SkyPilot raises $20M seed led by Lux for cross-cloud GPU routing

The pitch from Databricks cofounder Ion Stoica's new venture is not more GPUs, it is neutrality about whose GPUs. SkyPilot, run by CEO Zongheng Yang, has raised a $20 million seed led by Lux Capital with Coatue, Amplify Partners, Foundation Capital, Race Capital, and The House Fund also writing checks, Fortune reported. The framing being pushed by the founders and lead investor is 'the Switzerland of AI compute', a scheduler that sits above cloud providers and steers workloads to whichever accelerator is cheapest or actually available.

The numbers Yang uses for the pitch are concrete rather than aspirational. He told Fortune that customers spending on the order of $100 million a year on GPUs 'frequently' see more than 10% better utilization once workloads route through SkyPilot, which works out to roughly $10 million a year in efficiency savings for a buyer that size. The company also says its private preview customers are already managing more than 10,000 GPUs through the platform, and it lists Nubank, Abridge, Applied Compute, H Company, Archer and Hippocratic AI among its users.

The structural context matters here. Nvidia bought Run:ai for around $700 million in 2024 to bring GPU orchestration in-house, and Lux is betting that a vendor-neutral alternative is exactly what large AI buyers want in a world where Nvidia owns the hardware and increasingly the scheduler too. SkyPilot's integrations with neoclouds like CoreWeave and Nebius fit that theory: an independent broker is more useful to challengers than to hyperscalers whose lock-in it erodes.

The honest caveat is that most of what is verifiable today is the open source project, which has been on GitHub for years and, per the company, has passed 14 million downloads. Lux's Brandon Reeves is quoted saying the free version is 'probably like 1% of the way done', which is investor talk for 'the commercial product is still being built'. What the reporting does not give you is the paid product's shape, the business model, or which named customers actually sit behind the $100 million spend example Yang used.

The direction is the part worth watching. If a neutral scheduler genuinely commoditizes cloud choice for AI workloads, the winners are the buyers with large GPU bills and the neoclouds that get steered traffic they would not otherwise see. The losers are whoever assumed cloud lock-in would keep doing the retention work for them.