The experiment comes as the AI industry faces rising demand for computing power, electricity and data center capacity on Earth. 

[For more news, click here]

Google has put four of its AI processors into orbit, beginning an experiment that could change how the industry thinks about the physical infrastructure needed to run artificial intelligence. The company’s first Project Suncatcher satellite launched from California on October 1 aboard SpaceX’s Transporter-18 rideshare mission, carrying Google’s Tensor Processing Units, or TPUs, into space. Planet built the satellite, and Google says it has established contact with the spacecraft.  

Satellites can receive sunlight for much longer periods than solar panels on Earth, and an orbital computing system would not require the same land and grid infrastructure as a conventional data center. Before sending the hardware into orbit, engineers tested the TPUs against radiation and prepared the spacecraft for the physical stress of launch. The current mission allows them to see how the processors perform when those conditions come together in an actual spacecraft.  

Cooling presents one of the more difficult engineering problems. A data center on Earth can use air or liquid to carry heat away from processors, but there is no air in space. Google's prototype uses heat pipes and radiators to move heat through the spacecraft before releasing it into space. During the test, the TPU is expected to run in short bursts of around 15 minutes, giving engineers time to monitor the processor and thermal system before running it again. 

Radiation also creates a different set of problems for computing hardware in orbit. High-energy particles can interfere with electronics and cause errors, so Google's engineers are monitoring how the processors respond to the radiation and thermal conditions of space. Results from the mission should help determine whether the hardware can operatereliably over longer periods outside the controlled environment of a terrestrial data center.  

From One Satellite to a Computing Network 

Google's research extends beyond a single spacecraft; its published work describes a possible network of satellites carrying AI processors and communicating with one another through optical links. One research model uses 81 satellites flying in low Earth orbit, although Google has not announced plans to build that specific constellation. Such a system would depend on the satellites exchanging data quickly enough for their processors to work together as a distributed computing system.  

High-speed communication is central to that concept. Google has been studying free-space optical links for moving large amounts of data between satellites and has demonstrated an 800-gigabit-per-second connection between two terminals in ground testing. The company plans to test satellite-to-satellite optical communications with two spacecraft in 2027.  

Much of the interest in orbital computing comes down to electricity. Training and operating AI models requires large amounts of computing power, while data centers need a steady supply of electricity to keep those processors running. Google's research estimates that solar panels in suitable orbits could be up to eight times more productive than panels on Earth because of the amount and consistency of sunlight available in those orbits. 

Several practical obstacles remain before an orbital computing system could become a commercial proposition. Launching hardware into orbit is expensive, spacecraft have limited operating lives and repairs are far more difficult than they are in a terrestrial facility. Google's own research also identifies launch economics, radiation, thermal management, orbital dynamics and high-bandwidth communication as problems that need to be solved before the concept becomes practical.  

Suncatcher is still a small experiment, with four processors operating on a single satellite. Its value for Google will come from what the company learns about running AI hardware in orbit and whether those lessons support the more ambitious system described in its research. The results will help determine whether space-based computing remains a research project or becomes another option for meeting the industry's growing demand for AI infrastructure. 

Related Articles 

SpaceX's Stock Volatility Has a Simple Explanation Almost Nobody Can Sell It Yet 

A Decade After Falcon 9 China’s LandSpace Lands Its Own Rocket Booster 

AI Has a Power Problem Velaura Raised $110 Million to Solve It