Infrastructure automation startup SkyPilot Inc. launched today with $20 million in seed funding.

Lux Capital led the raise with participation from several other institutional backers. The funds were joined by Databricks Inc. Chief Executive Officer Ali Ghodsi, Google LLC Chief Scientist Jeff Dean and other prominent angel investors.

San Francisco-based SkyPilot commercializes an open-source project of the same name. It was developed at the University of California at Berkeley by the company’s founding team, which includes Databricks co-founder Ion Stoica. The software makes it easier for enterprises to manage the infrastructure they use in artificial intelligence projects. 

Companies often mix and match different types of AI infrastructure. For example, an enterprise might run inference workloads in the cloud while keeping its training environments and the associated datasets on-premises. In many cases, each infrastructure environment contains different chips and management tools.

Managing hardware assets with diverging technical parameters can be highly challenging. Usually, moving workloads between different infrastructure types requires extensive code changes. Additionally, automation workflows that run reliably in one environment might not work well in another.

SkyPilot’s eponymous open-source tool eases the task. It enables developers to manage infrastructure environments with different configurations through a single interface. According to the company, the software also automates several common maintenance tasks.

Developers can sometimes make an AI cluster more efficient by changing which server runs what workload. If a machine with 10 gigabytes of memory runs a 6-gigabyte workload, an engineer could replace it with two 5-gigabyte workloads. That optimization approach is known as bin packing. SkyPilot says that its software performs the task automatically to save time for users.

When an AI workload’s hardware requirements increase, the software automatically provisions additional infrastructure. It can also fix certain types of technical issues without human input.

SkyPilot makes money from the software with a paid version called the SkyPilot Platform. It includes a tool for creating inference sandboxes, virtual machines in which AI agents can run code without creating cybersecurity risks. Developers may customize the virtual machines by adding in components such as programming tools.

Launching an inference sandbox can take a significant amount of time in some cases. According to SkyPilot, its platform speeds up the task by loading virtual machines before they’re needed. That enables the software to activate a new sandbox in under a second when it’s requested by an AI agent.

The SkyPilot Platform also speeds up AI training runs. The company says that it can launch graphics card clusters with 5,000 chips in under a minute. For added measure, a tool called the SkyPilot GPU Manager regularly checks the accelerators’ health and fixes any technical issues it finds.

“Every organization is building custom intelligence around its own data and domains,” said SkyPilot co-founder and CEO Zongheng Yang. “The challenge is that the AI compute needed to build it is fragmented across clouds. SkyPilot gives frontier AI teams a single platform to manage that infrastructure so they can build custom intelligence faster.”

SkyPilot will use its funding round to enhance the paid and open-source versions of its platform. The company also plans to grow its headcount. 

Photo: SkyPilot