Summary

Google Research has published a preprint outlining Project Suncatcher, a proposed network of solar-powered satellites carrying TPUs for machine-learning computing in orbit. Its early work includes a bench-scale optical-link demonstration, satellite-formation modelling and radiation testing of a TPU.

Google Research has published a preprint describing Project Suncatcher, a proposed system of solar-powered satellites equipped with Google TPUs and linked to share machine-learning workloads. The design is a research moonshot, not an operational space data centre. Google’s early work includes a bench-scale optical communications demonstration, models of closely grouped satellites and radiation testing of a TPU. The company says its next milestone is a two-satellite learning mission, planned with Planet for launch by early 2027.

Contents

A solar-powered computing constellation

Project Suncatcher explores placing machine-learning computing in a constellation of satellites, likely in a dawn–dusk sun-synchronous low Earth orbit. Satellites in this orbit would receive near-constant sunlight, allowing solar panels to supply power while reducing the need for heavy batteries, according to Google.

The proposed system would distribute computing across many accelerators. That requires fast connections between satellites: Google’s analysis says data-centre-scale workloads would need inter-satellite links supporting tens of terabits per second. The team proposes using optical links with dense wavelength-division multiplexing (DWDM), which sends multiple data channels at different wavelengths, alongside spatial multiplexing.

Optical signals weaken with distance, so the proposed satellites would fly in a much tighter formation than current systems. Google says distances of kilometres or less could help meet the link’s power requirements. Its bench-scale demonstrator reached 800 Gbps in each direction—1.6 Tbps total—with one pair of transceivers. This is a component-level demonstration, rather than a test of a satellite network in orbit.

To examine whether such a compact constellation could stay together, Google modelled the orbital motion of satellites under Earth’s gravity and other effects. One illustrative configuration used 81 satellites at a mean altitude of 650 km, with a cluster radius of 1 km. In the model, the distance between next-nearest neighbours oscillated between about 100 and 200 metres over an orbit. Google’s models suggest that modest station-keeping manoeuvres may be sufficient to maintain the formation.

The team also tested Trillium, Google’s v6e Cloud TPU, in a 67 MeV proton beam. The test examined total ionizing dose (TID), the cumulative effect of radiation, and single-event effects caused by individual particles. Google reports that the high-bandwidth memory subsystem began showing irregularities after a cumulative dose of 2 krad(Si), nearly three times the expected shielded dose of 750 rad(Si) over a five-year mission. It reports no TID-attributable hard failures up to 15 krad(Si) on a single chip.

These results address specific engineering questions, but they come from a bench demonstrator, orbital modelling and proton-beam testing. The planned learning mission is intended to test how the models and TPU hardware operate in space and to validate optical links for distributed machine-learning tasks.

Costs and the planned learning mission

Google’s economic analysis projects that, with a sustained learning rate in launch costs, prices could fall below $200 per kilogram by the mid-2030s. At that level, the company estimates that launching and operating a space-based data centre could be roughly comparable to the reported energy costs of an equivalent terrestrial data centre, measured per kilowatt per year. This is a forward-looking estimate tied to the projected launch price, rather than a current cost comparison.

The company identifies thermal management, high-bandwidth communications with the ground and on-orbit reliability as challenges still to address. Its preprint argues that the basic concepts are not ruled out by fundamental physics or insurmountable economic barriers; the prototype mission is a next step in testing whether the proposed system can work in space.

Sources