Google's Prototype Satellite Runs AI in Orbit

Google now has a working data center chip in space. On October 1, a SpaceX Falcon 9 carried the company's first Project Suncatcher prototype satellite into orbit, and after about three weeks of commissioning, engineers uplinked their first prompt this week. The answer came back as the satellite passed over Svalbard, Norway: a haiku about being in space, generated by Google's open Gemma 4 model running on an orbital TPU. Google calls it the first AI response from a data-center-class TPU in orbit.

Worth being precise about what this is: a real launch and a real working prototype, but an experiment - one small satellite built with Planet Labs, not an operational data center.

One TPU at a time

The spacecraft carries four Google TPUs and generates up to 1 kW of electricity, but the power budget means it can only switch one chip on, for about 20 minutes at a time. Since the haiku, the team has run space-themed limericks and, more usefully, telemetry prompts asking the satellite to report on its own orbital conditions. The serious work starts now: measuring how the chips handle thermal stress and space radiation, whose "bit flips" can corrupt data silently. Ground tests in particle accelerators suggested inference workloads should cope; orbit gets the final say.

A moonshot, not a product

Suncatcher is classic Google long-game research, aimed at infrastructure that might make economic sense in the 2030s. The next milestone is a pair of interconnected satellites in 2027 with higher power and optical inter-satellite links, the piece Google bets will let future constellations run distributed AI workloads. The vision also depends on SpaceX's Starship becoming fully and rapidly reusable - Suncatcher's own researchers estimate that could take around 1,800 flights of learning-driven cost reductions. Google is not alone up there: startups like Starcloud and Cowboy Space are pursuing single-satellite inference, and SpaceX has orbital data center plans of its own.