Google's first orbital AI satellite will run its four TPUs in 15-minute bursts

Somebody on an infrastructure team will get this story forwarded on Monday, with a note asking whether orbital compute belongs in next year's capacity plan. Going by what Google has disclosed, it does not, and Google describes the flight in those terms itself. "This first launch is about seeing what works, identifying points of failure and applying those findings to future missions," wrote Travis Beals, senior director of Google's Paradigms of Intelligence research team ^[2].

The pitch and the payload are far apart. Google describes Suncatcher as a route to orbital AI data centers that could undercut ground facilities by running on solar energy above the atmosphere ^[5]. The end state is a large constellation linked to Earth by lasers ^[6]. What goes up on Oct. 1 is one refrigerator-sized box built by Planet Labs, a company that normally makes Earth-imaging satellites ^[1][12]. Its solar panels produce about a kilowatt ^[3]. Split evenly across the four TPUs, that comes to at most about 250 watts per chip, and the same kilowatt has to run the rest of the satellite ^[1].

Heat is the limit already visible. A vacuum has no airflow, so conventional cooling systems are out ^[13]. "We're working on a number of different approaches for this, including a combination of heat pipes and radiators to cool the chips," Beals said ^[9]. The flight unit still falls short. According to Beals, the chips will run workloads in bursts of about 15 minutes and then shut down to cool off ^[10]. The report does not say how long each cool-down lasts, so a sustained duty cycle cannot be worked out from it.

On Google's own testing, radiation is the smaller problem. According to Beals, Trillium TPUs running AI workloads under a proton beam at UC Davis's Crocker Nuclear Laboratory performed "remarkably well," despite a dose larger than a five-year mission would deliver ^[8]. Before that matters, the chips have to survive a 10-minute launch at up to 10 times Earth's gravity ^[7].

MVP rides as one part of a shared Falcon 9 payload ^[1]. A full constellation, SiliconANGLE expects, would likely need dedicated SpaceX launches ^[11]. Google announced Suncatcher in November 2025 in response to SpaceX founder Elon Musk's own plans for space-based data centers ^[4]. At scale, then, Google's launch schedule would sit with a company chasing the same idea ^[4][11].

For the person answering that Monday note, a 2x2 covers it. One axis is whether Google has published results for sustained operation, meaning chips working for days without stopping to cool. The other is whether a customer can buy the compute at a stated price and date. With neither, orbital compute goes on a watchlist. Sustained results without a product justify a call to the account team. A product without sustained results means waiting for flight data. Both together earn a line in the plan. Suncatcher sits in the first box on both counts, with 15-minute bursts ^[10] and a single experimental satellite ^[1]. I'd leave it out of next year's plan and read whatever cooling and error data Google releases after the flight. Waiting costs some lead time if the cooling results come back strong. Budgeting now is a bet on four chips and a kilowatt ^[3].