Google’s AI Is Officially Leaving the Planet
Google is putting its AI chips where the sun always shines: orbit.
The company announced Thursday that the first test satellite for Project Suncatcher, its research effort to find out whether space can host large-scale AI computing, launches next week. The refrigerator-sized prototype, called MVP, carries four of Google’s Trillium Tensor Processing Units and will ride to low Earth orbit on SpaceX’s Transporter-18 rideshare mission, lifting off October 1 from Vandenberg Space Force Base in California. Google built the satellite with Planet Labs, the San Francisco Earth-imaging company.
According to Reuters, the mission will measure how Google’s AI hardware handles launch forces, radiation, and the extreme temperatures of low Earth orbit — the first time the company’s chips will run anywhere but inside a ground-based data center.
Fifteen Minutes On, Then a Cooldown
Space is a terrible place to run a hot chip. There is no air, so fans are useless, and Google’s engineers had to design a cooling system built around heat pipes and radiators that dump heat straight into the vacuum.
The compromise: the TPUs can only run for about 15 minutes at a stretch. Then they shut down while the radiators shed heat, before firing back up. Travis Beals, Project Suncatcher’s senior director of product management, says the satellite’s solar panels generate roughly one kilowatt of power — about what a hair dryer draws — and the four TPUs together pack the compute of roughly one data center server.
Before the flight, Google blasted the chips with radiation at UC Davis’s Crocker Nuclear Laboratory, simulating more than five years of orbital exposure. As CNBC reports, the Trillium TPUs withstood a larger radiation dose than expected, and most radiation-induced bit flips — the data errors space radiation causes in electronics — could be fixed with a simple restart. But some things, Google admits, can only be tested in space.
Why This Matters
The prize is power. In low Earth orbit, satellites get near-constant sunlight and can generate up to eight times more solar power than panels on the ground — a tempting answer to the AI industry’s spiraling electricity appetite. SpaceX and startup Starcloud are chasing the same idea, and Nvidia announced its own space-bound AI module, Vera Rubin Space One, with Starcloud earlier this year.
Google is blunt that this is a research flight, not a product launch. The mission’s job is to collect in-orbit data and find failure points. Two more satellites go up in 2027 to test the high-bandwidth laser links that future orbital clusters would need. Google estimates launch costs would need to fall to around $200 per kilogram — and roughly 10,000 satellites would be needed to match a single gigawatt-scale data center — before orbital AI makes economic sense, a milestone it places in the mid-2030s. Jeff Bezos has put it even further out, at up to 20 years.
Still, the direction is unmistakable: the AI infrastructure race just gained a vertical dimension. And on October 1, the first four TPUs start climbing toward it.
Meanwhile in Space:
- AstroForge Puts an AI in Command of Its Next Spacecraft — With Zero Orders From Earth
- SpaceX Delays Starship’s First Orbital Flight to September 28
