Project Suncatcher is a research moonshot by Google aimed at determining if space can serve as a viable location for scalable machine learning infrastructure. The project is motivated by the potential for satellites in low Earth orbit to access near-constant sunlight, which could generate significantly more solar power than terrestrial data centers. To realize this, the team must first overcome the fundamental challenge of operating specialized AI hardware in the extreme conditions of space. The initial phase of this research involves a prototype satellite mission developed in partnership with Planet, which is scheduled to launch on a SpaceX Transporter-18 rideshare mission. This mission is designed to gather data on how Google's Tensor Processing Units (TPUs)—the company's custom AI chips—withstand the physical stresses of launch and the harsh environment of orbit. During a rocket launch, components experience intense vibrations and acceleration loads that can reach 50 to 100 times the force of gravity. To prepare, engineers subjected the hardware to vibration testing across all three axes to simulate launch frequencies. Once in orbit, the chips face additional hazards, including cosmic rays and solar events that can cause bitflips or other electronic errors. Testing at UC Davis’s Crocker Nuclear Laboratory using a proton beam has indicated that the Trillium TPUs can survive a total ionizing radiation dose exceeding that of a five-year space mission. Thermal management presents another significant hurdle. Because there is no air in a vacuum to facilitate cooling, the team is developing systems that combine heat pipes and radiators to diffuse the heat generated by the chips. Future iterations of the project aim to link clusters of satellites using high-bandwidth lasers to manage larger AI workloads, with a follow-up mission planned for 2027 to test this interconnectivity.
The project is a research moonshot by Google that explores utilizing the near-constant solar energy available in low Earth orbit to power advanced machine learning hardware.