Project Suncatcher: Google Sends Its First TPUs Into Orbit

A moonshot facing the test of reality

Announced the previous year, Project Suncatcher explores an idea simple to state and formidable to execute: hosting large-scale machine learning infrastructure in space. The main argument is energy-based, since in low Earth orbit a satellite catches near-constant sunlight and can generate up to eight times more electricity than on the ground.

After years of research, the project reaches a milestone with a first prototype satellite, developed with Planet and flown aboard SpaceX's Transporter-18 rideshare mission. The goal: collect real-world data on how TPUs behave under mechanical stress, radiation, and thermal swings.

Travis Beals compares the approach to Google's early work on self-driving cars and quantum computing: years of experimentation before any practical system emerges. The long-term ambition is to link multiple constellations to handle increasingly heavy AI workloads in orbit.

Surviving launch and radiation

The trip to low Earth orbit takes about ten minutes, with intense vibrations and accelerations of up to 10 g. Individual components, including TPU chips, can experience 50 to 100 g. The team shook the satellite on all three axes to reproduce launch frequencies, and admits being pleasantly surprised that everything held up.

On the radiation front, the TPUs were exposed to a proton beam at UC Davis's Crocker Nuclear Laboratory while running AI workloads. The team monitored the impact of errors like bitflips on calculations.

Initial result: Trillium TPUs withstand a total ionizing dose higher than what they would receive over a five-year mission. Google remains cautious: some things can only be tested in orbit.

Cooling without air

A TPU generates a lot of heat over a small surface area. On Earth, it's driven off with air or liquid; in a vacuum, only thermal radiation works, requiring a radically different approach.

Google is exploring several avenues, notably a combination of heat pipes and radiators. The system was validated in a thermal vacuum chamber simulating the space environment, and the flight will test these designs against reality.

Lasers to build a cluster in orbit

Eventually, each satellite will carry dozens of TPUs and fly in tight formation with others. To achieve the bandwidth AI requires, each craft must know its own position and that of its neighbors, and communicate via laser.

Space laser links already exist, but they're optimized for low bandwidth over long distances. Suncatcher needs the opposite: very high bandwidth over very short distances, with precision comparable to hitting a coin-size target from miles away, with both points in motion.

This piece will be tested in 2027 with two satellites in orbit, the program's next announced milestone.