Google exploring putting AI data centers in space — Project Suncatcher wants to harness in-orbit solar power to scale AI compute

Google just announced that it’s exploring the idea of putting AI data centers into orbit to take advantage of the sun’s solar output for power. According to Google Research, Project Suncatcher aims to have a constellation of solar-powered satellites with Google TPUs that communicate optically. This would allow the company to run a power-hungry data center without requiring the massive infrastructure needed to build one on land.

Power is currently one of the bottlenecks of the AI infrastructure build-out, with companies like Microsoft sitting on GPU inventories because it doesn’t have enough electricity for them. It's also causing electricity price hikes all over the U.S., burdening residents with the cost. Tech companies are investing in alternative technologies like small modular reactors and jet engines to solve their future power requirements, while current power supply challenges are quickly changing the makeup of power grids.

Solar power is a clean source of energy that data centers can use for power, but it takes up a lot of space and is subject to the day-and-night cycle and atmospheric conditions. Google suggests that by putting its satellite in a dawn-dusk sun-synchronous low-earth orbit (meaning the satellite would orbit the earth’s terminator, or the dividing line between day and night), it would be illuminated by the sun nearly 100% of the time. This would allow it to produce about eight times more power than ground-based solar panels, making space-based AI data centers potentially more scalable.

While the initial study has proven the viability of the concept, Google says that it still has significant engineering challenges to overcome. Some of these include developing a wireless link between satellites that can support massive amounts of data, controlling a compact constellation of satellites so that they work in unison and avoid colliding with one another and other orbiting bodies, ensuring the radiation resistance of semiconductors in space, and managing the costs of launching the entire infrastructure into space. The company says that it plans to launch two prototypes in 2027 to test the system and check its viability for machine learning.

Putting AI data centers might help reduce their impact on the electrical supply on Earth, but it also comes with its own set of issues. These include increasing the number of orbiting space junk around our planet and the possible disruption of ground-based astronomical observation, among others. The project is still in its early stages, though, so scientists and researchers have time to consider these things before launching a full-scale constellation in low-earth orbit.

Follow Tom's Hardware on Google News, or add us as a preferred source, to get our latest news, analysis, & reviews in your feeds.