Google's Project Suncatcher satellite is online and running AI in orbit
What's the deal? On 1 October 2026, Google launched the first prototype satellite of Project Suncatcher, its research moonshot exploring whether machine learning infrastructure can one day run in space. The refrigerator-sized satellite, developed in partnership with Planet Labs, flew on the SpaceXDealroom has a profile for this one. Try Dealroom → Transporter-18 rideshare mission and carries four Google Tensor Processing Units (TPUs). Planet will bring the spacecraft online, after which Google will power up the chips and run AI model workloads in 15-minute bursts, a limit set by heat management. The mission is planned to stay operational for a year and is led by Google senior director Travis Beals.
Update, 30 October: The satellite is online and has answered its first prompts from space. After about three weeks of commissioning, engineers sent the on-board Gemma 4 model its first prompt this week, asking for a haiku about being in space ("Blue marble below / Data streams in silent grace / Starlight feeds my core"), followed by other space-themed and telemetry prompts, TechCrunch and The Verge report. The spacecraft generates up to 1 kW, but power limits mean it can run only one of its four TPUs at a time, for about 20 minutes, which also complicates downlinking results. Google says the next tests cover thermal stress and radiation exposure: ground tests suggested radiation-induced bit flips matter little for inference but more for training. If the chips hold up, two interconnected satellites follow in 2027. The milestone drew wide attention on X, where one post sharing Google's video passed 1.3 million views.
What was tested on the ground? Google says its Trillium TPUs survived proton-beam testing at UC Davis's Crocker Nuclear Laboratory, withstanding a total ionising radiation dose greater than a five-year space mission would deliver. The satellite was shaken on all three axes to simulate launch loads of up to 10g (50–100g on individual components) and tested in a thermal vacuum chamber, using heat pipes and radiators to cool the chips without airflow. The next milestone is a pair of satellites in 2027 to test high-bandwidth, short-range laser links between spacecraft, a prerequisite for clusters in which each satellite carries dozens of TPUs.
Why it matters: Satellites in low Earth orbit can generate up to eight times more solar power than panels on the ground, while the International Energy Agency projects data centres will use around 3% of global electricity by 2030. Google's earlier research estimated that launch prices could fall below $200/kg by the mid-2030s, roughly where an orbital data centre's cost becomes comparable to the energy cost of a terrestrial one. Scale remains the open question: today's largest AI data centres draw up to a gigawatt, several thousand times the output of the International Space Station's solar arrays, and heat dissipation, radiation errors and downlink capacity all remain unsolved at that scale. SpaceX, which listed in June 2026, has also pitched orbital data centres to investors.
Read more: Google · Scientific American · NPR / WRVO · Google Research · TechCrunch · The Verge
