Google to send its AI chips to space for tests for first time
Google is preparing to send its own artificial intelligence processors into space for the first time. The company will launch a prototype satellite with TPU chips as part of the Project Suncatcher research project, which in the long run should show whether it is possible to place large computing power for AI directly in orbit.
The first experiment will take place as early as next week as part of the SpaceX Transporter-18 rideshare launch. The mission was prepared by Google together with satellite company Planet. The main goal is to check whether Tensor Processing Units, which Google uses to train and run AI models, can withstand real conditions of spaceflight.
During the mission, specialists will collect data on how the chips cope with vibration and overloads during launch, space radiation, vacuum, and sharp temperature swings. During satellite insertion into low Earth orbit, the equipment experiences accelerations up to 10 g, and individual electronic components can be subjected to loads tens of times higher.
Google has already carried out ground tests of the equipment. The satellite platform was tested on vibration tables simulating rocket launch, and the Trillium generation TPUs were exposed to proton radiation at the lab of the University of California, Davis. According to the company, the processors withstood a radiation dose exceeding the expected five-year orbital exposure.
Cooling remains a separate challenge. In ordinary data centers, heat from processors is removed by air or liquid systems, but in vacuum this approach does not work. For the space version, Google is testing a combination of heat pipes and radiators. Before launch, the system was tested in a thermal vacuum chamber, but its effectiveness still has to be confirmed directly in space.
Project Suncatcher was introduced by Google in 2025 as a long-term research project to create computing infrastructure for machine learning in orbit. The idea is to eventually cluster satellites with dozens of TPUs and use nearly continuous access to solar energy.
According to Google estimates, in low Earth orbit solar panels could potentially generate up to eight times more energy than similar installations on the Earth's surface. This could make space attractive for energy-hungry AI computing if engineers manage to solve the problems of cooling, communications, and cost of launching equipment.
In the future, satellites are planned to be connected by high-speed laser links. Such a system should provide sufficient bandwidth for distributed AI computations between multiple spacecraft. However, for this, the satellites will have to maintain extremely precise mutual positioning while moving in orbit.
The next major milestone of the project is planned for 2027. Google intends to send two spacecraft into space and test laser data transmission between them. In the long run, each satellite of a potential computing constellation could carry dozens of TPUs.
The company emphasizes that Project Suncatcher remains a research project for now, and there is no talk of moving Google data centers to space anytime soon. The first launch should primarily show how suitable existing AI hardware is for long-term operation beyond Earth.
Based on: Google, Google Research
