To AI or not to AI? Per-task energy and water demands of AI inference are orders of magnitude lower than human labor across all scenarios

Abstract The aggregate resource footprint of artificial intelligence infrastructure is growing rapidly, but per-unit comparisons against the human labor that AI augments or replaces have not been established. We construct a sector-decomposed lifecycle accounting framework for a representative US adult, covering transportation, aviation, residential energy, food systems, manufactured goods, and water supply, yielding a continuous power demand of approximately 6,700 W and daily water footprint of approximately 7,800 L. Normalized to a standard 40-hour working week, these demands correspond to 467 Wh and 22.7 L per working minute.

For large language model text inference on current hardware, we estimate 1.5 Wh and 3.9–4.6 mL per operational minute; for AI image generation, approximately 2.9 Wh and 7.5–8.9 mL per task. At a conservative 100:1 human-to-AI productivity ratio, the human lifecycle energy demand exceeds text AI inference by approximately 31,000-fold and image generation AI by approximately 16,000-fold; water demand exceeds both text and image generation AI by 70,000- to 1,900,000-fold depending on methodology. These results are robust across hardware, grid, and methodology assumptions and improve substantially at productivity ratios commensurate with observed AI task performance.

Our findings provide a missing per-unit baseline for evaluating the net resource implications of AI adoption.