The most ambitious infrastructure proposal of the AI era is not on Earth. Google's Project Suncatcher, as The New York Times reported from its San Francisco lab, aims to put AI data centers in space, powered by sunlight that never meets a cloud, a grid or a zoning dispute.
The first step is MVP, a refrigerator-size satellite commissioned from Planet Labs and loaded onto a Falcon 9 at Vandenberg. It carries four TPUs — roughly one server's worth of compute — fed by solar panels delivering about a kilowatt: the power of a hair dryer. Google's expectations are deliberately tempered; the mission is a physics test, not a product.
Two hard problems dominate. Radiation flips bits, changing a zero to a one at the circuit level. Since February 2025 Google has taken TPUs to the Crocker Nuclear Laboratory and blasted them with a cyclotron — magnets with Manhattan Project heritage — simulating five years of exposure. The promising result: restarting the chips usually resets the errors.
Cooling is the second. Fans do not work in vacuum, so the stack is layered like a pastry: TPUs on their motherboard, sheets of thermal interface material — putty like a Fruit Roll-Up — carrying heat into aluminum and copper, and a radiator panel exhausting it into space. The chips run about fifteen minutes per pass before resting; long enough for short Gemini queries.
The roadmap scales fast: two more satellites in 2027, then fleets of 80-plus flying in tight formation, talking to each other as a distributed data center — and a custom satellite the length of a soccer field on the drawing board. Google estimates orbital compute reaches cost parity with terrestrial data centers by the mid-2030s.
The project's champion, Blaise Agüera y Arcos, pitched it expecting rejection; Sergey Brin and Sundar Pichai greenlit it in the room. The best measure of success, one Google director said, will be when nobody notices their Gemini query was served from orbit.