouds or clusters. Its main limitation is scope. Checkpointing has to be written into your own training code, and it is not built to manage long-running Kubernetes services. Moving data between clouds can also cancel out the savings from picking cheaper GPUs.
Our Rating
AI Software7.6/10
Hosting & Infrastructure7.3/10
Who SkyPilot is best for
Best software for DevelopersGood fit
ML engineers define jobs in YAML through a CLI, with SSH and IDE access to the machines.
The open-source edition is free, and users report using it to burn through startup cloud credits from several providers.
The paid platform adds SSO, RBAC, airgapped deployment and SOC 2 Type II, but it has only recently become commercially available.
Infrastructure teams get one control plane with quotas, priority queueing and GPU health checks.
Pros & Cons
Pros
- One YAML job definition runs on AWS, GCP, Azure, Kubernetes, Slurm and neoclouds such as CoreWeave, Lambda and RunPod.
- When GPU capacity is unavailable, it automatically fails over to other regions or clouds and favours cheaper matching instances.
- Managed jobs recover from spot preemptions and crashes, which cuts down on manual babysitting.
- The core is free and open source under Apache 2.0 and runs inside your own cloud accounts.
- The enterprise edition adds SSO, RBAC, quotas and SOC 2 Type II for teams with governance needs.
Cons
- Checkpointing must be built into your training code, or a preempted job starts over from the beginning.
- It is not designed to manage long-running Kubernetes Deployments or StatefulSets.
- Moving data between clouds can erase the savings from choosing cheaper GPUs.
- The open-source edition requires you to run and maintain your own API server.
- The vendor publishes no pricing for the commercial platform.
Key Features
SkyPilot Pricing
Open Source
- Apache 2.0 license
- Self-hosted API server
- Autostop for idle compute
- Managed jobs with preemption recovery
- Supports 20+ clouds plus Kubernetes and Slurm
- Cloud compute billed separately by your providers
SkyPilot Platform (Enterprise)
- SOC 2 Type II
- Sandboxes for AI agents
- SSO, RBAC and workspaces
- Contact sales or request a demo
- BYOC, private VPC and airgapped deployment
- GPU health monitoring with auto-remediation
- Quotas, policy enforcement and cost reporting
Frequently Asked Questions
No. SkyPilot orchestrates compute in your own cloud accounts, Kubernetes clusters, Slurm clusters or on-prem machines, so your providers bill you for the compute directly.
The open-source project is free under the Apache 2.0 license. The commercial SkyPilot Platform has no public pricing, and you request a demo to get a quote.
The documentation lists 25 backends, including Kubernetes, Slurm, AWS, GCP, Azure, OCI, CoreWeave, Lambda Cloud and RunPod.
The vendor shows a SOC 2 Type II badge for the SkyPilot Platform. The platform also supports SSO, RBAC and airgapped deployment.
Yes. SkyServe deploys model endpoints with autoscaling, a gateway and TLS, but one competitor's comparison describes its cost optimization for services as basic.
UC Berkeley researchers, including CEO Zongheng Yang and Databricks co-founder Ion Stoica, founded the company. It raised a $20M seed round led by Lux Capital in July 2026.
Similar Tools
See all alternatives
Bare metal cloud provider that offers dedicated servers, GPU servers and virtual machines on demand, billed hourly or monthly, for developers and infrastructure-heavy businesses.
Amazon's public cloud platform, with over 200 pay-as-you-go services for compute, storage, databases, networking and AI, used by startups through large enterprises and governments.
Northflank is a Kubernetes-based developer platform for deploying services, jobs, databases and GPU workloads on its managed cloud or on customers' own cloud accounts.
Ranked near SkyPilot in AI Software
See SkyPilot in the listSkyPilot is #3 in AI Software by our score.
