Every hour an AI team spends fighting infrastructure is an hour the frontier doesn't move.
Today, SkyPilot is out of stealth. We help frontier AI teams build intelligence faster by removing their biggest bottleneck: AI compute fragmentation.
Custom intelligence is stalled by fragmented AI compute # Shortly before ChatGPT's release, I was a PhD researcher at Berkeley in the lab that incubated Databricks and Spark, running lots of training jobs on GPUs from AWS. We used AWS because they gave our lab free compute.
The problem was, everyone in the lab was using AWS, too. The lucky ones sometimes got GPUs. The unlucky got no-capacity errors. It was our first glimpse into the GPU shortage.
Our lab did receive credits from other providers. GCP, Azure, and some unknown "specialized clouds" (we call them AI neoclouds now). No one used them. Each provider's setup was a 10-step wiki page no one wanted to follow. New concepts, new APIs, but the worst part was migrating our precious workloads and state. It's too complicated.
We wanted to use one system to manage all our AI compute. Not five.
So we started building it in open source. We called it SkyPilot. We asked: Can we build a system to unify all our compute into a single pool — a "sky" of cloud compute?
Fast forward to 2026, SkyPilot is used by hundreds of companies. Working with them, it became clear the problem we saw is now orders-of-magnitude more severe in industry.
Since GPU demand far outpaces supply, AI teams have to get GPUs anywhere they can. They then firefight fragmented *compute — *three hyperscalers, six neoclouds, a dozen Kubernetes clusters, many accelerator SKUs. Researchers burn time on siloed workload setup (yes, migrating workloads or state is still a pain). Infra gets paged when GPUs go down. Frontier AI teams move slowly even on the fastest compute ever built.
Companies put up with this because they have to build *custom intelligence — *agents, apps, and models that are trained on each company's data, workflows, and knowledge. Intelligence only they can build. Intelligence that beat closed models on quality and cost.
**We believe the next winners in AI will win by custom intelligence. **They will control their AI stack: their compute, data, and intelligence. They will need more compute. Not less.
To accelerate building custom intelligence for every organization, an open, provider-agnostic AI compute layer capable of running frontier workloads is needed.
We built SkyPilot for this reality.
Turning fragmented compute into one AI supercomputer # SkyPilot is a control plane that turns fragmented compute into one "AI supercomputer". It abstracts your AI compute — across neoclouds or hyperscalers — into one pool of resources, so they can be centrally managed and efficiently used.
While SkyPilot started as a research-y open source project, it's now running a meaningful share of AI compute across hundreds of leading companies:
- Top deployments surged past 1,000+ nodes and10,000+ GPUs - GPU hours consumed on SkyPilot grew 35% MoM, 6x in last 6 months - 14M+ downloads (~6M in last 3 months alone)
- 280+ contributors
SkyPilot is BYOC (Bring Your Own Compute). It offers native support for most AI workloads: interactive development, jobs, batch inference, evals, large-scale training, and RL. All workloads run on your compute, with your frameworks of choice.
Frontier teams use SkyPilot to build custom intelligence # Hundreds of leading organizations already run on SkyPilot — from top neolabs to Fortune 500 enterprises. They're building custom intelligence that leads their vertical, with some users seeing 10x faster time-to-intelligence and double-digit increase in GPU utilization.
Abridge is defining the standard for AI in healthcare, trusted by 300+ major health systems in the US. With SkyPilot, they achieved 10x faster AI experimentation:
"SkyPilot is easy to setup, light weight to use and configurable in just the right places. We can now launch ten experiments in the time it used to take me to set up one."
Applied Compute is at the frontier of building intelligence for enterprises. Their team trains RL models and agents across clouds using SkyPilot:
"We originally spent some time investing in our own home-grown GPU scheduling solution, but SkyPilot gave us all of those features and more out of the box - gang scheduling, multi-node jobs, SSH access - in a unified interface across all our clouds. Scaling to new GPU clusters now takes minutes instead of weeks, and our researchers launch 1000s of jobs across all our clouds in seconds."
