Water Accounting Emerges as a Key Issue for US AI Data Centers

TechNews has published an analysis of how water use at US AI data centers should be assessed, contrasting direct facility consumption with indirect water use associated with electricity generation.

Citing LBNL estimates, the article places direct US data-center water consumption at about 47 million gallons per day, or roughly 17 billion gallons annually. When water associated with electricity generation is included, the estimate rises to about 580 million gallons per day—more than 12 times the direct figure. These are different accounting boundaries rather than directly interchangeable measures of facility water withdrawal.

The article also references a USGS estimate of 133 billion gallons per day of water use by US thermal power generation in 2015, placing the electricity-related component of data-center water accounting within the wider power-sector context.

Cooling architecture is central to the issue as AI rack power densities reach 100 kW and above. The article characterizes conventional whole-room air conditioning as unsuitable at these densities and describes direct-to-chip liquid cooling systems that use coolant distribution units (CDUs) and separate closed loops for server coolant and facility water.

Under suitable ambient-temperature conditions, such systems can use dry cooling with virtually no external water consumption, according to the article. That outcome remains conditional on weather and system configuration rather than being a universal attribute of liquid cooling.

For operators, investors and site-selection teams, the analysis indicates that water risk cannot be evaluated through on-site consumption alone. Liquid cooling, CDUs and closed-loop designs could become increasingly important in water-constrained markets, while the reported water footprint will depend on whether the assessment includes operational water at the facility, electricity-system water use, or both.