Is the future of AI in space? Google unveils Suncatcher: a project to test the feasibility of data center infrastructure in space.
The energy consumption of data centers is a controversial topic in the tech sector. The demand for AI technology is causing energy bills to rise rapidly, and new, increasingly larger data centers are sprouting up like mushrooms. The need for more capacity is driving tech companies to search for creative solutions.
Tech tycoons like Jeff Bezos and Elon Musk announced ambitious plans to build data centers in space, but Google seems to be actively pursuing this path as well. In a
AI in Space
With Project Suncatcher, Google is investigating the technical and economic feasibility of ‘space data centers.’ This concept involves constellations of small satellites powered by solar energy and equipped with optical interconnections and TPU chips that perform AI calculations outside Earth. Such a system could, if proven viable in reality, provide an alternative to energy-intensive data centers on Earth, Google claims.
Space offers several advantages for energy. The closer solar panels can be placed to the sun, the more sunlight they capture. In a sun-synchronous orbit, solar panels receive up to eight times more sunlight than on Earth, providing a constant power supply. Moreover, space-based AI infrastructure requires no water or land.
However, there are some potential technical obstacles to overcome. One is achieving stable and faster networks between satellites. Google achieved inter-satellite connections with a capacity of 1.6 Tbps in laboratory tests using existing optical DWDM technology, provided the distance between satellites remains limited to about 200 meters.
Radiation and High Costs
Besides satellite network density, radiation is a concern. Google is extensively testing the radiation resistance of its Trillium TPUs. The results are again promising: the chips would be resistant to the expected radiation level over a five-year period. Memory errors remained below a level acceptable for inference, though Google emphasizes that further research outside the lab is needed.
Finally, launching satellites is a costly endeavor. Google is hopeful that low-orbit launch costs will drop below $200 per kilogram by 2035 over the next decade. If this prediction comes true, the operational costs for space data centers would not be much higher than on Earth.
Practice will have to show whether Google’s project is feasible, and there’s only one way to find out. Google plans to launch the first two prototype satellites into space in 2027. In later phases, the design will evolve further into integrated systems combining solar power, cooling, and computing power into a single unit.