AI and data centers

Artificial intelligence systems rely on vast data centers filled with energy-intensive servers that generate substantial heat during the processing of prompts and generation of responses. 

Cooling these facilities often involves water-based systems that evaporate significant volumes to dissipate heat, while additional water is consumed upstream in the electricity generation process powering the centers. 

Early research focused on models like GPT-3 estimated that producing roughly ten to fifty medium-length responses could require the equivalent of a standard 500-milliliter bottle of freshwater, encompassing both on-site cooling and indirect water use tied to power production. 

This calculation was later extrapolated in media coverage to a single 100-word output from more advanced systems such as GPT-4, yielding figures near half a liter depending on the data center location, local climate, grid efficiency, and cooling technology employed. 

Variations arise because water intensity differs sharply by region—drier areas or those relying on thermal power plants demand more—while factors like server utilization and model size further influence totals. 

Although individual prompts appear modest, the cumulative effect across billions of daily queries worldwide raises concerns about strain on local water resources, especially in drought-prone regions hosting hyper-scale facilities. 

Efficiency gains in newer hardware, improved cooling methods, and optimized models have reduced per-query estimates substantially in recent years, yet the broader environmental footprint of expanding AI infrastructure continues to warrant careful scrutiny and sustainable design practices.

Source: Li, P., Yang, J., Islam, M. A., & Ren, S. Making AI Less “Thirsty”: Uncovering and Addressing the Secret Water Footprint of AI Models. arXiv:2304.03271 (University of California, Riverside).