Project Suncatcher Flies Google AI Chips Into Orbit to Learn If Compute Can Live on Sunlight

Google will send its first AI chips into space next week as the opening flight of Project Suncatcher, a long-term research effort to find out whether machine learning hardware can run in low Earth orbit on solar power. A refrigerator-sized prototype named MVP, built with Planet Labs and carrying four Tensor Processing Units, is booked on SpaceX’s Transporter-18 Falcon 9 rideshare from Vandenberg, with a target date of October 1. The satellite is not a working orbital data center. It exists to measure what launch vibration, radiation, and vacuum heat do to Google’s chips once they leave the ground.

Project Suncatcher was introduced in November 2025 as a way to work backward from a future of clustered, solar-powered satellites that talk to each other over lasers. Sitting in a dawn-dusk sun-synchronous orbit, panels can absorb near-constant sunshine and generate up to eight times the electricity you’d get on Earth, which is the entire rationale for leaving the planet in the first place. The plain, unvarnished reason Google is pushing this is because our present AI usage consumes a lot of electricity and land, whereas the Sun is the solar system’s largest energy source. The first flight is essentially the smallest feasible iteration of that concept, with a hair dryer’s equivalent of electricity, roughly 1 kilowatt of solar power, and four Trillium TPUs performing Gemini workloads in brief bursts.

Planet had already built the bus. Google, which has invested in the company, asked to fly sooner than the two custom satellites planned for 2027, so the chips went into an existing spacecraft instead of waiting for a purpose-built pair. So they just stuffed the chips into an existing vehicle rather than waiting for a pair to be manufactured specifically for the task. Yes, it’s a challenging ride; launch is only ten minutes of nonstop turbulence. The spaceship experiences sustained weights of 10g, while individual components like as the TPUs can weigh 50-100g or more. To put that to the test, ground crews physically shook the satellite back and forth on all three axes and were relieved to discover that the hardware could withstand it.

That still does not get us to orbit, where the only thing that can remove heat is a thin vacuum, and there is no atmosphere to deflect radiation. Radiation testing began at UC Davis’ Crocker Nuclear Laboratory, where researchers conducted AI operations on Trillium TPUs inside a proton beam while keeping a tight eye on how many bit flips and long-term damage the chips sustained. Google claims that the chips withstood more ionizing radiation than a usual five-year mission and did not break altogether, albeit memory proved to be the weakest spot. That’s somewhat useful, but also a bit incomplete.

The real test will come when the satellite is launched and we can observe how solar activities and cosmic rays behave in space. Then there’s the issue of cooling, which means the heat pipes and radiators must be able to dump the TPU’s waste heat into the vacuum. On the first flight, the system can only handle roughly 15 minutes of computation time before the radiators overheat and the processors shut down.

Next year, two more satellites will be launched to test the component that will transform isolated boxes into proper clusters: high-bandwidth laser communications over short distances. Each craft will need to know where it is and where its neighbors are, as well as be able to keep its beam on point. Google compares it to trying to hit a coin-sized target from miles away while both ends move. On the ground, they have a system that is already moving 800 gigabits per second each way, or 1.6 terabits total, via a single pair of transceivers. A study preprint already envisions a future in which there are 81 satellites, each carrying many more processors, within a one-kilometer radius, using free-space optics to replace the fiber that connects racks in a typical hall.