Making AI Less 'Thirsty': Uncovering and Addressing the Secret Water Footprint of AI Models

Overview

The paper was published in Communications of the ACM in 2025 (vol. 68, no. 7, pp. 54-61); the arXiv:2304.03271 v5 preprint (26 March 2025) was read. Its DOI is 10.1145/3724499. The authors are Pengfei Li, Jianyi Yang, Mohammad A. Islam and Shaolei Ren. The method sums on-site cooling water (via WUE) and water embedded in electricity (via PUE x EWIF) and separates withdrawal from consumption.

The method is applied to GPT-3 training and inference across Microsoft US data-centre locations, and it projects global AI water withdrawal for 2027. Global AI demand is projected to account for 4.2-6.6 billion m3 of water withdrawal in 2027. Training GPT-3 in Microsoft's US data centres can consume 5.4 million litres of water in total, including 700,000 litres of scope-1 on-site consumption. GPT-3 'drinks' a 500 ml bottle for roughly 10-50 medium-length responses.

With good water quality, roughly 80% of cooling-tower withdrawal is evaporated (consumed). Data centres evaporate about 1-9 L per kWh of server energy, from 1 L/kWh (Google global annualised) to 9 L/kWh (a large Arizona facility in summer). Outside-air cooling with evaporative assist needs water when outside air exceeds about 85 F or is too dry, and about 70% of that added water is consumed, per Meta's report.

The US average electricity water intensity factor (EWIF) used is 3.14 L/kWh; in Table 1, total water per kWh ranges from 4.814 L (Iowa) to 9.629 L (Arizona). The spatial and temporal variation in WUE and EWIF makes 'when and where' a lever for water-aware scheduling and siting.

Key points

1. Global AI demand is projected to account for 4.2-6.6 billion m3 of water withdrawal in 2027.
2. Training GPT-3 in Microsoft's US data centres can consume 5.4 million litres of water in total, including 700,000 litres of scope-1 on-site consumption.
3. GPT-3 'drinks' a 500 ml bottle for roughly 10-50 medium-length responses.
4. With good water quality roughly 80% of cooling-tower withdrawal is evaporated; data centres evaporate about 1-9 L per kWh of server energy.
5. Outside-air cooling with evaporative assist needs water above about 85 F or when air is too dry; about 70% of the added water is consumed (per Meta).
6. US average EWIF is 3.14 L/kWh; in Table 1 total water per kWh ranges from 4.814 L (Iowa) to 9.629 L (Arizona).
7. Water should be reported alongside carbon.

Using it in design

• Separate withdrawal from consumption in water outputs, and default cooling-tower consumption to about 80% of withdrawal when water quality is good.
• Use location-specific WUE and EWIF, and show how a site's monthly WUE varies through the year.
• Support water-aware load placement and scheduling as an operations scenario for AI training halls.

Who should read it

Designers, sustainability analysts and researchers calculating the water footprint of AI data centers.

Related standards

• ISO/IEC 30134-9 (WUE)

Publication details

Authors

Pengfei Li; Jianyi Yang; Mohammad A. Islam; Shaolei Ren

Publisher

Communications of the ACM (Association for Computing Machinery (ACM))

Year

2025

Document number

arXiv:2304.03271

Volume

68

Issue

7

Pages

54-61

Document language

English

Source type

Scientific

Access

Free to read

DOI

10.1145/3724499

Licence

arXiv distribution permission — reuse conditions remain with the rights holder

Citation