A new Harvard analysis mapped where data-center water actually goes.
Listen to this article:
Recent social media posts have joked that using an LLM is a bit like turning on a hose, draining a bottle of water, or flushing the toilet. The idea being that every time an LLM answers a prompt, it heats up a microchip in a data center and evaporates water to cool down the facility. Water per prompt is an intuitive way to make AI’s resource demands tangible, but it also compresses a complicated infrastructure system. To more accurately understand AI’s water footprint, 2026 HBS AI Institute associate Francesca Dominici and her co-author Gianluca Guidi mapped operational water consumption across 472 U.S. hyperscale facilities, separating water used directly for cooling from water consumed through the electricity systems that power them. In “The Hidden Water Geography of U.S. Hyperscale Data Centers in the AI Era,” their analysis suggests that understanding AI’s growing water footprint requires looking beyond how much water it uses to examine how that water is consumed and where the burden ultimately falls.
Key Insight: Two Water Footprints
“Operational water consumption arises through two pathways.” [1]
The physical footprint of AI infrastructure begins with electricity. The 472 hyperscale facilities studied represent more than 20,000 megawatts of capacity and, under the researchers’ baseline scenario, approximately 116 terawatt-hours of annual electricity demand. But electricity demand creates two different connections to water. The first is direct cooling water, which the researchers call Scope 1. Data centers generate enormous amounts of heat, and many cooling systems consume water through evaporation. The second is easier to overlook: electricity-related water, or Scope 2. Generating electricity can itself consume water, such as when power plants use water for cooling or when reservoir losses are attributed to hydroelectric generation. To estimate both pathways, the researchers combine facility electricity demand with the standardized Water Usage Effectiveness (WUE) metric for cooling, the combination of energy sources supplying each regional electricity grid, technology-specific water-consumption factors, and geospatial data on water stress.
Key Insight: Most of the Water is Somewhere Else
“Scope 2 therefore contributes about 75% of the national total, roughly three times the Scope 1 contribution.” [2]
Under the study’s baseline assumptions, the 472 facilities consume approximately 300 billion liters of operational water per year. Roughly 74 billion liters come from direct cooling, while about 226 billion are attributed to generating the electricity those facilities consume, meaning that Scope 2 is far more dominant than Scope 1. The size of the estimate depends substantially on one methodological choice: how to attribute reservoir evaporation to hydropower. The researchers include that evaporation in their baseline calculation, but note that attributing it to electricity generation is debated. The baseline assigns hydropower a water-consumption factor of 8 liters per kilowatt-hour. When the researchers instead assign hydropower no reservoir-evaporation burden, estimated Scope 2 water falls by 43%, from about 226 to 128 billion liters annually (still 63% of the total). The authors caution that their analysis uses annual averages, does not capture seasonal fluctuations or facility-specific measured cooling performance, and excludes the water associated with manufacturing, construction, and supply-chain activities.
Key Insight: The Water Burden Moves When You Follow the Power
“The places where cooling-side changes would save the most water are not the places where electricity-side changes would save the most water.” [3]
Mapped side-by-side, the footprints barely overlap. On-site cooling pressures show up most strongly in parts of the West and South-central U.S., especially where water is already scarce. Once the researchers trace water through the power system, however, much of the burden shifts toward grid territories farther east. Electricity-related water is also much more concentrated. More than half of direct cooling water is spread across the 16 largest-contributing hydrologic basins, but just three electricity-grid regions account for 59% of estimated electricity-related water. In discussing attempts to reduce water consumption, the researchers note that cooling-side reductions depend on local decisions about cooling architecture, reclaimed or non-potable water, sourcing, and drought planning. Electricity-side reductions instead depend on energy generation mix, procurement, utility resource planning, and transmission.
Why This Matters
Every AI capability rests on a larger physical and operational system that may introduce dependencies the business does not directly see or control. As organizations scale AI, business leaders and executives should ask not only whether a technology performs well or delivers ROI, but what infrastructure it depends on, where those dependencies are concentrated, and how they could affect cost, resilience, reputation, or future growth. That makes infrastructure awareness part of AI strategy, and the companies best positioned to scale AI may be those that consider it from the beginning, rather than discovering it only after AI becomes critical to the business.
References
[1] Guidi, Gianluca, and Francesca Dominici, “The Hidden Water Geography of U.S. Hyperscale Data Centers in the AI Era,” arXiv preprint arXiv:2607.02531 (2026): 1. https://doi.org/10.48550/arXiv.2607.02531
[2] Guidi and Dominici, “The Hidden Water Geography of U.S. Hyperscale Data Centers in the AI Era,” 2.
[3] Guidi and Dominici, “The Hidden Water Geography of U.S. Hyperscale Data Centers in the AI Era,” 6.
Meet the Authors
Gianluca Guidi is a Postdoctoral Research Fellow at the Harvard T.H. Chan School of Public Health.
Francesca Dominici is Clarence James Gamble Professor of Biostatistics, Population, and Data Science at the Harvard T.H. Chan School of Public Health and a 2026 Associate at the HBS AI Institute.