In October 2026, Google successfully received the first space-based AI responses from a prototype satellite carrying an on-board Tensor Processing Unit, marking a major milestone for Project Suncatcher as the company tests machine learning models in orbit.
Project Suncatcher Puts Tensor Processing Units in Orbit
Google has completed the initial space-based test of its custom AI hardware. Ground researchers confirmed the achievement during the final week of October.
The experimental setup utilizes a Gemma 4 model running directly on a specialized Tensor Processing Unit. When the spacecraft passed over Svalbard, Norway, the mission team tasked the onboard processor with writing a short poem about space. The satellite executed the request in orbit and beamed the resulting data down to the ground station.
Subsequent tests evaluated additional poetic prompts alongside diagnostic telemetry reporting the hardware’s operational status. Google designed Project Suncatcher to test whether these TPU chips can survive the physical launch phase while managing severe thermal and radiation profiles in low Earth orbit.
Thermal Management in a Vacuum Environment
Operating high-performance silicon in space presents severe engineering hurdles. Without an atmosphere to dissipate heat via conventional convection, the hardware relies entirely on specialized conduction and radiation mechanics.
Google engineers paired heat pipes with external radiators to manage thermal loads inside the vacuum of space. The current prototype mission, developed in partnership with satellite operator Planet, is actively collecting live performance data regarding these cooling systems under extreme temperature swings.
The 2027 Laser Inter-Satellite Link Experiments
With the initial processing validation complete, the roadmap shifts toward multi-node orbital communication. Google plans to execute a follow-up experiment in 2027 that will deploy and link two distinct satellites.
This upcoming test aims to verify high-bandwidth laser communications between spacecraft. The architecture will allow orbital nodes to share computational workloads directly without routing every packet back through terrestrial ground stations first.
