Sustainable AI infrastructure: A scenario-based forecast of water footprint under uncertainty

Overview

The rapid expansion of AI and cloud computing is creating a significant but often overlooked impact on global water resources. The study distinguishes water consumption for operational use at the facility, off-site water consumption related to electricity generation, and embodied water associated with hardware manufacturing and supply chains. Total water is modelled as on-site W_on = e(t) × WUE(t), plus off-site W_off = e(t) × PUE(t) × EWIF(t), plus embodied water divided by equipment lifespan; each parameter declines at a scenario-specific annual improvement rate.

A Monte Carlo framework inspired by Bayesian methods samples uncertain server energy and applies scenario growth and efficiency-improvement rates, run for business-as-usual (BAU), moderate-intervention and sustainable-transformation pathways. Mean projected global data-centre water use (billion L/day) is 4.18 (2030) and 28.11 (2050) under BAU, 3.06 and 8.36 under moderate intervention, and 1.86 and 1.38 under sustainable transformation. Under BAU, water demand rises nearly sevenfold between 2030 and 2050; improvements in WUE, EWIF and embodied water could cut 2050 use by up to 95% against BAU.

The sensitivity analysis finds compute growth, WUE and EWIF dominate the outcome: against the 2050 BAU baseline of 28.11 billion L/day, a ±20% shift in WUE changes demand by 4.69 billion L/day (16.7%), EWIF by 4.24 (15.1%) and embodied water by 1.29 billion L/day. A qualitative comparison puts evaporative cooling as high water and high efficiency, air-based cooling as low water and moderate-low efficiency, direct liquid cooling as low water and very high efficiency, and immersion as very low water and very high efficiency.

Limits: the model works at global scale with sparse empirical data and expert-informed assumptions. The authors call for closed-loop and immersion cooling, standard water accounting and climate-aware siting, and argue that this trajectory is not inevitable: improvements in cooling performance, alternative water sources and siting could substantially curb demand.

Key points

1. Total water is modelled as on-site W_on = e(t) × WUE(t), plus off-site W_off = e(t) × PUE(t) × EWIF(t), plus embodied water divided by equipment lifespan; each parameter declines at a scenario-specific annual improvement rate.
2. Mean projected global data-centre water use (billion L/day): 4.18 (2030) and 28.11 (2050) under BAU, 3.06 and 8.36 under moderate intervention, 1.86 and 1.38 under sustainable transformation.
3. Under BAU, water demand rises nearly sevenfold between 2030 and 2050; improvements in WUE, EWIF and embodied water could cut 2050 use by up to 95% against BAU.
4. Against the 2050 BAU baseline of 28.11 billion L/day, a ±20% shift in WUE changes demand by 4.69 billion L/day (16.7%), EWIF by 4.24 (15.1%) and embodied water by 1.29 billion L/day.
5. Qualitative comparison: evaporative cooling is high water and high efficiency; air-based cooling low water and moderate-low efficiency; direct liquid cooling low water and very high efficiency; immersion very low water and very high efficiency.
6. Compute growth, WUE and EWIF dominate the outcome.

Using it in design

• In water-footprint planning, calculate on-site (WUE), off-site electricity (PUE × EWIF) and embodied hardware water as separate items.
• In cooling selection, weigh water and efficiency together: evaporative cooling is high water, closed-loop and immersion low water; a ±20% shift in WUE changes 2050 BAU demand by 16.7%.
• Site by regional climate and grid EWIF and use standard accounting in water reporting.

Who should read it

Designers, sustainability managers and policy specialists planning the water impact of AI infrastructure.

Document contents

1. 1. Introduction
2. 2. Drivers of AI data centre water consumption (2.1 AI water footprint: impact at scale, 2.2 AI water footprint: quantification)
3. 3. Cooling systems and their role in operational water consumption (3.1 Air-based cooling systems, 3.2 Evaporative cooling towers, 3.3 Liquid and immersion cooling, 3.4 Adoption rate of cooling technologies)
4. 4. Regional variations in data centre water consumption
5. 5. Probabilistic scenario-based forecasting (5.1 Probabilistic scenario-based forecasting, 5.2 Probabilistic scenario projections from scenario-based simulations)
6. 6. Discussion and recommendations (6.2 Probabilistic scenario-based framework: strengths and limitations, 6.3 Interpretation of results and policy implications)
7. 7. Conclusions

Related standards

• ISO/IEC 30134-9 (WUE)

Read the document

Publication details

Authors

Manuel Herrera; Xiang Xie; Andrea Menapace; Ariele Zanfei; Bruno M. Brentan

Publisher

Journal of Cleaner Production (Elsevier BV)

Year

2025

Volume

526

Pages

146528

Document language

English

Source type

Scientific

Access

Open licence — downloadable

DOI

10.1016/j.jclepro.2025.146528

Licence

CC BY 4.0 — share with attribution

Licence evidence

eartharxiv.org

Citation