Google’s Satellite AI Generates Its First Response From Orbit
A Project Suncatcher satellite ran Google’s Gemma 4 model in orbit, marking an early test of the company’s plans for space-based AI computing.
Google’s experimental Project Suncatcher satellite has generated its first AI response in orbit and sent it to Earth. After engineers prompted an onboard Gemma 4 model to write a haiku about being in space, Google reported that the satellite’s Tensor Processing Units, or TPUs, had successfully run the request. The result is an early test of the company’s proposal to put AI computing infrastructure in space, not a demonstration of a full orbital data center.
Table of Contents
What the satellite did
The satellite launched earlier this month. According to TechCrunch, it traveled on a SpaceX Falcon 9 rocket alongside more than 100 other spacecraft, then spent about three weeks in commissioning while engineers checked its subsystems. The first prompt from the ground arrived this week, TechCrunch reported.
Google also tried space-themed limerick prompts and requests for information about conditions around the satellite, according to the reports. A Google spokesperson told TechCrunch that the spacecraft was operating normally. The Verge reported that Google had successfully booted the satellite’s TPUs before running the Gemma 4 prompt.
The limits of the orbital test
TechCrunch reported that the prototype carries four TPUs and can generate up to 1 kilowatt of electricity. It can operate only one TPU at a time, and its power demands limit that operation to about 20 minutes at a stretch. Those constraints add to the difficulty of running computing tasks and communicating with Earth from orbit.
The next tests will focus on how the chips handle temperature extremes and radiation in space. Google has already tested TPUs using particle accelerators on Earth, TechCrunch reported, but orbital conditions will provide a direct measure of how the hardware performs. Radiation can alter values stored in memory and potentially corrupt data without setting off standard monitoring tools. Ground testing suggested that such effects would pose less of a problem for running an existing model than for training one, according to TechCrunch.
What comes next
Google says reliable TPU operation in orbit would clear the way for a 2027 test involving two interconnected satellites. TechCrunch reported that those future spacecraft are intended to generate more power and demonstrate optical links between satellites. The longer-term aim is a network of nearby spacecraft capable of handling substantial AI inference workloads together.
For now, the haiku establishes a narrower milestone: a model ran on the prototype’s hardware in space and returned a response. Whether the chips can work reliably under sustained orbital conditions—and whether linked satellites can share larger workloads—remains to be tested.
Sources
This story was compiled by AI from the reports below. Read the originals for the full details.