A prototype satellite constructed by Planet Labs left the launch pad in California aboard a SpaceX vehicle, marking Google's initial attempt to operate its custom Tensor Processing Unit beyond Earth. The experiment will assess whether the chip can sustain a continuous kilowatt of power, maintain thermal stability, and execute a suite of machine-learning models without failure.
Travis Beals, the Google executive overseeing Project Suncatcher, explained that the satellite will activate its TPU in short, fifteen-minute intervals to avoid overtaxing power and cooling resources. The current platform follows Planet Labs’ standard bus, but a follow-up mission planned for next year will field two purpose-built compute satellites linked by a laser communication system.
Project Suncatcher is framed as a long-term moonshot aimed at assembling a constellation of eighty-one satellites that operate in close formation, delivering parallel processing capability comparable to terrestrial racks. Beals emphasized that inter-TPU bandwidth and latency become critical when scaling workloads, and that the required launch cadence exceeds the capacity of existing launch vehicles.
Google released a peer-reviewed white paper in the journal Joule that examines the economics of orbital compute. The authors note a roughly twenty-percent annual cost-reduction trend for SpaceX rockets since the Falcon 1 era, projecting launch prices near $200 per kilogram by 2035. Achieving that price point, they argue, would demand about 370,000 tons of payload, equivalent to roughly 1,800 Starship missions over a decade.
Reaching a total of 1,800 flights presents a substantial operational hurdle. To date, Starship has completed fewer than five launches per year, while SpaceX has suggested an eventual hourly launch cadence by 2029. The timeline remains speculative, and the required frequency far exceeds the vehicle’s historic flight rate.
Radiation tolerance testing required Google to expose its TPU to a particle accelerator after discovering that the chip’s layout provided more shielding than anticipated. The accelerated tests showed a very low error incidence,about one error per million inference operations,supporting confidence that the processors can handle typical AI inference workloads for the projected five-year satellite lifespan, though they would struggle with large-scale training tasks.
A correction note clarified that the original headline misstated the Starship launch learning-curve estimate as 1,600; the accurate figure cited in the analysis is 1,800 launches.