Google's computing satellite runs Gemma 4 in orbit and delivers a haiku
The Project Suncatcher prototype is meant to show whether AI chips in space can cope with heat and radiation, on the road to data centres in orbit.
Google has for the first time received an answer from an AI model running on board a satellite. The Project Suncatcher craft was launched earlier this month on a SpaceX Falcon 9 rocket together with more than a hundred other spacecraft. After about three weeks of system checks, engineers this week sent a first prompt to Gemma 4, the Google DeepMind model loaded on the satellite: write a haiku about life in space. According to The Verge, drawing on a Google blog post, other prompts followed, including requests to report data about conditions in orbit.
TechCrunch describes the technical limits. The satellite carries four of Google's own AI chips, known as TPUs, and generates up to 1 kilowatt of power. Because a TPU draws a lot of power, only one can be switched on at a time for about twenty minutes, which makes communication with the ground even trickier. A spokesperson told TechCrunch that the satellite is operating normally.
In the coming period Google will study how the chips respond to temperature swings and cosmic radiation, which can flip bits in memory and silently corrupt data. Tests on Earth in particle accelerators suggested radiation is mainly a problem for training models and less so for running them, but that expectation now has to be tested in practice. If the test succeeds, two satellites linked to each other by lasers will follow in 2027. According to TechCrunch, this sets Google apart from other plans for computing centres in space, in which satellites each work on their own.