Google Unveils Project Suncatcher To Explore Solar-Powered AI Satellites For Orbit-Based Machine Learning

In Brief

Google just unveiled Project Suncatcher, a moonshot research project exploring the use of solar satellites equipped with its AI chips for running AI workloads in orbit.

Technology company Google announced Project Suncatcher, a research initiative investigating the deployment of solar-powered satellites equipped with AI chips to run AI workloads in orbit, using sunlight to reduce the energy demands of Earth-based data centers

The project envisions compact constellations of satellites carrying Google TPUs, interconnected via free-space optical links, offering potential for large-scale computation while limiting impact on terrestrial resources

Initial findings are detailed in a preprint paper titled “Towards a future space-based, highly scalable AI infrastructure system design,” which addresses key challenges such as high-bandwidth satellite communication, orbital dynamics, and radiation effects on computing

Project Suncatcher continues Google’s tradition of pursuing ambitious, high-impact scientific and engineering projects.

Assessing Feasibility Of ML Infrastructure For Space-Based AI Satellites

According to the announcement, the proposed system envisions a network of satellites operating in a dawn–dusk sun-synchronous low Earth orbit to maximize continuous solar exposure and minimize reliance on heavy batteries

Achieving this vision requires overcoming several technical challenges. First, inter-satellite links must reach data center-scale bandwidth, supporting tens of terabits per second, which is feasible using multi-channel dense wavelength-division multiplexing (DWDM) and spatial multiplexing in close satellite formations. Bench-scale tests have already demonstrated 800 Gbps one-way transmission per transceiver pair

Second, maintaining tightly clustered satellite formations demands precise orbital control. Using physics models based on Hill-Clohessy-Wiltshire equations and refined with differentiable simulations, the team has shown that clusters with satellites hundreds of meters apart can remain stable with modest station-keeping maneuvers

Third, the TPU accelerators must tolerate space radiation; tests of Google’s Trillium v6e Cloud TPU showed that components remained operational under doses highly above expected five-year mission exposure

Finally, economic feasibility hinges on declining launch costs, which projections suggest could drop below $200 per kilogram by the mid-2030s, potentially making space-based AI data centers comparable in cost per kilowatt-year to terrestrial facilities.

Google Explores Feasibility Of Space-Based AI With Plans For Prototype Satellite Mission

Initial assessments indicate that space-based machine learning computation is feasible and not fundamentally limited by physics or prohibitive costs, though substantial engineering hurdles remain, including thermal regulation, high-bandwidth ground communications, and reliable on-orbit operation

In order to address these challenges, a learning mission in collaboration with Planet is planned, targeting the launch of two prototype satellites by early 2027 to test TPU performance in space and validate optical inter-satellite links for distributed ML workloads. In the longer term, large-scale gigawatt constellations could adopt more integrated satellite designs that combine compute architectures optimized for space with tightly coupled solar power collection and thermal management, similar to how modern system-on-chip technology advanced through smartphone innovation.