Google is exploring a bold new frontier for artificial intelligence: space-based computing. The company’s latest initiative, Project Suncatcher, investigates the potential of using solar-powered satellites in low Earth orbit to create a new kind of data center—harnessing the sun’s energy in ways that are far more efficient than on the ground. The sun produces over 100 trillion times more energy than humanity currently consumes. In space, solar panels can generate up to eight times more power than those on Earth, operating almost continuously without the need for large battery storage. This abundance of clean, consistent energy is at the heart of Google’s vision: to use space as a platform for scaling AI workloads in a sustainable and high-performance way. The proposed system would consist of constellations of small satellites, each equipped with powerful processors and connected via high-speed laser-based optical links. These satellites would operate in a sun-synchronous low Earth orbit, where they remain in near-constant sunlight, minimizing the need for energy storage and maximizing power generation. A key challenge for any space-based data center is achieving the high-speed, low-latency communication required for large-scale machine learning. Google’s research shows that data center-level performance is possible using advanced technologies like dense wavelength division multiplexing and spatial multiplexing. The team has already demonstrated a bench-scale version of this system, achieving a total transmission rate of 1.6 terabits per second—proof of concept for the high-bandwidth connections needed. However, such performance requires satellites to fly in extremely tight formation, separated by just a few kilometers. Maintaining these precise configurations at an altitude of about 650 kilometers presents a significant engineering challenge. Google developed detailed physics simulations to model the effects of Earth’s gravitational field, atmospheric drag, and other forces. The results suggest that only minor station-keeping maneuvers would be needed to keep the satellites in stable, coordinated orbits. Another critical question is whether existing hardware can survive the harsh space environment. Google tested its Trillium v6e Cloud TPU chips and found they are surprisingly resilient. The processors withstood radiation doses nearly three times higher than expected over a five-year mission before showing any irregularities. While the High Bandwidth Memory systems were more sensitive, they only began to fail after exposure to 2 kilorads—well above the 750 rads expected for a shielded mission. The economic viability of the project hinges on the continued decline of launch costs. Google’s analysis projects that with advancements in launch technology, prices could fall below $200 per kilogram by the mid-2030s. At that point, the cost of launching and operating a space-based data center could become comparable to the energy costs of a similar Earth-based facility. If these technical and financial hurdles can be overcome, space-based AI computing could offer a sustainable, high-performance solution to the growing energy demands of machine learning—ushering in a new era where the next generation of artificial intelligence isn’t built on Earth, but above it.