Google plans to deploy solar-powered satellites with AI chips in low Earth orbit for space-based ML data centers
The project aims to generate round-the-clock power directly from the Sun in space
Falling launch costs could make space-based AI operation cost-competitive with terrestrial data centers, it adds
Early prototypes are planned for launch by 2027 to test chip performance in partnership with Planet
Global technology giant Google has announced Project Suncatcher, a new research ‘moonshot’ to explore the use of solar energy to power next-generation artificial intelligence (AI) systems in space.
Under Project Suncatcher, it plans to send an interconnected network of solar-powered satellites equipped with its AI chips called Tensor Processing Units (TPUs) into a low-earth orbit (LEO) to operate in space.
Google believes that the rapidly growing demand for electricity by the AI industry will need to be met by solar energy, as the Sun is the largest source of power in the solar system, emitting more than 100 trillion times humanity’s total electricity production. “At some point in the future, the best way to power AI will likely thus be to more directly tap into that enormous source of energy,” reads Google’s research.
Space-based solar power is not a new idea. Satellites in space already use solar cells to generate electricity from the Sun. However, beaming this power back to the Earth is the missing piece of the puzzle (see NASA Releases Report On Space-Based Solar Power).
But Google doesn’t want to beam this power back to the Earth. “Instead of transmitting power to Earth from space, we propose a future that includes space-based ML “data centers” consisting of many solar-powered satellites networked via free-space optical inter-satellite links,” says Google. “While there are a number of challenges that would need to be addressed to realize this “moonshot,” in the long run it may be the most scalable solution, with the additional benefit of minimizing the impact on terrestrial resources such as land and water.”
Google researchers say the concept could one day create a distributed network of solar-powered satellites capable of running complex AI models in real time. In its research paper, Google shares its initial studies, key research areas such as satellite constellation design, communication systems, and initial learnings from radiation testing of TPU chips under space conditions.
As for costs, US data centers spend about $570–$3,000 per kW each year on power, according to its research. If launch costs of LEO fall to $200 per kg, the lifetime cost of sending and operating equipment in space could be similar to these terrestrial energy costs.
It now plans to launch 2 prototype satellites by early 2027 to test its hardware in orbit, in partnership with satellite imagery and geospatial solutions company Planet.