Google Has Put Four AI Chips In Orbit To Find Out If They Survive

In this storyGoogleSpaceX

Google launched a prototype satellite carrying four of its Tensor Processing Units into low Earth orbit on 1 October, the first time it has sent its AI accelerators into space.

The refrigerator-sized spacecraft, built with Planet Labs, flew on SpaceX's Transporter-18 rideshare mission from Vandenberg Space Force Base as one of around 130 payloads, and was deployed about 61 minutes after liftoff. Google has made contact and is beginning in-orbit testing.

The mission is part of Project Suncatcher, announced in November 2025, which Google describes as a research effort towards solar-powered AI data centres in orbit. It will assess how the TPUs handle the physical stress of spaceflight along with radiation and thermal extremes. The chips will run in 15-minute bursts to avoid straining the satellite's power and thermal management systems.

Ground testing with a proton beam at the Crocker Nuclear Laboratory at the University of California, Davis, suggested the chips could tolerate more ionising radiation than five years in orbit would be expected to deliver. The satellite also carries a new cooling system, since there is no airflow in space to disperse the heat the chips generate.

Google plans two further satellites next year to test laser links between spacecraft, and has described operating in clusters of 81.

Travis Beals, who leads the project, pointed to solar energy as the attraction, saying every other power source humanity has tapped is a tiny fraction of a per cent by comparison.

SpaceX has said it expects to begin deploying orbital AI compute satellites as early as 2028, and filed with the US Federal Communications Commission in January for a constellation of up to a million. Starcloud flew an Nvidia H100 last November.

Four chips, firing in 15-minute bursts, is the correct scale at which to read this.

A single terrestrial AI campus is measured in hundreds of megawatts. Google's prototype cannot run its processors continuously without straining the satellite's power and thermal systems. The gap between the two is not a matter of adding satellites, and Google's own research says as much: its estimate is that SpaceX's Starship would need roughly 1,800 launches before orbital data centres approach the energy cost of an equivalent facility on the ground, per kilowatt per year.

That figure is the most useful thing the company published. It is a learning-curve estimate rather than a plan, and it frames the project honestly as a bet on launch costs falling by an order of magnitude rather than on chips working in vacuum.

What the mission actually proves, if it works, is narrower and still worth knowing. The ground testing with a proton beam suggested the TPUs could handle more radiation than five years in orbit would deliver. That is a simulation. Radiation in orbit arrives in a different distribution, and the only way to check is to put the hardware there.

The cooling test is arguably the harder problem. Vacuum does not carry heat away, it traps it, and every watt a chip generates has to be radiated. Running the TPUs in bursts is an admission that the thermal design is not yet solved at continuous load.

The next milestone is more informative than this one: two satellites next year, communicating by laser, which is where a cluster either becomes a computer or stays a collection of chips.

The reason this is being attempted at all is terrestrial rather than technical.

Data centres on Earth have run into limits that have nothing to do with silicon. Scotland introduced mandatory environmental impact assessments for every facility above 50MW last month, after campaigners noted that four pending applications would draw as much power as a nuclear station. Grid connections are queued, water for cooling is contested, and communities object.

Orbit has none of those problems. Sunlight is constant, vacuum handles cooling in principle, and nobody lives underneath. That is the whole case, and it explains why Google, SpaceX and a handful of startups are pursuing the same idea simultaneously.

The problems are different rather than absent. The cooling system Google is testing exists because vacuum is an insulator, not a coolant, and heat can only leave by radiating away. A failed chip cannot be swapped by a technician. And laser communication across hundreds of miles and down through the atmosphere is slower than fibre.

For readers here the connection worth drawing is to TakeMe2Space, which four days ago signed for two satellites on a 2028 Falcon 9 carrying 10 Nvidia Thor GPUs each. The Hyderabad company is solving a narrower problem: process Earth observation imagery on board and downlink conclusions rather than pixels, where the constraint is bandwidth and the economics already work.

Google is attempting something categorically larger and, by its own arithmetic, further away. Both flew on SpaceX. Neither is yet a data centre.

For daily, sharp analysis of the biggest moves in the Indian business and startup ecosystem, follow StartupFox.