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

Google launched its first prototype orbital compute satellite today aboard a SpaceX rocket from California, marking the first time the company has sent an advanced chip into space. The satellite, built by Planet Labs, is intended to demonstrate that a Google Tensor Processing Unit can operate in orbit by supplying a kilowatt of continuous power, cooling the chip, and running models to identify failures. Travis Beals, the Google executive managing Project Suncatcher, said ground testing cannot fully substitute for flight testing.

Once commissioned, the satellite will run its TPU in 15-minute bursts to avoid straining power and thermal systems. Google and Planet Labs are working on a demo expected to fly next year with two satellites purpose-built for advanced compute, which will attempt to collaborate via a laser communications link. Google envisions a network of 81 satellites flying in close formation and processing in parallel. Beals said bandwidth and latency between TPUs matter greatly for multi-rack workloads, and that the project looks ahead to workloads five years out. Google released a peer-reviewed version of its orbital data center white paper on Thursday, to be published in Joule.

The paper's authors, who state their analysis is not an economic feasibility study, argue SpaceX has achieved a roughly 20% annual launch price-reduction learning curve since Falcon 1 and expect launch prices near $200 per kilogram by 2035. Based on Falcon 9 payload volumes, they estimate that trajectory would require Starship to deliver 370,000 tons to orbit, about 1,800 launches over 10 years, or 180 annually, assuming 200 metric tons per mission. SpaceX has never flown Starship more than five times in a year; Elon Musk has suggested an hourly flight rate in 2029.

The updated research also indicates Google's chips will likely survive space radiation. After redoing particle accelerator tests, the company reported slightly more logic circuitry errors but remains confident the chips can handle large inference workloads over a satellite's five-year lifespan. Beals said the error rate is very low for typical inference, around one in a million, but was already problematic for mega-scale training runs involving thousands of chips over months. Google is also a major investor in SpaceX.