Google’s orbital data center testbed is alive. The Suncatcher prototype launched earlier this month on a SpaceX Falcon 9 alongside more than a hundred other spacecraft, and after about three weeks of commissioning it answered its first prompt.
The request went to Gemma 4, the DeepMind model loaded before flight, and asked for a haiku about being in space. Engineers followed with space-themed limericks and, more usefully, asked the craft to report on orbital conditions. Google said the satellite is operating nominally.
One chip at a time
Four Tensor Processing Units travel aboard, but the array generates about 1 kW, enough to run only one at once for roughly twenty minutes. That constraint complicates the already awkward job of downlinking results from orbit.
The campaign now turns to physics. Radiation can flip bits in memory without tripping ordinary monitoring, and ground tests in particle accelerators suggested inference workloads would shrug off the damage better than training would. Space is where that prediction gets tested.
The real bottleneck is a rocket
Two more satellites with newer chips are planned for 2027, using higher power budgets and optical links between spacecraft so they can share work. Google’s Suncatcher researchers have estimated that SpaceX would need about 1,800 Starship flights to reach the same learning-driven cost curve earlier vehicles achieved.
TechCrunch notes Alphabet holds a significant stake in SpaceX, and that the parent company’s shares slid this week on AI-related ad pressure.
