Space-Based versus Land-Based AI Data Centres: A Comparative Environmental Assessment Centred on Google's Orbital Compute Programme

This paper compares space-based AI compute against terrestrial hyperscale data centres across six environmental parameters, using Google's published research programme on solar-powered TPU constellations in dawn-dusk low-Earth orbit as the reference case (Arcas et al., 2025). The comparison uses a dual functional unit — 1 MW-year of delivered IT power, and 1 MW of IT power over 1-, 5- and 10-year missions — under a cradle-to-grave boundary that includes launch and re-entry for the orbital case. Three findings emerge. First, the carbon result is functional-unit dependent: per kWh of energy delivered to the payload, orbital systems are 1.5–1.95× more carbon-intensive than terrestrial ones (52.1–66.3 versus 34.0 gCO₂e/kWh) (Ohs et al., 2025); but per MW of IT capacity against real-world grids, launch-phase emissions of 1,032–2,585 t CO₂e are repaid within the first year and amount to only 5.5–13.7% of five years of terrestrial operation at the 395.65 gCO₂e/kWh global data-centre average (Vries-Gao, 2025). Break-even occurs at grid intensities of 21.6–54.1 gCO₂e/kWh over five years. Second, water and land are where space wins categorically: no evaporative cooling, and no terrestrial land conversion of the tens-to-hundreds of hectares a hyperscale facility requires (Hlabisa, 2025). Third, the binding physical constraint in orbit is not compute silicon but radiative heat rejection — 33–67 cm² of radiator per watt of waste heat, or 3,333–6,667 m² per MW (Vicente, 2021) — which converts thermal inefficiency directly into launch mass and launch carbon.