Google plans to send artificial intelligence (AI) data centers into space and operate them as a constellation of satellites that use sunlight as their primary energy source. Their plan for the 2030s, recently outlined in a document awaiting peer review, is called Project Suncatcher.
With this offering, Google aims to address the biggest current limitation of artificial intelligence: its enormous power consumption. As machine learning systems become more powerful, they also require more electricity and more investment. According to the company’s calculations, if the space industry maintains its pace of progress, by the middle of the next decade it will be possible to directly use solar energy from low Earth orbit.
Three of the biggest US tech giants announced record profits and infrastructure spending, fueling speculation of a possible AI bubble.
AI in space, inside the swarm
“Project Suncatcher” plans to launch dozens of artificial intelligence servers into space, which will be powered by solar radiation, rather than relying on large data centers on Earth. These swarm servers will communicate with each other via high-speed optical communication channels, spinning in a sun-synchronous fashion. In other words, Google’s minisatellites will move in such a way that they stay in sunlight most of the time.
The document reports that Google has already passed the first viability tests. The researchers modeled a cluster of 81 satellites, distributed within a radius of one kilometer at a distance of 100 to 200 meters, that remained stable during their orbit. In the lab, they have also achieved transfer rates of up to 1 terabit per second (Tbps) using commercially available optical technology. In addition, they tested their specialized AI TPU chips under radiation conditions equivalent to five years in orbit, and they survived without permanent failure.
Radiation, temperature and transmission speed, the main problems
Before launching the prototype into space, Google admitted that it must finalize these tests. The company needs to improve the maneuverability of its satellite formations to cope with the Earth’s complex orbital dynamics, increase transmission speeds to 10 Tbit/s and strengthen the resistance of its chips to the solar wind.
Google noted that the cost of a space launch into low orbit should be below $200 per kilogram. The price is calculated based on the total cost of launching the rocket divided by the payload it carries. Today, it costs $67 million to launch a Falcon 9, the rocket NASA uses to deliver the International Space Station, the equivalent of $2,500 to $3,000 per kilogram of cargo. SpaceX previously said it would reach $100 per kilogram by 2035.
Among the unsolved problems that need to be solved are the dissipation of heat in the vacuum created by the chips and how to manage communications on Earth in the event of atmospheric turbulence.
“Our initial analysis shows that the basic concepts of ML computing (Machine learning) space-based is not hindered by fundamental physics or insurmountable economic barriers. However, significant engineering challenges remain, such as temperature management, high-bandwidth ground communications, and system reliability in orbit,” the company said in a statement.
