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

Technical Summary: Per-task Energy and Water Demands of AI Inference vs. Human Labor

Problem Statement
While the aggregate environmental footprint of artificial intelligence (AI) infrastructure is rapidly growing, a critical analytical gap remains: the lack of established per-unit comparisons between AI inference and the human labor it augments or replaces. Prior studies have focused on macroscopic sectoral trajectories (e.g., total water and carbon emissions of US AI deployments by 2030) or specific training costs. However, these aggregate analyses cannot answer whether AI adoption results in a net resource saving or a net increase. Without a common functional unit, it is impossible to distinguish between the Jevons paradox (where efficiency gains lead to increased aggregate consumption) and genuine resource displacement. This paper seeks to establish a missing per-unit baseline to evaluate the net resource implications of AI adoption.

Methodology
The authors construct a sector-decomposed lifecycle accounting framework to compare a representative US working adult against AI inference tasks.

1. Functional Unit: The study normalizes all demands to a "working minute."
   • Human Baseline: A continuous power demand of ~6,700 W and a daily water footprint of ~7,800 L are calculated for a representative US adult across seven sectors: personal transportation, aviation, residential energy (electricity and gas), food systems, manufactured goods, and municipal water/waste. These are normalized to a standard 40-hour work week (2,400 minutes), yielding 467 Wh and 22.7 L per working minute.
   • AI Baseline: Inference demands are calculated for Large Language Models (text) and image generation on current hardware (e.g., NVIDIA H100).
2. Water Accounting Decomposition: The study introduces a two-component decomposition for AI water demand:
   • Direct Cooling: Measured via Water Usage Effectiveness (WUE).
   • Indirect Electricity Generation: Calculated based on grid water intensity.
   • The "Hydro Paradox": The methodology accounts for hydroelectric reservoir evaporation. The authors find that including evaporation reverses the water intensity ranking between the US national grid and the California grid, as California's higher reliance on hydro (13.4% vs. 6% national) results in a higher comprehensive water intensity (3.068 L kWh⁻¹ vs. 2.288 L kWh⁻¹).
3. Comparative Scenarios: The analysis spans a matrix of hardware configurations, cooling technologies (WUE ranging from 0.2 to 9.0 L kWh⁻¹), and electricity grid mixes (US national vs. California, with and without hydro evaporation). Productivity ratios are tested, starting with a conservative 100:1 (human-to-AI) ratio and extending to observed performance ratios (up to 3,600:1).

Key Results
The study finds that the per-task resource demands of AI inference are orders of magnitude lower than those of human labor across all tested scenarios.

• Energy:
  • Text Inference: Requires ~1.5 Wh per operational minute.
  • Image Generation: Requires ~2.9 Wh per task.
  • Comparison: At a conservative 100:1 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. Even at a 1:1 productivity parity, the human lifecycle uses 311 times more energy for text and 161 times more for image generation.
• Water:
  • Text Inference: Ranges from 1.2 mL to 16.9 mL per minute depending on WUE and grid methodology.
  • Image Generation: Ranges from 2.3 mL to 32.8 mL per task.
  • Comparison: Human water demand exceeds text AI by 134,000- to 1,925,000-fold and image generation AI by 69,000- to 995,000-fold, depending on the specific WUE and grid assumptions.
• Robustness: These conclusions hold even under "worst-case" assumptions (e.g., older V100 hardware, high WUE, and comprehensive hydro evaporation accounting). The gap widens significantly when using productivity ratios observed in real-world AI performance (e.g., 240:1 to 3,600:1).

Significance and Claims
The paper claims to provide the first per-unit baseline necessary for a complete policy analysis of AI adoption. Its primary significance lies in reframing the debate from aggregate consumption to per-unit efficiency relative to the displaced activity.

• Policy Implication: The authors argue that inferring a net increase in societal resource use solely from growing aggregate AI footprints is misleading without a human comparator. If AI displaces a more resource-intensive human activity, the net result is a resource saving, even if the absolute number of AI queries grows.
• Methodological Contribution: The study highlights the "hydro paradox," demonstrating that grids perceived as sustainable (like California's) may carry higher lifecycle water intensities when reservoir evaporation is included. It recommends that AI infrastructure disclosures explicitly distinguish between three water scopes: (i) on-site evaporative loss (WUE), (ii) indirect electricity water excluding reservoir evaporation, and (iii) indirect electricity water including reservoir evaporation.
• Limitations: The authors acknowledge that their human budgets represent US national averages, exclude embodied energy in semiconductor manufacturing, and assume 100% human productivity during working hours. They also note that systems-level rebound effects (where AI enables more total work rather than substitution) are not accounted for.

In summary, the paper concludes that while data center water and energy use are growing, the per-task resource cost of AI is negligible compared to the human labor it replaces, suggesting that AI adoption could represent a net reduction in resource intensity per unit of output.