Large technology companies are expanding their AI data centers in full swing, grabbing land and electricity. However, this has resulted in many problems such as high energy costs and logistical management. For example, Jeff Bezos, Elon Musk and othe...
Large technology companies are expanding their AI data centers in full swing, grabbing land and electricity. However, this has resulted in many problems such as high energy costs and logistical management. For example, Jeff Bezos, Elon Musk and others have envisioned putting AI chips in space. Google has proposed a new plan to explore new fields, Project Suncatcher, to create a scalable low-orbit TPU network.
Google envisions using solar-powered satellites as the basis, equipped with TPU (Tensor Processing Unit), and connecting nodes into a distributed network through free-space optical links. In a suitable orbit, the solar panels on the satellite are continuously exposed to sunlight and are 8 times more efficient than on the ground, reducing reliance on batteries due to almost continuous power generation.
"In the future, space may be the best place to expand AI computing," Google senior director Travis Beals wrote in an official blog post. Google also published a preprint paper that has not yet been peer-reviewed, detailing Google's progress so far in this effort.
Google needs to overcome multiple engineering challenges before Project Suncatcher can become a reality. One of the challenges is ensuring good communication between these satellites. To compete with data centers on land, "satellite-to-satellite connections need to be supported at tens of terabits per second," Beers said. Grouping a series of satellites into a tight formation, within a few kilometers of each other or even closer, can help achieve this goal. This is much closer than current satellites operate, and risks such as space debris caused by collisions also increase.
In addition, it is important to ensure that the TPU is strong enough to withstand the higher doses of radiation in space. Google conducted radiation tolerance testing of the sixth-generation Trillium TPU and stated that these AI chips survived a total ionizing dose (TID) equivalent to a five-year mission life without permanent failure.”
▲ Simulate satellites forming a tight formation in orbit.
The development of space around the Earth has undergone great changes in recent years. The new generation of satellite clusters represented by Starlink has proven that it is feasible to relay network communications through orbital systems. Similar methods can be used to deploy high-performance AI chips in space.
The current cost of launching TPUs into space is quite high. A cost analysis conducted by Google shows that by the middle of 2030, the cost per kilowatt/year of launching and operating a space data center may be roughly the same as that of a data center of the same size on the ground. Google said it plans to work with the company Planet Labs to launch several prototype satellites by 2027 to test hardware in orbit.
🙂 Only possible because of SpaceX’s massive advances in launch technology!
— Sundar Pichai (@sundarpichai) November 4, 2025
▲ Project Suncatcher was praised by Musk, and Google CEO Pichai did not forget to praise SpaceX for its huge progress in satellite launch technology.
If Project Suncatcher is successful, a data center based on the space environment will be established. In this way, Google hopes to achieve the goal of using solar energy around the clock. Ideally, this will use an almost infinite, clean energy source, so that Google can continue to develop AI and other advanced technologies without having to worry about rising power plant emissions and electricity bills due to surges in power demand in ground data centers.
