Google Put an AI Computer in Space to See If It Works
Google sent an AI chip into space for the first time on October 1, 2026, on a SpaceX rocket from California. TechCrunch
The satellite was built by Planet Labs. It carries a Google Tensor Processing Unit, or TPU, a chip made for AI tasks. It will test whether that chip can work in orbit. The work is part of Project Suncatcher, Google's plan for large groups of computer satellites around Earth, led by executive Travis Beals.
Survival is the test. Speed can wait.
After early checks, the satellite will turn on its TPU in 15-minute bursts. Short bursts keep power use and heat inside the limits of a small satellite that was not built as an AI server. Space has no air to carry heat away, so cooling is like running a laptop with the vents blocked.
The launch was a shared ride. The SpaceX vehicle carried more than 100 payloads, including missions from Satlyt and Cowboy Space Company. Google's prototype flew on a Falcon 9, according to launch reporting. CNBC
This October flight was the first in-orbit test of Project Suncatcher.
A two-satellite demo and an 81-satellite design
Google and Planet Labs are already building the next step. A demo planned for next year will use two satellites built for heavier computing.
Those two satellites will try to link by laser and work together. Laser links can move a lot of data without heavy gear and without radio licenses, but holding the link between two moving satellites is hard. The demo will give early data on whether computers on separate satellites can stay in sync.
Google's longer-term design calls for 81 satellites flying close together and processing in parallel. The current prototype is much smaller. It can answer simple AI questions from space, a capability described before launch. The New York Times
Google has described Project Suncatcher in company materials as an early test to scale AI computing in space and as its moonshot to put AI in space.
Launch cost as the gating factor
With the hardware, Google released a peer-reviewed version of its white paper on space data centers for publication in Joule.
The paper is about launch costs. Google researchers estimate SpaceX has cut prices by about 20 percent per year since the Falcon 1. At that pace, they expect prices near $200 per kilogram by 2035.
The paper applies the same math to Starship. Researchers estimate Starship would need to lift 370,000 tons to orbit to follow a cost path like Falcon 9. At 200 metric tons per flight, that is about 1,800 launches over the next 10 years, or 180 per year.
Google is a major investor in SpaceX.
The broader context here is that Google sees launch rate, not chip speed, as the main limit. The TPU test can check radiation effects, heat, and power cycling in orbit. It cannot by itself lower cost per kilogram. Only frequent, reusable flights do that.
In my view, the 15-minute bursts and the move to purpose-built satellites show where the hard part sits. On the ground, power and cooling shape data centers as much as chips. In orbit, with no air and no power grid, that link is even tighter.
Worth flagging for software work, parallel AI work across moving satellites needs a steady link, a way to work around a satellite that drops out, and software that treats motion as part of the network. The two-satellite test next year should show if Google and Planet Labs can hold the link steady enough.
Looking ahead, the hopeful case is that strong solar power and cheaper launch could open a new home for AI. That needs Starship to fly often and cheaply, satellites that handle steady heat, and AI tasks that fit spread-out machines working in bursts. One test does not answer those questions. This flight starts gathering the data to find out.
