Google runs Gemma model in orbit in Project Suncatcher satellite test Google

Google completed the first orbital neural network inference under Project Suncatcher, transmitting a prompt from Earth and running its Gemma foundation model on a spaceborne Tensor Processing Unit.

The satellite, built with Planet and launched into low Earth orbit via SpaceX Transporter-18, downlinked an AI-generated haiku as it passed over an Arctic ground station in Svalbard, Norway.

What it is

• In-orbit hardware execution. Google deployed data-center-class TPU silicon aboard a compact satellite to test machine learning computation outside Earth's atmosphere.
• Solar power potential. Orbital data centers can capture unfiltered starlight nearly continuously, yielding up to eight times more solar power per panel area without terrestrial grid contention or water cooling constraints.
• Thermal and radiation testing. The mission evaluates vacuum heat dissipation through custom radiators alongside silicon resistance against cosmic ray bit-flips.
• Economic viability criteria. Google research indicates space data centers will become economically competitive with terrestrial facilities only if orbital launch costs drop tenfold to around $200 per kilogram, which the company estimates will not occur before the mid-2030s.
• Cluster roadmap. Google plans to deploy twin interconnected satellites in 2027 to evaluate high-speed free-space laser cross-links for distributed orbital compute.

Orbital test specifications

Sources

• Google Research: Exploring a space-based scalable AI infrastructure
• Google Blog: Project Suncatcher Overview