Unequal Feasibility: Quantifying the Physical Limits of Low-Carbon AI Deployment Across Six European Electricity Grids

The carbon footprint of AI inference is determined first by the physical characteristics of the electricity grid rather than by scheduling or software optimization. This study quantifies the best physically achievable emissions across six European power systems by identifying the cleanest strictly contiguous multi-day windows from 2022 to 2024 using legally reported hourly data from ENTSO-E. In 2024, feasible emissions vary by a factor of 19.2, from 0.084 gCO2per inference in Sweden to 1.615 g in Poland, under fixed model energy and power usage effectiveness (PUE) assumptions. This gap persisted across window durations (3 to 30 days), model tiers, and PUE values, and was statistically validated using non-overlapping bootstrap confidence intervals (95% CI,n= 1000). All six countries show steady improvement from 2022 to 2024, yet their relative rankings remain unchanged, indicating a structural rather than temporary disparity. The resulting feasibility benchmark provides a reproducible foundation for siting AI infrastructure, sustainability reporting, and grid-aligned policy design.