Study finds EU data centre efficiency metrics may worsen AI's environmental harm

A new preprint from Oxford researchers (Onitiu, Wachter, Mittelstadt) tackles the growing environmental burden of AI data centres, arguing that current EU regulations inadvertently worsen the problem. The paper focuses on the recast Energy Efficiency Directive's benchmarks—PUE (Power Usage Effectiveness) and WUE (Water Usage Effectiveness)—which data centres must report. The authors coin the term 'efficiency paradox': optimising for these metrics can lead to retrofitting larger facilities that consume more electricity and fresh water, especially for generative AI workloads. For example, lowering PUE by adding more cooling equipment may increase overall energy use, while reducing water consumption can shift the burden to local grids.

The paper proposes three policy interventions to fix the framework: (i) measures to reveal and certify genuine efficiency improvements, (ii) mandatory documentation of trade-offs when PUE or WUE scores improve, and (iii) a monitoring framework to track diminishing returns and countereffects over time. These steps aim to make the EU's common rating scheme 'fit-for-purpose' by balancing sustainability with AI innovation. The authors warn that without such corrections, hyperscale data centre expansion will continue to strain resources under a false sense of ecological progress.