Summary
Google’s Project Suncatcher studies whether solar-powered satellite constellations could support large-scale AI computing. Its early work includes a bench-scale optical-link demonstration and radiation testing of a TPU; a two-satellite learning mission with Planet is planned for early 2027.
Google Research has outlined Project Suncatcher, a research effort to assess whether solar-powered satellite constellations could provide computing capacity for machine-learning workloads. The company’s November 4, 2025 announcement describes a proposed system and early component tests, not an operating orbital data centre. Google says its next milestone is a learning mission with Planet, planned to launch two prototype satellites by early 2027.
The idea is to place Google TPUs—specialised processors used for machine-learning tasks—on satellites that communicate through free-space optical links. These links transmit data as light through open space. Google argues that an orbit with near-continuous sunlight could provide more consistent solar power and reduce the need for heavy batteries.
The design depends on close-flying satellites
Large machine-learning workloads are split across many accelerators, which must exchange data quickly. Google says links comparable to those in terrestrial data centres would need to carry tens of terabits per second between satellites. Its analysis considers multi-channel dense wavelength-division multiplexing, which sends multiple data channels over different light wavelengths, alongside spatial multiplexing.
At those bandwidths, the received signal would need to be much stronger than in typical long-range optical links. Because received power falls with the square of distance, the proposed satellites would fly in close formation—kilometres apart or less. Google reports that a bench-scale demonstrator using one transceiver pair achieved 800 Gbps in each direction, or 1.6 Tbps total. That is a component-level link test, rather than a demonstration between satellites in orbit.
The company also modelled the motion of a proposed constellation. In one illustrative configuration, 81 satellites orbit at a mean altitude of 650 km, with the cluster extending to a radius of 1 km. The model puts next-nearest-neighbour distances at roughly 100–200 metres as Earth’s gravity changes their relative positions. Google’s analysis suggests that formations hundreds of metres apart may need only modest station-keeping manoeuvres to remain stable in the intended sun-synchronous orbit.
Tests, costs and the planned in-orbit step
To examine radiation effects, Google tested its Trillium, or v6e Cloud TPU, in a 67 MeV proton beam. The high-bandwidth memory subsystems were the most sensitive: irregularities began after a cumulative dose of 2 krad(Si), nearly three times the expected shielded dose for a five-year mission, which Google gives as 750 rad(Si). The company reports no hard failures attributable to total ionizing dose up to 15 krad(Si) on a single chip. These results are from proton-beam testing, not years of operation in orbit.
Google’s economic analysis suggests launch prices could fall below $200 per kilogram by the mid-2030s if a sustained learning rate continues. At that price, the company says, launching and operating a space-based data centre could become roughly comparable, on a per-kilowatt-per-year basis, to reported energy costs for an equivalent terrestrial data centre. The projection depends on future launch-price reductions; the announcement also identifies thermal management, high-bandwidth ground communications and on-orbit reliability as engineering challenges.
The planned two-satellite mission with Planet is intended to test how the models and TPU hardware perform in space and to validate optical links for distributed machine-learning tasks. It is the next step toward assessing the concept in orbit, rather than a deployment of a scaled computing constellation.