Google's Project Suncatcher runs Gemma 4 on a TPU in orbit

Google's Project Suncatcher hit its first milestone. A prototype satellite launched on October 1 received a command from the ground, booted its data-center TPU, ran a prompt through Gemma 4 and sent a complete response back to Earth. The response was a haiku, which Sundar Pichai posted: "Blue marble below, / Data streams in silent grace, / Starlight feeds my core."

Suncatcher is a long-term research moonshot asking whether machine learning can one day scale using abundant solar energy in orbit, sidestepping terrestrial power and cooling limits that constrain data-center growth. This test proves the basic loop of uplink, compute and downlink works with production TPU hardware.

The context is the energy crunch behind AI infrastructure. Hyperscalers this week reported soaring capex, and AWS, Microsoft and Samsung all tied record numbers to AI demand that still exceeds supply. Space compute is one of several exotic answers to that bottleneck, alongside nuclear deals and new grid buildouts.

Skeptics are loud. Engineers note only one of four TPUs can run at a time, for about 15-20 minutes before overheating, and launch costs remain huge. Pichai himself listed open challenges from radiation bit-flips to orbital thermal management. This is a proof of concept, not a product.