Google announced Project Suncatcher last year as a moonshot to move AI computing off Earth, where data centers’ enormous electricity needs have made them increasingly contentious. The premise is seductive: solar-powered satellites could someday run AI workloads above the atmosphere rather than draw power from terrestrial grids.¹

That long-range vision now faces a far more basic test. On October 1, Google plans to send an experimental, refrigerator-sized satellite aboard SpaceX’s Transporter-18 Falcon 9 rideshare. Built on a Planet Labs spacecraft and carrying four of Google’s Tensor Processing Units, the prototype has only about one kilowatt of solar power—roughly the draw of a microwave or hair dryer.¹

The crucial question is not whether Gemini can run in orbit, but how long. Google will test the same TPU hardware used in its Earth-based servers against launch vibration, high g-forces, radiation and extreme temperatures. “Some things can only be tested in space,” the company says of the remaining unknowns.²

Cooling is likely to be the early bottleneck. Heat pipes, aluminum and copper components, and radiators are meant to shed the accelerators’ heat into space—but the chips can run only in bursts of about 15 minutes before shutting down so the system can catch up.¹

Google chief executive Sundar Pichai has presented the flight as an open-ended experiment: “Can our TPUs survive and operate in space? Well, we’re going to find out.”Sundar Pichai (sundarpichai)³ The company expects the satellite to operate for only a few months, while still eyeing two custom launches in 2027.¹

Elon Musk, by contrast, offered no such caveat, declaring that space-based capacity would “obviously round up to 100% of all compute.”Elon Musk (elonmusk)⁴ For now, Suncatcher’s grand future rests on a small satellite—and a quarter-hour at a time.