Google tests AI chip in space, but the rocket fleet it needs isn't ready yet
Google launched a prototype satellite carrying its Tensor Processing Unit chip aboard a SpaceX rocket to test orbital data center technology. A new peer-reviewed study shows that achieving the plan's scale will require thousands of Starship launches.
Google sent one of its advanced computing chips into space for the first time today, launching a prototype satellite aboard a SpaceX rocket from California. Built by Planet Labs, the satellite will test whether Google's Tensor Processing Unit (TPU), its answer to Nvidia's GPUs, can supply continuous power, stay cooled, and run models reliably in orbit.
The mission is part of Project Suncatcher, led by Google executive Travis Beals. Once commissioned, the satellite will run its TPU in 15-minute bursts to avoid overloading its power and thermal systems. Next year, Google and Planet Labs plan a follow-up demo with two more purpose-built satellites that will communicate via a laser link. The long-term vision involves a constellation of 81 satellites flying in close formation and processing workloads in parallel.
The rocket bottleneck
The same SpaceX rocket carried more than a hundred other payloads, including missions from Satlyt and Cowboy Space Company, but Beals describes Suncatcher as a long-term "moonshot" aimed at infrastructure and AI workloads that don't exist yet.
On Thursday, Google also published a peer-reviewed white paper on orbital data centers, set to appear in the journal Joule. The researchers note that SpaceX has achieved roughly a 20% annual cost reduction since the Falcon 1's debut and estimate launch prices could approach $200 per kilogram by 2035. Reaching that point, they calculate, would require Starship to complete around 1,800 launches over the next decade — about 180 per year, each carrying 200 metric tons — far beyond the five launches per year Starship has managed at most so far. SpaceX has suggested far higher flight rates are possible in the future.
On a more encouraging note, updated radiation testing suggests Google's chips should survive the five-year lifespan of a satellite for typical inference workloads, though error rates could become problematic for massive, months-long training runs involving thousands of chips.
