Towards a future space-based, highly scalable AI infrastructure system design

A new proposal envisions fleets of satellites in low-Earth orbit running machine learning computations on solar power, according to a preprint posted to arXiv. The design uses Google tensor processing units and free-space optical links to create a space-based AI infrastructure. The paper, authored by Thomas Fischbacher and submitted in November 2025, explores a scalable compute system that would place satellites equipped with solar arrays and Trillium TPU accelerator chips into close-proximity formations [1]. The Sun is the largest energy source in the solar system, the authors note, and the design aims to tap that power directly for AI workloads [1]. To maintain high-bandwidth, low-latency communication, the satellites would use inter-satellite links via free-space optics and fly in tight clusters. The paper illustrates the formation-flight approach with an 81-satellite cluster contained within a 1 km radius [1]. High-precision machine-learning models would control the large-scale constellations [1]. Radiation testing of the Trillium TPUs indicates they can survive a total ionizing dose equivalent to a 5-year mission life without permanent failures, and the chips have been characterized for bit-flip errors [1]. Launch costs remain a critical variable. A learning-curve analysis in the paper suggests that the cost of launching payloads to low-Earth orbit could fall below $200 per kilogram by the mid-2030s [1]. The proposal arrives as demand for AI compute and energy continues to grow. Foundation models — large machine-learning models trained on vast datasets — are often highly resource-intensive, with the most advanced examples costing hundreds of millions of dollars to develop, driven by the need for sophisticated infrastructure and extended training times on hardware such as GPUs [4]. Large language models, a common type of foundation model, underpin modern chatbots and are typically based on transformer architectures [3]. The preprint was posted on arXiv, a repository that hosts scholarly papers before peer review and is part of a broader ecosystem of open-access platforms that assign digital object identifiers to establish provenance [5][7]. The paper has not been evaluated by a scientific journal [5].