Project Suncatcher: Google’s A.I. Chips Are Now Computing in Orbit

Google’s first “Project Suncatcher” satellite is up there now, roughly 650 kilometers above Earth, and it’s already talking back. The prototype launched a few days ago on SpaceX’s Transporter-18 mission, and project lead Travis Beals said the team has confirmed contact and the craft is operating as expected.

This is not a polished space data center. It’s a refrigerator-sized test rig built with Planet, the Earth observation company that is also operating the satellite. On board are four Trillium-generation Tensor Processing Units, the same kind of chips Google uses for Gemini models on the ground. For now, they’re running Gemma in 15-minute bursts, long enough to see whether the hardware can survive launch, radiation, and brutal temperature swings.

Google says the chips already handled radiation doses in ground tests at UC Davis that were well beyond what a five-year mission would be expected to see. The company has also published the underlying research in Joule, which gives this more weight than a glossy moonshot video usually does. Over the next few weeks, the team will pull data from orbit and adjust the design.

The bigger plan is much more ambitious. Google wants two more satellites next year to test whether the chips can be linked with high-bandwidth lasers, and eventually it imagines swarms of satellites carrying dozens of TPUs doing large A.I. jobs together. The pitch is simple: space gets more solar energy than Earth, because there’s no night, clouds, or atmosphere getting in the way.

But heat is the catch. In a vacuum, there’s no air to whisk waste heat away, so it has to be radiated out through heat pipes and radiators. That is why the satellite only computes for 15 minutes at a time before cooling down. And then there’s the money problem. Google says orbit would need launch costs below $200 per kilogram by the mid-2030s to compete, and Beals told NPR he doesn’t see it being cheaper in the next five years.