Google thinks SpaceX’s Starship has to launch 1,800 times before space data centers get off the ground

Google launched a Planet Labs-built satellite carrying one of its Tensor Processing Units into orbit aboard a SpaceX rocket to test whether the advanced AI accelerators can operate in space. The flight is intended to validate that the TPU can be powered (about a kilowatt of continuous power), cooled, and run inference models reliably under real orbital conditions after ground testing. During commissioning the chip will be run in 15-minute intervals to limit stress on the satellite's power and thermal subsystems.

The prototype is the first step in Project Suncatcher, Google’s long-term effort to develop large-scale compute clusters in orbit. Google and Planet Labs plan a follow-up demo next year with two satellites purpose-built for heavier compute loads and connected by a laser communications link to explore cooperative processing. The current launch shared its rocket with more than 100 other payloads, including projects from Satlyt and Cowboy Space Company.

In a peer-reviewed white paper published in Joule, Google laid out analysis of how compute reaches orbit and how launch costs might evolve. The paper assumes SpaceX’s historical learning curve — about a 20% cost reduction per year since Falcon 1 — could continue, leading to launch prices near $200 per kilogram by 2035. To reach that point through Starship, Google’s calculation estimates Starship would need to carry a cumulative 370,000 tons into orbit, which at 200 metric tons per flight implies roughly 1,800 launches over the next decade (about 180 launches per year).

Google acknowledges that this ramp would be a major scale-up from current Starship flight rates (the vehicle has not flown more than five times in a year to date), and notes that while SpaceX projects much higher cadence, those projections are uncertain. On the hardware side, Google repeated radiation testing of its TPUs using a particle accelerator after discovering earlier test setups overestimated shielding; the revised tests showed slightly more logic errors but still indicate low error rates for typical inference workloads over a five-year satellite lifespan. The company cautioned that error rates could be more problematic for sustained, massive-scale training runs.