H Company is one of Europe's leading AI labs, training computer-use agents that are pushing the boundaries of agents. They use SkyPilot to run RL across 2,000+ GPUs:
"SkyPilot is now our standard AI infrastructure layer, powering all our training and enabling us to scale online RL to 2,000+ GPUs on K8s, previously impossible on Slurm."
Nubank is one of the world's largest digital banks, serving over 100 million customers across Latin America. Nubank's AI team runs AI workloads at scale on SkyPilot:
"SkyPilot provides a flexible orchestration layer that adapts to Nubank’s infrastructure and operational requirements, rather than constraining us to a rigid platform. Equally valuable is the strength of the SkyPilot community and the responsiveness of its support. Together, these have made SkyPilot a foundational part of how we run AI workloads at scale."
Announcing SkyPilot Platform # We punted off building a commercial product for years. The open source "worked".
But the same questions kept being raised by frontier teams. They love using SkyPilot, but can they also get optimizations to maximize GPU fleet utilization, the performance to scale to 1,000s or 10,000s of GPUs, high availability in production, multi-project governance, and a 24/7 managed experience?
So today, we're announcing SkyPilot Platform — the AI compute platform for frontier AI teams. It is optimized for managing large GPU fleets and frontier workloads.
**One platform for all your AI compute. **Bring all your GPU clusters across neoclouds and hyperscalers and get standardized GPU management: validation, monitoring, scheduling, and utilization optimizations. All your AI compute, under one managed control plane.
Frontier workload support. SkyPilot Platform speeds up building custom intelligence:
AI-native development— GPU devboxes for humans or coding agent fleets** Pre-training at scale**— 1,000+ node hero runs with minimum interruption— fast sandboxes for RL rollouts and agentsSandboxes— deploy your models on your computeProduction multi-cluster serving
Enterprise-ready. High Availability (HA), team and quota management, SSO and RBAC, and SOC 2 compliance are available.
20x faster than open source. Customers have seen 20x performance improvements over SkyPilot open source, which means you move faster and scale to more workloads.
Some customers are using SkyPilot Platform to manage 10K+ GPUs and support 200+ researchers. Open source users can switch to the platform with a server URL change. We're now opening access to select new customers. Request a demo → to get started.
Our cloud partners # The next decade of AI won't run on one cloud. The frontier AI teams know this. That's why the top neoclouds and hyperscalers have partnered with us — because their customers, running the most demanding AI workloads in the world, need to move freely across providers. SkyPilot ships with support for 20+ clouds. Here are some examples:
Nebius actively contributes to SkyPilot open source and offers seamless integrations for joint customersCoreWeave offers SkyPilot on CoreWeave Kubernetes Services (CKS) and SUNKLambda's Managed Kubernetes and cloud API are natively supported by SkyPilot** AWS**has SkyPilot integrations for EKS, EC2, S3, and SageMaker Hyperpod
"AI teams need to scale from one experiment to thousands of GPUs without rebuilding their workflow.SkyPilot makes that practical with a portable orchestration layer across clouds, regions, and GPU pools, turning scarce capacity into usable AI infrastructure."
These partnerships mean we are able to offer a plug-and-play experience for frontier AI teams — regardless of what providers they choose.
Our mission and $20M funding # Our mission is to accelerate the world's most ambitious AI teams. Every hour they spend fighting infrastructure is an hour the frontier doesn't move. We intend to give them those hours back.
We’ve raised over $20M led by Lux, with participation from Amplify, Coatue, Foundation, Race, The House Fund, and some of the industry's top AI operators like Ali Ghodsi (CEO, Databricks), Jeff Dean (Chief Scientist, Google), Guillermo Rauch (CEO, Vercel), Amjad Masad (CEO, Replit), Clem Delangue (CEO, HuggingFace), and more.
We're hiring in Engineering and GTM to deliver the best platform for the AI frontier.
If you firefight AI compute, let's build. — Zongheng, Romil, Zhanghao, Ion, Scott