SkyPilot
Open-source AI compute platform that runs training, RL, batch inference and endpoints across Kubernetes, Slurm and 20+ clouds
Overview
SkyPilot pitches "one platform for all your AI compute: Kubernetes, Slurm, 20+ clouds." You describe a job once and SkyPilot finds capacity across your clusters and cloud accounts, then handles launch, queueing and cleanup. The current product covers development boxes, pre-training, post-training and RL, batch inference, sandboxes and endpoints, with GPU monitoring, quotas, priority queueing, GPU sharing and health checks with auto-remediation. It began in 2022 at UC Berkeley's Sky Computing Lab (the research behind it appeared at NSDI 2023). The open-source core is Apache-2.0 (latest v0.14.2, October 28, 2026); in July 2026 the team launched a company with a $20M seed led by Lux Capital and sells SkyPilot Platform, an enterprise tier with BYOC, air-gapped installs, SSO, RBAC and cost reporting. The old skypilot.co address now redirects to skypilot.ai.
The Verdict
Who Should Use SkyPilot?
Best For
- Teams chasing scarce GPUs across several clouds, neoclouds and on-prem clusters
- Research groups that mix Kubernetes and Slurm and want one CLI
- Training and RL runs that need to move to wherever capacity is cheapest
Not Ideal For
- Experiment tracking or model registry (pair it with MLflow or W&B)
- Teams on a single cloud happy with that cloud's own ML service
- Buyers who need a public price for the enterprise tier
What's Great
- One job spec runs on 20+ clouds, Kubernetes and Slurm
- Free Apache-2.0 core with frequent releases
- Large-scale users named on the site, including NVIDIA, Meta, Mistral and Nubank
- Near-zero-config launch of multi-GPU nodes, per a practitioner write-up
Watch Out For
- Dashboard was slow under heavy use (later releases improved API server speed)
- Enterprise features (SSO, RBAC, air-gapped) are in the paid platform with no public price
- Few independent reviews yet; no G2 profile found
Pricing
View all features & details
Key Features
- Single CLI and scheduler across clusters and clouds
- GPU monitoring, quotas and priority queueing
- GPU sharing and health checks with auto-remediation
- Workspaces and cost reporting (Platform)
- SOC 2 Type II (Platform)
Workloads
- Development environments
- Pre-training, post-training and RL
- Batch inference
- Sandboxes and endpoints
Real-World Usage
Named Users
- Meta FAIR (large Slurm fleet), NVIDIA, Mistral, Amazon Robotics
- H Company: online RL on 2,000+ GPUs
- Nubank, Abridge, Hippocratic AI, Sarvam
Company
- Out of UC Berkeley's Sky Computing Lab; Ion Stoica involved
- $20M seed led by Lux Capital, July 2026
How It Compares
| Feature | SkyPilot | Ray | Kubeflow | Modal |
|---|---|---|---|---|
| Main job | Find and run compute across clouds and clusters | Distributed Python compute engine | Kubernetes ML platform | Serverless GPU cloud |
| Where it runs | 20+ clouds, Kubernetes, Slurm | Any cluster | Kubernetes | Modal’s cloud |
| Open source | Yes, Apache-2.0 | Yes, Apache-2.0 | Yes, Apache-2.0 | No |
| Paid option | SkyPilot Platform (custom) | Anyscale (usage-based) | Managed distributions | Hosted service only |
Sources: skypilot.ai · GitHub · Releases · The Next Web
Is this your tool? Add a badge