Google is preparing to test an unconventional idea that, if successful, could transform the future of AI infrastructure: sending chips designed to run AI workloads into space. The aim of this experiment is to examine the feasibility of building large-scale computing systems off Earth.
According to the company, the first prototype satellite of the research project “Project Suncatcher” is scheduled to launch on October 1, aboard SpaceX’s Transporter-18 mission and in collaboration with the satellite company Planet; a timing that is several months earlier than the initially planned early-2027 date. This refrigerator-sized satellite will carry four of Google’s TPU chips and will run the company’s Gemini model workloads at an altitude of several hundred kilometers above the Earth’s surface.
Google, which first introduced this project in November 2025, explained in a blog post published on September 24 by Travis Beales, the company’s head of “Patterns of Intelligence,” that the goal of the first mission is to answer a simple question: can Google’s AI hardware survive and operate in space?
Why space? Running advanced AI models requires massive amounts of energy and computing power; a reality that has pushed tech companies to look for new ways to expand data centers. According to Google, sun-synchronous orbits can produce up to eight times more energy than comparable solar panels on the ground, because satellites in such an orbit are exposed to sunlight almost continuously.
In a draft scientific paper, Google also outlined a more ambitious vision: a cluster composed of 81 satellites within a one-kilometer radius connected by inter-satellite optical links, flying at an average altitude of about 650 kilometers in a denser configuration than any existing or past satellite constellation; although the company emphasized that the final scale of the plan may change.
Technically, two main challenges face the project: space radiation and cooling. Google tested its Trillium TPU chips with 67 MeV proton beams at the University of California, Davis, and said the chips endured a radiation dose higher than what would be expected on a five-year space mission. For cooling, since there is no airflow in the vacuum of space to remove heat by conventional means, Google focused on a combination of heat pipes and passive radiators to transfer heat generated by the chips to radiating surfaces.
In the next step, planned for 2027, Google intends to test communication between two satellites using free-space optical links; a technology that in ground experiments has achieved a two-way transfer rate of 1.6 terabits per second and serves as a precursor to the larger concept of satellite constellations, each carrying dozens of TPU units.
Google is not the only player in this arena. Startup Starcloud earlier in November 2025 sent an NVIDIA H100 GPU into orbit, and SpaceX has declared space AI infrastructure as an important part of its IPO plans; SpaceX CEO Elon Musk has previously said the company will soon enter the space data center field. Nevertheless, experts say Google’s Project Suncatcher, with defined planning and a clear schedule, has turned it from a mere idea into a real operational program.