Zongheng Yang launches SkyPilot with $20M to unify fragmented AI compute

Why it matters

AI teams are assembling GPU capacity across incompatible providers and clusters. SkyPilot is betting the control plane that coordinates that supply can become a valuable software layer independent of any cloud or chip vendor.

Zongheng Yang and Ion Stoica launched SkyPilot on Tuesday with a $20 million seed led by Lux Capital, turning a UC Berkeley open-source project into a commercial control plane for AI teams running GPUs across multiple clouds and clusters.

The financing, first reported by Fortune, included Coatue and Amplify Partners. SkyPilot's announcement also named Foundation Capital, Race Capital and The House Fund as investors. Operator angels include Databricks CEO Ali Ghodsi, Google Chief Scientist Jeff Dean, Vercel CEO Guillermo Rauch, Replit CEO Amjad Masad, Hugging Face CEO Clem Delangue and dbt Labs CEO Tristan Handy. SkyPilot did not disclose its valuation, revenue, pricing or headcount.

Yang is SkyPilot's CEO and the founder responsible for turning the research into a business. He earned a UC Berkeley computer science PhD under Stoica, worked on TensorFlow and deep-learning research at Google Brain, and interned at Databricks when it had roughly 10 employees, according to Fortune. The other founders are CTO Zhanghao Wu, CPO Romil Bhardwaj, board chairman Stoica and Berkeley networking researcher Scott Shenker.

That division of roles matters. Fortune's headline trades on Stoica's record as a cofounder of Databricks, Anyscale and Conviva. SkyPilot's daily execution sits with three researchers Stoica advised at Berkeley: Yang, Wu and Bhardwaj. They spent years building and distributing the underlying software before asking customers to buy a managed platform.

Research credits became the product thesis

SkyPilot began with an irritation inside Berkeley's Sky Computing Lab. Researchers had donated credits from AWS, Google Cloud, Microsoft Azure and specialized providers, yet much of that capacity went unused because each provider required different setup, APIs and migration work.

Yang's team wanted one interface for the entire pool. The researchers released the open-source SkyPilot project on November 16th, 2022, after developing it for more than a year. A subsequent USENIX research paper described SkyPilot as an intercloud broker that could select infrastructure based on capacity and cost, provision clusters and move jobs among regions and providers.

The problem grew as AI companies began assembling capacity wherever GPUs were available. A single team may now combine hyperscalers, specialized GPU clouds, Kubernetes environments, Slurm clusters and several generations of accelerators. Stoica told Fortune that expanding Databricks from one cloud to two consumed about a year of engineering work. SkyPilot is betting that AI infrastructure teams will pay to avoid repeating that work for every provider and cluster.

SkyPilot Platform sits above infrastructure that customers already control. It does not resell GPU time. SkyPilot instead provides a common layer for provisioning, scheduling, monitoring, validating hardware and moving workloads across providers. The platform supports development environments, training, reinforcement learning, batch inference and model serving, according to SkyPilot.

That bring-your-own-compute design gives Yang a clean position in the market: SkyPilot can work with cloud vendors without requiring customers to purchase capacity from one favored supplier. SkyPilot says it supports hyperscalers, specialized GPU clouds, Kubernetes, Slurm and virtual machines, with CoreWeave and Nebius named as integration partners. A separate Hugging Face integration lets customers run workloads across supported clouds while keeping model data in Hugging Face storage.

The open-source base gives SkyPilot a head start

The SkyPilot repository gives SkyPilot a public open-source base. SkyPilot says the software has recorded more than 14 million downloads and attracted over 280 contributors. Those figures are company-reported, and downloads do not show how many organizations run SkyPilot in production.

SkyPilot also says hundreds of organizations use the software. Its named users include Abridge, Applied Compute, H Company and Nubank. The largest deployments have exceeded 1,000 nodes and 10,000 GPUs, according to SkyPilot, while some commercial platform customers use it to manage more than 10,000 GPUs for over 200 researchers.

Those deployments form the sales funnel for SkyPilot Platform. Yang wrote that the founders postponed commercialization because the open-source software already worked, then encountered demands for high availability, governance, managed support and better utilization across large GPU fleets. The commercial product adds those controls alongside SSO, role-based access and quota management.

Yang told Fortune that SkyPilot can frequently improve utilization by more than 10% for customers spending $100 million a year on GPUs. That would translate to $10 million in annual savings, though the figure is SkyPilot's calculation and Fortune did not identify the customer data behind it. SkyPilot separately claims some customers have seen performance improvements of up to 20 times compared with the open-source edition.

The economic pitch is stronger than simple cloud price shopping. Large AI companies often commit capital to reserved or owned GPU capacity. Once that hardware is secured, idle time, failed jobs and fragmented scheduling become operating expenses. Yang is selling SkyPilot as a way to recover capacity customers have already paid for.

Neutrality faces a well-funded test

GPU orchestration already has a powerful incumbent. Nvidia agreed to acquire Run:ai in 2024, with Fortune placing the price at roughly $700 million, and later opened the software. Run:ai manages shared GPU infrastructure through Kubernetes and is integrated with Nvidia's enterprise products.

SkyPilot's differentiation is its wider abstraction across clouds, Kubernetes, Slurm and virtual machines, paired with a claim of neutrality among compute providers. That argument will matter only if SkyPilot can maintain reliable behavior across infrastructure it does not control. The project's public GitHub tracker includes reports involving provisioning failures, hardware mismatches and provider-specific configuration problems, the kind of operational edge cases inherent in supporting many backends.

Lux is financing a familiar Berkeley commercialization pattern. Stoica helped turn Apache Spark research into Databricks and Ray into Anyscale. SkyPilot applies the same playbook to the fragmented layer beneath AI models: build an open system in the lab, seed adoption among researchers, then sell reliability and centralized management to organizations operating at larger scale.

Yang now has to prove that neutrality can become a durable software business rather than a feature cloud vendors absorb. The seed round buys SkyPilot time to build enterprise operations and sales around an open-source project that already has distribution. It also puts SkyPilot between AI teams seeking flexibility and infrastructure providers that would prefer those teams to stay inside one stack.