Paper comparison
The three papers answer different questions. Carry each paper’s study type and limitation with any result you cite.
Supports: an estimated U.S. AI-server water-use range of 731–1,125 million m³ per year, with 71% indirect and 29% direct in the modeled total.
Does not support: treating the result as facility meter data or a universal per-prompt value.
Supports: more than 10,000-fold variation in estimated water use across data-center workloads, shaped by workload efficiency, cooling, electricity, climate, and utilization.
Does not support: assigning one water value to every AI request or commercial service.
Data centers: an emerging threat to freshwater biodiversity in the United States
Supports: screening potential pathways involving thermal effects, lower-quality discharge, water stress, entrainment, and altered flows.
Does not support: claiming demonstrated ecological or drinking-water harm at a particular facility.
| Paper | Evidence form | Scale | Best use in this module | Required qualifier |
|---|---|---|---|---|
| Xiao et al. | Projection/model | United States | Direct versus indirect AI-server water use | Modeled scenario, not meters |
| Lei et al. | Workload assessment | Workload and infrastructure combinations | Why estimates vary | Broad data-center workloads |
| Jager and Yoon | Risk screening | Data-center water and ecosystem pathways | What local quality evidence to request | Potential risk, not demonstrated harm |