SkyPilot

SkyPilot is an open-source AI compute platform that unifies Kubernetes clusters, Slurm clusters, on-premises machines and more than 25 cloud providers into a single pool, so AI and platform teams can launch training, fine-tuning, batch inference and serving workloads anywhere GPUs are actually available instead of hand-managing each environment separately.

SkyPilot is an AI compute platform that presents Kubernetes clusters, Slurm clusters, SSH-accessible on-premises machines and more than 25 cloud providers as a single pool of infrastructure, so an AI team can launch a training run, a fine-tuning job, a batch inference sweep or an agent sandbox with one YAML file or one Python call and let the system decide where it actually lands. It began as an open-source project in the UC Berkeley Sky Computing Lab, the same lab behind Spark and Ray, and the Apache-2.0 repository carries roughly 10,400 GitHub stars, with v0.12.0 in March 2026 adding Slurm support, job groups and improved data mounting. The core abstractions are clusters (virtual collections of VMs or pods you can SSH into and attach an IDE to), jobs (sky exec on an existing cluster, or sky jobs launch for managed jobs that auto-recover and fail over to another region or provider when capacity disappears) and services for multi-replica model serving spread across locations and pricing tiers. SkyPilot Inc. came out of stealth on 21 July 2026 with a $20 million seed round led by Lux Capital, alongside a commercial SkyPilot Platform that layers gang scheduling, priority queues, GPU sharing and quotas, GPU health checks with auto-remediation, sub-second inference sandboxes, SSO, RBAC, cost reporting and a unified dashboard on top of the open-source core. The vendor lists Nubank, Meta, NVIDIA, Amazon Robotics, Shopify, Mistral, Abridge, HeyGen and Hippocratic AI among more than 30 enterprises using it.

Category

Infrastructure & Cloud

Pricing

Freemium, Contact for pricing

Target Market

CTOs, VPs of Engineering, Platform Engineers, ML Infrastructure Teams, Data Scientists, Enterprise Developers

Deployment

Open-source, Self-hosted, Hybrid, Multi-cloud

Headquarters

San Francisco, United States

Customers

30+ enterprises per the vendor, including Nubank, Meta, NVIDIA, Amazon Robotics, Shopify, Mistral, Abridge, HeyGen and Hippocratic AI

• ✓
  Unified multi-infrastructure interface

  Presents Kubernetes, Slurm, SSH node pools and 25+ clouds as one compute pool addressed through a single YAML or Python API.

• ✓
  Managed jobs with auto-recovery and failover

  sky jobs launch retries and relocates a job to another region or provider when nodes are preempted or capacity disappears mid-run.

• ✓
  Advanced scheduling

  Gang scheduling, multi-node jobs, priority queueing, GPU sharing, quotas and workload binpacking to raise utilisation on fleets you already pay for.

• ✓
  GPU Manager health checks

  Commercial platform monitors accelerator health and auto-remediates bad nodes, which is the failure mode that silently wastes large training runs.

• ✓
  Services for model serving

  Runs multi-replica inference endpoints spread across regions and pricing tiers, with autoscaling and failover between them.

• ✓
  Sandboxes for agent code

  Launches isolated inference and agent execution sandboxes on your own Kubernetes in under a second, per the vendor.

• ✓
  Team governance controls

  SSO, RBAC, policy enforcement, per-team cost reporting and a shared dashboard via a team-wide API server deployment.

• •
  Chasing scarce GPU capacity across providers

  Submit one job definition and let SkyPilot find available H100 or B200 capacity across clouds, regions and neoclouds rather than manually polling each console.

• •
  Fault-tolerant pre-training and RL runs

  Run multi-day distributed training on spot or preemptible capacity, with managed jobs restarting and relocating the run automatically after node loss.

• •
  Consolidating Kubernetes and Slurm

  Give research users one submission path while the platform team keeps central scheduling, quota and cost visibility across both orchestrators.

• •
  Raising utilisation on an existing GPU fleet

  Apply binpacking, priority queues and idle-resource cleanup to squeeze more throughput from committed capacity, which the CEO frames as a 10%-plus utilisation gain.

• •
  Portable batch inference and evaluation sweeps

  Fan a large evaluation or embedding job out across whichever clusters and cloud accounts are cheapest and idle at the time.

The open-source core is Apache-2.0 and free to self-host with no seat or node fee. SkyPilot Platform publishes no list price at all: the site routes every commercial enquiry to a demo booking, so expect a negotiated enterprise agreement. Either way SkyPilot bills nothing for compute itself. You keep paying your clouds, neoclouds and Kubernetes providers directly, and the platform's pitch is that better utilisation of that existing spend more than covers its own cost.

This page was written from 8 sources, 7 on domains other than skypilot.ai.

1. 1.siliconangle.com — skypilot nabs 20m ease ai infrastructure management
2. 2.fortune.com — skypilot from databricks cofounder raises 20m to be the swit
3. 3.github.com — skypilot
4. 4.github.com — issues
5. 5.hn.algolia.com — search
6. 6.hpcwire.com — skypilot launches with 20m to accelerate custom intelligence
7. 7.skypilot.ai — skypilot.aivendor
8. 8.docs.skypilot.ai — overview