The news of November 6thGoogle The official launch of the data centre landing programme is intended to place all its calculations in space. It's called Project Suncatcher, Solar Catcher Project。
Google's idea is simple: instead of fighting for depleted resources on Earth, it's better to go to space with solar energy. This whole new moon landing programme has only one goal, creating a solar-powered, scalable AI infrastructure in space。
Specifically, Google, in order to solve the energy problems of the data centres, their plan was to launch a solar-powered, loaded Google self-study TPU A SATELLITE CONSTELLATION OF CHIPS (FOR THE COMPUTATION OF GPUS LIKE WEAVERDA) CREATES AN "ORBIT AI DATA CENTRE" IN SPACE。
AT THE SAME TIME, THEY PROPOSED SOLUTIONS TO THE PROBLEM OF AI:
In space, how to achieve low delayed data transmission:
formation flight + laser communication. Satellites are only at a kilometre scale or closer to each other, and each of them is equipped with solar arrays, radiation cooling systems and high-bandwidth optical communication modules in a analog constellation of 81 satellites。
At such a close distance, they can achieve high-speed connectivity through free space light communication (FSO ISL, Free-Space Open Inter-Satellite Links). Google disclosed in its paper that their presentation had successfully achieved a two-way transmission rate of 1.6 Tbps。
Resistance to "radiation": Google found that its TPU could run in LEO for five years without permanent damage. It is therefore planned to launch two prototype satellites by 2027, in cooperation with Planet, to test the operational environment。
Data echo: The problem of delay Google chooses the "morning-synchronous orbit" which, although solar power is full, the paper acknowledges that this will increase to some ground position delays; the current top record of "ground-air" light communications for bandwidth bottlenecks is also NASA 200 Gbps, created in 2023。
It is worth mentioning that Google has calculated in its paper that if SpaceX's launch cost can be reduced to $200/kg (projected around 2035), the unit power cost of the space data centre would be equal to that of the ground data centre, approximately $810/kW/year, with a complete overlap between $570-30000/kW/year in the United States mainland data centre。
In other words, when a rocket is cheap enough, space is better suited than the Earth for a data centre。The reality, however, is that the current launch price is more than 10 times that of this ideal price。