Google's Project Suncatcher is scheduled to launch its first prototype satellite aboard SpaceX's Transporter-18 rideshare mission, putting Tensor Processing Units into low Earth orbit for the first time. The immediate goal is narrow: test whether TPU chips survive the physical stress of launch and the radiation and thermal extremes of space. The long-term ambition is vast — constellations of satellites running AI workloads powered by near-constant sunlight. The physics case is straightforward. In low Earth orbit, satellites access up to eight times more solar power than ground-based installations. If compute-per-watt economics ever favor orbital deployment, that energy advantage could be decisive. But energy abundance is the easy part. The engineering challenges disclosed here — cooling electronics in a vacuum, maintaining laser interconnects between moving satellites with coin-sized targeting precision — are each standalone hard problems that ground-based data centers solved decades ago with air, fiber, and fixed foundations. The hardware survival results are genuinely informative. TPU chips endured vibration testing at 50-100 g-force across three axes. Radiation testing at UC Davis's Crocker Nuclear Laboratory showed Trillium TPUs surviving total ionizing doses exceeding a five-year space mission equivalent. These are real numbers from real facilities, not simulation. The proton beam tests monitored bitflip errors during active AI workloads — a meaningful test protocol. Cooling is where the honest uncertainty lives. TPUs generate concentrated heat, and vacuum eliminates convective cooling entirely. The team is testing heat pipes and radiators in thermal vacuum chambers, but acknowledges the in-orbit test will be the first real data point. Future satellite designs envision dozens of TPUs per satellite — a thermal management problem that scales nonlinearly with chip density. The interconnect challenge is equally formidable. Satellite clusters need high-bandwidth, short-distance laser links between platforms moving at orbital velocity. Existing space-based laser communication systems optimize for the opposite regime: low bandwidth across large distances. The 2027 two-satellite test will be the first attempt at the required configuration. What's conspicuously absent from the announcement is any economic analysis. There is no comparison of cost-per-FLOP between orbital and terrestrial compute, no discussion of launch costs per TPU, no mention of latency penalties for space-based inference serving Earth-side applications, and no analysis of how orbital compute competes with simply building more ground-based solar farms in deserts. The framing as analogous to early autonomous driving and quantum computing research is strategic — it pre-emptively frames any skepticism about near-term viability as missing the long-term vision. This is a genuine research program with real hardware tests and disclosed results, wrapped in a corporate communications package that elides the hardest question: not whether it's possible, but whether it's ever economically rational compared to terrestrial alternatives. The 8× solar advantage means nothing if launch, cooling, interconnect, and maintenance costs eat it ten times over.
