Google Tests AI Model Inference on Orbital Satellite

Google shared milestones from its Project Suncatcher orbital computing research. Following an October 1 launch of a prototype satellite developed with Planet Labs, the team successfully used an onboard TPU to run Gemma 4 model inference. The refrigerator-sized satellite carries four proprietary TPUs and solar panels providing roughly 1 kilowatt of power. Constrained by power consumption, only one TPU can currently operate for approximately 20 minutes at a time, aiming to verify hardware reliability in space. Google plans to continue assessing thermal stress and radiation exposure, targeting a 2027 launch of two satellites to test high-speed optical links.

Key takeaways

• Google successfully ran Gemma 4 inference in orbit using a datacenter-grade TPU aboard a prototype satellite co-developed with Planet Labs.
• The satellite carries four proprietary TPUs powered by roughly 1 kilowatt of solar panels, currently allowing one TPU to run for about 20 minutes per session.
• Google plans further testing under space thermal stress and radiation, alongside a planned 2027 launch of two satellites to evaluate high-speed optical communications.