If you want to satiate AI’s hunger for power, Google suggests going to space (arstechnica.com)

Google today unveiled Project Suncatcher, a moonshot to put clusters of its Tensor Processing Units (TPUs) into low‑Earth orbit satellites — solar‑powered “swarms” linked by free‑space optical (laser) communications — to run training, generative content, speech/vision models and predictive workloads off‑planet. Early tests show Google’s TPUs can survive intense space radiation, and the company says LEO platforms could tap near‑continuous sunlight and high‑bandwidth optical links to scale ML compute beyond terrestrial constraints. The plan is significant because AI compute demand is ballooning and terrestrial data centers face enormous electricity, cooling and water bottlenecks; Project Suncatcher aims to offload some of that footprint. But it raises hard engineering and operational tradeoffs: thermal management and heat rejection in vacuum (no convection), on‑orbit reliability and maintenance, launch emissions, satellite lifetimes, and the need for robust optical ground links and constellation management. If solved, space‑based compute could expand global ML capacity and resilience, but it also introduces new latency, cost, regulatory and environmental considerations that will shape whether orbital compute becomes a complement to — rather than replacement for — Earthbound data centers.