Google Tests Running AI in Space With New Satellite

Google has launched a refrigerator-sized satellite equipped with four Tensor Processing Units (TPUs) as part of its Project Suncatcher research initiative. The satellite will run a version of Google’s Gemma AI model for roughly 15-minute intervals due to heat constraints. Google intends to evaluate how its chips respond to the stresses of launch, radiation, and extreme temperatures in orbit.

The company's ultimate goal is to construct clusters of satellites functioning collectively as orbital data centers, powered by solar energy. These future satellites could house dozens of TPUs and communicate with other satellites and Earth using lasers. Google intends to dispatch two additional satellites next year to evaluate these connections.

The current prototype was created in collaboration with satellite firm Planet and deployed aboard a SpaceX rocket. The mission is projected to run for approximately a year. Significant hurdles remain, particularly in cooling, as conventional air- and water-based cooling methods are ineffective in space. Google is experimenting with pipes and radiators to dissipate heat from the chips.

Repairs would also be considerably more difficult than in traditional data centers, while launches remain costly. Google asserts that space-based data centers are unlikely to be cheaper than terrestrial facilities within the next five years. Other firms are also investigating comparable concepts. Starcloud previously launched an Nvidia H100 chip into space, while SpaceX has indicated it could initiate the deployment of orbital AI compute satellites as early as 2028.

Google characterizes Project Suncatcher as a long-term "moonshot," rather than something anticipated to supplant Earth-based data centers in the near future.