📋 Research Report | Compiled February 2026 | Sources: Lawrence Berkeley National Laboratory · IEA · ScienceDirect · Brookings · MSCI · CEEW · WEF · Lincoln Institute


AI Data Centers & The Freshwater Crisis

Consumption volumes · Cooling physics · Corporate disclosures · Regional vulnerabilities · Mitigation pathways · Investment risk


Key Statistics

Metric Value Source
US direct cooling water use (2023) 17B gal / 64B L LBNL 2024
US indirect water use via power plants (2023) 211B gal / 800B L LBNL 2024
Big Tech water growth 2020→2023 +61% Surfshark / Corp reports
Global AI system water footprint (2025 est.) 313–765B liters ScienceDirect Dec 2025
Mid-century multiplier without mitigation 7× current ScienceDirect Apr 2025

Executive Summary

AI data centers have vaulted freshwater consumption from a peripheral infrastructure footnote to a first-order geopolitical and ecological variable. This report synthesises peer-reviewed science, hyperscaler sustainability disclosures, policy documents, and investment-risk analyses to map the problem across three scales — the physics of cooling, the geographies of scarcity, and the economics of mitigation — and to chart credible reduction pathways.

The central finding is probabilistic rather than binary: the trajectory of water consumption is highly malleable, with outcomes ranging from a seven-fold mid-century multiplication (business-as-usual evaporative cooling + fossil-electricity) to near-zero freshwater draws (closed-loop chip-level liquid cooling + renewables). The distance between those poles is a function of capital allocation, regulatory pressure, and siting discipline — all three currently trending in ambiguous directions.

"About two-thirds of data centers built since 2022 have been sited in water-stressed regions." — Bloomberg Intelligence / Lincoln Institute, October 2025

The India case is the most acute near-term crystallisation of the tension: $100B+ in committed hyperscaler investment converging on cities — Hyderabad, Pune, Bengaluru — that already face chronic water deficits, with regulatory frameworks that barely mention water.


1 · The Physics of Thirst: Why AI Cooling is Water-Intensive

1.1 Heat Generation Scales Super-Linearly with AI Density

Conventional data center racks drew 5–10 kW. By 2023 average rack density had reached 36 kW; NVIDIA's 2025 GTC roadmap projects the Rubin Ultra NVL576 rack at 600 kW by 2027. Every watt of compute becomes heat that must be expelled, and the cheap, scalable method for decades has been evaporative cooling — moving water-laden air through cooling towers where ~80% of withdrawn water is lost to atmosphere.

Water absorbs heat 3,000× more efficiently than air, making liquid-based cooling thermally superior — but legacy evaporative designs trade thermal efficiency for water expenditure, not conservation.

1.2 Three Accounting Scopes

Scope Description 2023 US Estimate
Scope 1 — Direct On-site chillers, cooling towers, humidification, direct-to-chip loops 17B gal (64B L)
Scope 2 — Indirect (power) Thermoelectric power plant steam cycles supplying the grid 211B gal (800B L) — ~12× Scope 1
Scope 3 — Embodied Semiconductor fab water (ultrapure water per chip), construction Not systematically tracked