Evidence Notes

Card 1: Terms, Boundaries, and Local Evidence

Four distinctions

Direct/indirect describes where a pathway occurs. Withdrawal/consumption describes what happens to water. A direct withdrawal can be partly returned and partly consumed.

Quantity is not quality

• Water quantity concerns the amount, source, timing, withdrawal, consumption, and return flow.
• Water quality concerns properties such as temperature, dissolved solids, treatment chemicals, and whether discharge meets permit limits and protects the receiving water.

An annual total does not reveal the season of use, the condition of returned water, or effects on a particular stream.

Professor Wen's water-scientist lens

Before judging a local proposal, ask for:

• cooling technology and operating schedule;
• source water and seasonal withdrawal and consumption volumes;
• electricity mix and where that electricity is generated;
• cooling-water blowdown chemistry and discharge temperature;
• discharge route, permit limits, and compliance monitoring;
• receiving-water flow, temperature, existing stress, habitat, and other users;
• any additional water infrastructure, who would pay for it, and whether special rates or discounts are proposed.

These questions are also emphasized in the AAAS EPI Center, GovAI Coalition, and National League of Cities (2026) community decision guide, which asks local officials to request the cooling system, peak and annual withdrawal and consumption, electricity-related water, source water, and infrastructure obligations. Open the official AAAS EPI Center resource page.

Guide limit: This public-facing practitioner guide was vetted for use by community decision-makers, but it is not a peer-reviewed research paper and does not count as one of the three scientific studies. It helps students ask for evidence; it does not establish a facility's impacts.

Evidence limit: The three paper cards can frame good questions. They cannot substitute for a facility design, permit application, meter record, discharge sample, or local ecological assessment.

Card 2: Xiao et al. — A Modeled U.S. AI Projection

Source: Xiao, T. et al. (2025). “Environmental impact and net-zero pathways for sustainable artificial intelligence servers in the USA.” Nature Sustainability. https://doi.org/10.1038/s41893-025-01681-y

What the study adds

The authors model environmental impacts of AI servers in the United States under different scenarios. Their estimated annual AI-server water footprint is 731–1,125 million cubic meters. In the modeled result, approximately 71% is indirect, associated with electricity generation, and 29% is direct, associated with the data-center site.

Claims this card supports

• In this U.S. projection, indirect water use is larger than direct water use.
• A water-footprint estimate changes when location, electricity supply, and facility choices change.
• Electricity-related water can be a major part of an AI system's total water footprint.

What this card does not establish

• It is not a set of meter readings from every U.S. AI data center.
• It does not imply that every prompt uses the same amount of water.
• It does not predict the fictional Central New York facility's exact water use.
• It does not demonstrate local water-quality or ecological harm.

Use the words modeled, projected, and in this U.S. study when citing the result.

Card 3: Lei et al. — Why Workload Estimates Vary

Source: Lei, N. et al. (2025). “The water use of data center workloads: A review and assessment of key determinants.” Resources, Conservation & Recycling, 219, 108310. https://doi.org/10.1016/j.resconrec.2025.108310. Official DOE/LBNL record and accepted manuscript.

What the study adds

The authors assess how data-center workload water use changes with computing efficiency, utilization, cooling, climate, location, and electricity supply. Across the conditions they examine, estimated workload-level water use varies by more than 10,000-fold. They connect this range to more than 1,000-fold variation in water use per unit of electricity and about 10-fold variation in workload efficiency.

Claims this card supports

• One universal “water per prompt” value hides major differences in workload, facility, grid, and location.
• Server efficiency alone cannot describe the full water footprint.
• A comparison is meaningful only when system boundaries and assumptions are visible.

What this card does not establish

• The article concerns data-center workloads broadly, not one named commercial AI service.
• The more-than-10,000-fold range is not a claim that one ordinary prompt literally changes by that factor everywhere.
• It does not supply local meter, permit, discharge, or ecosystem data for Central New York.

Card 4: Jager and Yoon — Screening Water and Ecosystem Risks

Source: Jager and Yoon (2026). “Data centers: an emerging threat to freshwater biodiversity in the United States.” Water Biology and Security, article 100585, corrected proof. https://doi.org/10.1016/j.watbs.2026.100585

What the study adds

The authors use a screening-level framework to identify pathways by which data centers could affect water resources and aquatic ecosystems. These pathways include water stress and altered flows, warm or lower-quality discharge, and risks to aquatic organisms associated with water intake and changing habitat conditions.

Claims this card supports

• Water impacts can involve quality and ecology as well as quantity.
• Site, watershed, and receiving-water conditions matter to risk.
• A local review should examine intake, discharge, season, water stress, and sensitive ecosystems.

What this card does not establish

• It does not demonstrate that the fictional facility has harmed drinking water or fish.
• Screening a potential pathway is not the same as measuring a site-specific impact.
• The article covers data centers broadly—including AI and other computing demand—not AI alone.
• A conclusion about local harm would require facility, permit, monitoring, and receiving-water evidence.