How much of a problem is AI's water use?
US data centers consumed an estimated 66 billion liters of water in 2023, representing less than 1% of national consumption, but this figure is projected to rise sharply with AI expansion AI model efficiency has improved dramatically: the widely cited "500 ml per query" estimate from 2024 is now outdated, with Google reporting only five drops per median Gemini query in 2025 Strategic siting of data centers in water-rich, renewable-energy-abundant regions (e.g., Montana, Nebraska) instead of drou
Analysis
TL;DR
- US data centers consumed an estimated 66 billion liters of water in 2023, representing less than 1% of national consumption, but this figure is projected to rise sharply with AI expansion
- AI model efficiency has improved dramatically: the widely cited "500 ml per query" estimate from 2024 is now outdated, with Google reporting only five drops per median Gemini query in 2025
- Strategic siting of data centers in water-rich, renewable-energy-abundant regions (e.g., Montana, Nebraska) instead of drought-stricken areas could reduce AI's future water footprint by up to 86%
- Liquid cooling technology in new AI data centers operates as a closed-loop system, eliminating evaporative water loss compared to traditional evaporative cooling towers
- By 2030, US data centers could consume 731–1,125 billion liters annually, with the high-end estimate equivalent to New York City's entire annual drinking water supply
Why It Matters
This article directly addresses an emerging tension between AI infrastructure growth and environmental sustainability, making it critical for AI practitioners and policymakers who must balance computational expansion with resource responsibility. The data on water consumption and the available mitigation strategies provide actionable guidance for companies building AI infrastructure, while the local protest movements in the Southwest signal growing community opposition that could delay or block projects if unaddressed.
Technical Details
- Cooling technologies: Traditional pre-AI data centers use evaporative cooling towers that lose water to the atmosphere, while modern AI data centers from Amazon, Microsoft, and Google increasingly deploy closed-loop liquid cooling systems that circulate fluid directly over processors with minimal to no water loss
- Projected consumption: Fengqi You's group at Cornell estimated 731–1,125 billion liters annually by 2030; the 2025 Xiao et al. study in Nature Sustainability modeled low, mid, and high-demand scenarios for US water footprint from 2024–2030
- Efficiency gains: AI model efficiency improvements have drastically reduced per-query water estimates—from 500 ml per email (GPT-4 era, 2024) to approximately five drops per median Gemini query (2025)
- Mitigation levers: Shifting to renewable energy (solar/wind require negligible water vs. coal/gas steam-turbine cooling), strategic siting away from drought zones, and advanced cooling technologies could collectively reduce AI's water footprint by up to 86%
- Regional impact: Current Texas data centers consume less than 1% of state water demand, but projections suggest significant increases during peak summer months when supplemental misting cooling is activated
Industry Insight
- Companies expanding AI infrastructure should prioritize data center siting in regions with abundant renewable energy and low water stress (e.g., Montana, Nebraska, parts of Texas and South Dakota) to mitigate both regulatory risk and community opposition, potentially reducing water footprint by up to 86%
- The rapid improvement in model efficiency means per-query resource estimates become obsolete quickly; organizations should invest in next-generation liquid cooling and closed-loop systems rather than relying on outdated evaporative cooling, which will become increasingly untenable in drought-prone regions
- Water consumption narratives are highly polarized—ranging from "AI has no water problem" to viral claims of 500 ml per query—so practitioners should cite updated, company-specific data (e.g., Google's five-drop figure) and contextualize within broader industrial usage to engage constructively with stakeholders and policymakers
Disclaimer: The above content is generated by AI and is for reference only.