Google launches its first TPU chip into orbit

Google launched its prototype orbital compute satellite, developed under Project Suncatcher, aboard a SpaceX rocket. Built by Planet Labs, the satellite carries a Google Tensor Processing Unit (TPU) to prove the hardware can operate in the harsh environment of space. To prevent overheating and power strain, the satellite will run the TPU in 15-minute bursts, requiring one kilowatt of continuous power. Google and Planet Labs are already planning a second demonstration next year with two purpose-built satellites communicating via a laser link.

This launch is the first step in Google's long-term vision of creating an orbital data center consisting of 81 satellites flying in close formation and processing data in parallel. However, scaling this infrastructure depends on dramatic reductions in launch costs. In a peer-reviewed white paper published in Joule, Google researchers estimate that SpaceX must lower launch prices to near $200 per kilogram by 2035. Achieving this rate requires SpaceX's Starship to fly 370,000 tons of payload over 1,800 launches over the next 10 years—averaging 180 flights annually, assuming each mission carries 200 metric tons.

For AI practitioners, the viability of space-based compute depends heavily on hardware reliability. Ground testing in a particle accelerator revealed that while space radiation causes slightly more errors in the TPU's logic circuitry, the chip remains highly reliable for standard workloads. The error rate is roughly one in a million, which Google deems perfectly acceptable for running large inference models over a satellite's five-year lifespan. However, this error rate makes the current setup unsuitable for mega-scale AI training runs that require thousands of chips to run continuously for months.