Google announced Project Suncatcher, a research initiative to launch AI chips into space aboard solar-powered satellites, aiming to establish orbital data centers that utilize continuous solar energy to address terrestrial data center energy demands and emissions.

The project represents Google’s effort to overcome limitations in powering AI infrastructure on Earth. Traditional data centers consume substantial electricity, contributing to increased power-plant emissions and elevated utility costs. By relocating components to space, Google seeks to tap into solar power available nearly around the clock. Satellites in orbit avoid nighttime interruptions and atmospheric interference, enabling more consistent energy generation. This approach aligns with the company’s broader AI development goals, where computational demands continue to escalate.

Details of the project appear in a preprint paper released by Google, which outlines initial advancements without peer review. The paper, along with a blog post, emphasizes the potential of space for expanding AI capabilities. Travis Beals, senior director for Paradigms of Intelligence at Google, stated in the blog post, “In the future, space may be the best place to scale AI compute.” This sentiment echoes in the preprint, reinforcing the project’s focus on orbital computing.

Central to Project Suncatcher are Google’s Tensor Processing Units, or TPUs, designed for AI workloads. These units would orbit Earth on satellites equipped with solar panels. The panels in space generate electricity almost continuously, achieving productivity levels eight times higher than equivalent panels on the ground. This efficiency stems from uninterrupted sunlight exposure above the atmosphere, free from weather variations or day-night cycles that limit terrestrial solar installations.

Communication between satellites poses a significant technical obstacle. To rival ground-based data centers, inter-satellite links must support data transfer rates of tens of terabits per second. Achieving such bandwidth requires precise coordination, with satellites positioned in formations mere kilometers or less apart. Current satellite operations maintain greater distances, so this proximity demands advanced maneuvering capabilities. Google identifies this closeness as essential for high-speed data exchange in a constellation network.

The tighter formations also introduce risks from space debris. Existing orbital junk from past collisions already threatens active satellites. Closer groupings amplify collision probabilities, necessitating robust collision avoidance systems and ongoing orbital debris monitoring. Google acknowledges these challenges as critical to the project’s feasibility, requiring innovations in satellite design and trajectory control.

Radiation exposure in space presents another hurdle for electronic components. Unlike Earth environments, orbits subject hardware to intense cosmic and solar radiation, which can degrade performance. Google conducted tests on its latest Trillium TPUs to assess durability. Results show the units endure a total ionizing dose comparable to five years in orbit without experiencing permanent failures. These evaluations confirm the TPUs’ resilience under simulated space conditions, supporting their use in long-duration missions.

Launch and operational expenses currently make space-based systems cost-prohibitive. However, Google’s analysis projects cost reductions over time. By the mid-2030s, expenses for launching and maintaining a space data center should align closely with energy costs for a comparable Earth facility, measured on a per-kilowatt-year basis. Factors include declining launch prices and improved satellite manufacturing efficiencies.

To advance the initiative, Google collaborates with Planet, an Earth observation company. The partners plan a joint mission to deploy prototype satellites by 2027. This test will evaluate hardware performance in actual orbital conditions, gathering data on power generation, communication, and radiation effects to refine future implementations.

Featured image credit