The language of the text is English.
Researchers at Google are exploring a radical approach to mitigate the growing environmental and resource footprint of artificial intelligence by moving data centers into orbit around the Earth. A new study outlines a long-term vision for a vast, space-based AI infrastructure powered directly by the sun. This system would operate as a distributed supercomputer composed of satellite constellations using Google’s Tensor Processing Unit (TPU) hardware. The initial findings suggest the concept is not only technologically feasible but could also become economically competitive with terrestrial data centers within the next decade.
To overcome the limitations of land, water, and power grids on Earth, the proposed system would leverage the constant and abundant solar energy available in space. The concept involves placing fleets of satellites into a sun-synchronous low-Earth orbit (LEO), which provides almost continuous sunlight for the solar panels. Rather than attempting to build a single massive station, the architecture is modular, relying on arrays of smaller, interconnected satellites. This design simplifies deployment and allows the computational network to be expanded incrementally over time.
A primary technical obstacle for a space-based supercomputer is achieving the extremely high-bandwidth, low-latency communication that large-scale AI models demand. The research paper proposes a solution using formation flying, where satellites are arranged in tight clusters with separations ranging from a few kilometers down to several hundred meters. This proximity would enable the use of commercial optical networking technology to establish inter-satellite links capable of terabit-per-second throughput, vastly outperforming current satellite communication systems. The study successfully modeled a stable 81-satellite cluster to demonstrate the viability of this formation.
To determine if standard hardware could survive the harshness of space, Google tested its latest Trillium TPU accelerators by subjecting them to proton beam radiation simulating the LEO environment. The results indicated that the chips could withstand a total radiation dose equivalent to a five-year mission without catastrophic failure. While the radiation induced manageable data errors, particularly in the High Bandwidth Memory, the rate was low enough to be acceptable for many AI applications. This finding is significant, as it suggests that costly, specialized radiation-hardened components may not be necessary.
The ultimate feasibility of the project is heavily dependent on the economics of space launch. The analysis concludes that if launch costs to LEO decrease to approximately $200 per kilogram, a target anticipated to be reachable by the mid-2030s with fully reusable rockets, the cost of launching the solar power infrastructure could rival the ongoing energy expenses of a ground-based facility. While substantial challenges such as thermal management in a vacuum, high-speed communication with Earth, and on-orbit repair strategies must still be solved, this research establishes a foundational blueprint for a future where AI computation is powered from space.
https://services.google.com/fh/files/misc/suncatcher_paper.pdf