Hyperscalers are gobbling up land for AI data centers and now Google is mapping out property in space. Known as Project Suncatcher, the plan is to launch solar-powered networks of orbiting Tensor Processing Units (TPUs) — Google’s custom-developed machine learning accelerator chips — connected by free-space optical links. Apparently, Google feels that deploying high-performance TPUs in space would not be that different from the satellites that relay Internet communications from orbit. A big advantage over Earth-based AI factories is proximity to the sun, which emits “more power than 100 trillion times humanity’s total electricity production,” per Google.

“In the right orbit, a solar panel can be up to 8 times more productive than on earth,” according to a Google research blog that emphasizes this would offer “tremendous potential for scale” while minimizing the impact on overextended terrestrial resources. “By focusing on a modular design of smaller, interconnected satellites, we are laying the groundwork for a highly scalable, future space-based AI infrastructure.”

The company shared its progress in the pre-publication scientific paper Towards a Future Space-Based, Highly Scalable AI Infrastructure System Design, which analyzes the challenges involved in coordinating orbital dynamics, controlling the effects of radiation on computing and maintaining high-bandwidth communication between satellites.

“On Earth, the nodes in a data center communicate via blazing-fast optical interconnect chips,” Ars Technica explains, noting that “high-speed communication among the orbiting servers will require wireless solutions that can operate at tens of terabits per second.” Early terrestrial testing has achieved bidirectional speeds up to 1.6 Tbps, which Google thinks can eventually scale up.

“The fact is, terrestrial data centers are dirty, noisy, and ravenous for power and water,” writes Ars Technica. These traits have led to community opposition to having one in their neighborhood. Putting them in space could solve a lot of problems.

But “it would be pretty pricey to send those TPUs into space at the moment,” observes The Verge, citing a Google cost analysis that suggests the cost of “launching and running a data center in space could become ‘roughly comparable’ to the energy costs of an equivalent data center on Earth on a per-kilowatt/year basis by the mid-2030s.”

Working with satellite imaging company Planet Labs, Google hopes to launch a pair of prototype TPU-powered satellites by early 2027. “This mission is designed to test the viability and performance of Google’s TPUs in the harsh environment of space, and the ability of two such spacecraft to work in concert, flying in tandem with high bandwidth cross link communications,” according to a Planet blog post.