Google today unveiled Project Suncatcher, a research effort to explore running machine learning workloads on a constellation of solar-powered satellites equipped with its Tensor Processing Units. The company says a dawn-dusk sun-synchronous low Earth orbit could allow solar panels to generate up to eight times more power than on Earth, reducing the need for batteries and enabling near-continuous operation.

The satellites would communicate via free-space optical links, potentially at tens of terabits per second, and fly in close formation — hundreds of meters apart — to distribute ML tasks across many chips. Google says it has already radiation-tested TPU v6e (Trillium) chips, finding them “surprisingly radiation-hard” for space: they only began to show irregularities after a cumulative dose nearly three times the expected five-year mission dose, with no hard failures observed.

Launch costs would need to fall below $200 per kilogram by the mid-2030s for space-based data centers to approach the energy costs of terrestrial ones on a per-kilowatt-per-year basis. Google’s preprint describes initial work on satellite design, control, and communication.

Partnering with Planet, Google plans to launch two prototype satellites by early 2027 to test how models and TPU hardware operate in space and to validate optical inter-satellite links for distributed ML. The company acknowledges engineering challenges remain, including thermal management, high-bandwidth ground communications, and on-orbit reliability.