Google to launch TPUs into orbit in first test of space-based AI compute

Google is moving its Project Suncatcher moonshot from concept to hardware, launching a prototype satellite carrying Tensor Processing Units on SpaceX’s Transporter-18 rideshare mission in partnership with Planet. The bet behind the program is energy: satellites in low Earth orbit sit in near-constant sunlight and can harvest up to eight times more solar power than equivalent panels on the ground, which Google sees as a path to eventually clustering constellations of satellites into scalable machine learning infrastructure.

The first mission is a survivability check rather than a working data center. TPUs face brutal conditions—launch vibration and acceleration loads that hit individual chips at 50 to 100 g, plus radiation from cosmic rays and solar events. Ground testing has been encouraging: the hardware passed three-axis vibration trials, and proton-beam tests at UC Davis showed Trillium TPUs tolerating a total ionizing dose exceeding what a five-year mission would deliver. Cooling remains the harder unsolved problem, since a vacuum offers no airflow and heat can only be shed through radiators, pushing Google toward heat-pipe-and-radiator designs validated so far only in thermal vacuum chambers.

The real test of the networking vision comes in 2027, when Google plans to fly two satellites and attempt high-bandwidth laser links over short distances—an inversion of existing space optical systems built for low bandwidth across long ranges, and a pointing challenge Google likens to hitting a coin-sized target from miles away with both endpoints moving. For now, the effort is deliberately incremental: prove the silicon endures orbit, catalog the failure modes, and feed that data into later launches. Whether orbital compute ever becomes economical is unproven, but the program signals how seriously hyperscalers are treating the energy and cooling limits of terrestrial AI infrastructure.