Google launches its chip into orbit, says space data centers need thousands of Starship flights
Google sent its first Tensor Processing Unit chip into space aboard a SpaceX rocket to test performance in orbit, while a new peer-reviewed paper outlines how many Starship launches would be needed to make orbital data centers viable.
Google launched a satellite carrying its Tensor Processing Unit (TPU) chip into orbit for the first time on Thursday, lifting off aboard a SpaceX rocket from California. Built by Planet Labs, the satellite is part of Google's Project Suncatcher, an effort to build large-scale compute clusters in Earth orbit.
Travis Beals, who manages the project, said ground testing can only go so far and that real-world validation is essential. The satellite will run its TPU in 15-minute bursts to avoid overloading its power and cooling systems. A more advanced demo involving two purpose-built satellites communicating via laser link is planned for next year.
Thousands of launches needed
Google envisions a constellation of 81 satellites flying in close formation to process workloads in parallel. The company acknowledges this depends on rockets becoming far cheaper and more frequent than they are today. A newly published, peer-reviewed white paper — set to appear in the journal Joule — examines SpaceX's pricing trends, noting the company has achieved roughly a 20% annual cost reduction since its first Falcon 1 flight. Based on this trajectory, Google estimates launch costs could fall to around $200 per kilogram by 2035.
Reaching that point would require SpaceX's Starship to complete about 1,800 launches over the next decade, carrying a combined 370,000 tons of payload — roughly 180 launches a year, assuming each carries 200 metric tons. Starship has never flown more than five times in a single year, though Elon Musk has suggested the vehicle could eventually fly as often as once an hour by 2029.
The research also offered encouraging news on radiation resistance. Updated particle accelerator tests showed a slightly higher error rate than initial results suggested, but Google remains confident its chips can handle large-scale inference workloads throughout a satellite's five-year lifespan. The company noted that massive training runs involving thousands of chips over extended periods would be more vulnerable to radiation-related errors.
