SkyPilot

What it is for

SkyPilot is an Apache-2.0 AI compute control plane: write a YAML, sky launch it onto Kubernetes / Slurm / a cloud GPU, then optionally serve or queue managed jobs. Use it when the job is run training or inference on infra you already pay, not when you need a hosted chat seat.

What makes it unique: the cover is SkyPilot’s official docs dashboard still — Sky Clusters with training-multinode / dev-alice / inference on Kubernetes and AWS — not a pricing card.

Key features

• Launch — sky launch -c mycluster hello_sky.yaml provisions (or reuses) a cluster, syncs workdir, runs setup then run (Quickstart). One job; AWS / GCP / K8s / Slurm targets are infra SKU, not five products.
• Managed jobs — sky jobs launch puts the same YAML on auto-recovered clusters so you can fire 10s–1000s of runs (Managed jobs; Quickstart “Scaling out”). A queue, not an interactive cluster.
• SkyServe — sky serve up service.yaml load-balances replicas across regions/clouds (sky serve status / sky serve logs) (Serving Models). A service, not a train. Endpoints (PD-disagg, scale-to-zero) is the next-gen of this in early access — not a second public product.
• Dashboard — sky dashboard lists every cluster and managed job across clouds (Quickstart “Access the dashboard”; official clusters still). An ops desk, not a launch.
• Team API server — a shared SkyPilot API so the org sees one fleet, with optional SSO / RBAC on the commercial Platform (Team Deployment; Platform). A control plane, not a job.

uv pip install "skypilot[...]" / sky --help are install. You bring the GPU / cloud bill — that is the meter. GPU type names (B200 / H100) are SKU.

Who it's for

Best fit: ML infra teams who already have GPUs or cloud accounts and will run sky launch (or the Hello Sky notebook), not a team looking for a $20 chat plan.

Pricing

Verified 2026-10-22 (Asia/Shanghai) from the Apache-2.0 LICENSE, docs.skypilot.co, and SkyPilot Platform. There is no hosted SKU table.

• Software (OSS) — Apache 2.0. License fee $0. You pay AWS / GCP / RunPod / the Kubernetes bill.
• SkyPilot Platform — limited early access. Request a demo. No public seat dollar.

Do not invent a SkyPilot Cloud seat. GPU Compass (gpus.skypilot.co) compares other clouds’ published $/hr — those dollars are not SkyPilot’s.

Limitations / costs

• You bring the GPU. The CLI is $0; H100 hours are not.
• Platform has no printed $. Confirm the quote.
• SkyServe is beta for internal/R&D; production Endpoints is early access.
• A control plane, not a GPU marketplace. For a published $/hr card, see Runpod or Salad.

Alternatives

• Runpod — a GPU marketplace when you want to buy hours, not orchestrate ones you already have.
• Modal — serverless functions when you do not want to manage clusters.
• vLLM — the serving engine SkyServe can deploy, not the control plane.

Official tutorial

Official SkyPilot channel exists. The long meetup talk is not a first-hour lesson we embed. Use written Quickstart (sky launch -c mycluster hello_sky.yaml) plus Managed jobs and Serving Models.

Website

https://skypilot.ai

Repository

https://github.com/skypilot-org/skypilot

Source

Catalog row curated for AI Tools Nav. Cover is SkyPilot’s official docs Clusters dashboard (training-multinode on Kubernetes). License fee $0; Platform custom, no seat table. No first-party first-hour YouTube embed.

Last verified

2026-10-22 (Asia/Shanghai) editorial check.