Boundary-consistent screening of electricity, carbon and water footprints for generative-AI inference: a reproducible methodological note

Abstract

Environmental footprint estimates for generative artificial-intelligence (AI) inference are often compared despite differences in workload, reporting statistic and infrastructure accounting boundary. This methodological technical note proposes a conservative screening protocol for deciding whether published electricity, carbon and water estimates can be compared directly, compared only conditionally, or should remain incomparable. We demonstrate the protocol using two distinct evidence types: a published model-based estimate of 0.31 Wh for standard inference and 3.91 Wh for long-reasoning inference, and a separately reported production-system median of 0.24 Wh per text prompt. For one million modelled requests, a hypothetical 10% long-reasoning share yields 670 kWh, compared with 310 kWh under an all-standard workload. At an explicitly assumed electricity emissions factor of 0.4 kg CO₂e/kWh, these totals correspond to 268 and 124 kg CO₂e, respectively. Water-consumption illustrations are kept separate from production measurements because their coefficients are not facility-specific observations. The proposed screening protocol requires alignment of functional unit, workload, reporting statistic, infrastructure boundary, grid accounting and water-consumption scope before numerical comparison. The examples demonstrate how apparently conflicting published footprints can arise from incompatible measurement conventions rather than contradictory evidence. No new metered data, independent cross-provider validation or priority claim is offered. The contribution is a reproducible comparison checklist and transparent boundary-sensitive demonstration intended to reduce inappropriate cross-study inference.