data centers in space to meet demands of AI use on earth
Google explores building data centers in space using constellations of solar-powered satellites to meet the computing demands of AI machines and models on Earth. Named Project Suncatcher, the research project expects to use the energy directly from the sun to run the data centers equipped with solar panels, which, as a result, can handle more computing power and energy used by machine learning systems and language models on Earth.
The proposal describes an AI space-based data center made of many small satellites, each equipped with solar panels, computing chips, and communication systems. These satellites would orbit the Earth in a formation, creating a cluster that works like one large computer network. The system would use Google Tensor Processing Units, which are chips made for machine learning and have already been tested for radiation and can survive about five years in orbit without permanent damage. The design includes solar arrays to collect energy, heat pipes and radiators to control temperature, and free-space optical communication links to connect the satellites.
image courtesy of NASA, via Unsplash
Google’s project suncatcher uses solar-powered satellites
An additional design element for Google’s AI data centers in space includes optical links that use light instead of radio waves to send data for faster communication between satellites and reduce delay. Each satellite would fly close to others in a formation about one kilometer wide, and the formation would be managed by machine learning models that control position and movement to avoid collisions. The satellites would orbit in constant sunlight to help the solar panels collect power without interruption, and the position also reduces the time needed to send data back to Earth (but the company says some delay in communication might still happen depending on ground locations).
Cooling is a major part of the design. Because space has no air, heat from the chips must move through solid materials to the radiators, releasing heat into space, so the team plans to use advanced thermal interface materials to transfer heat efficiently without mechanical parts. The plan also focuses on modular satellite design by building smaller satellites instead of one large structure to reduce the risk and costs. If one satellite fails, others can continue operating. Launching smaller units is also cheaper, as future launch costs to low-Earth orbit are expected to fall to about 200 USD per kilogram by the 2030s. The research describes building AI data centers in space as a milestone since they could perform like a normal Earth-based data center. Later on, Google plans to test real satellite communication and flight control and, over time, improve each part, including better cooling, stronger chip protection from radiation, and automated fault recovery.
image courtesy of Allison Saeng, via Unsplash
image generated on Gemini using the prompt ‘solar-powered satellited’
image courtesy of Ian Battaglia, via Unsplash
project info:
name: Project Suncatcher
research: here
The post google plans to build AI data centers in space using solar-powered satellites appeared first on designboom | architecture & design magazine.