California data centre water fears exaggerated, study finds
Professor Jay Lund argues media narratives on artificial intelligence water usage are largely speculative, with current estimates representing less than one per cent of total human consumption.
Concerns regarding the water consumption of artificial intelligence facilities in California are largely speculative, according to a new analysis by Professor Jay Lund of the University of California – Davis. The study challenges prevailing media narratives that suggest AI data centres pose a significant threat to the state's water supply, indicating that current usage levels are negligible when measured against total human consumption.
Lund's assessment, published in the California WaterBlog, indicates that annual water evaporation from existing data centres in the state ranges between 32,000 and 290,000 acre-feet. This volume represents only between 0.055 per cent and 0.7 per cent of California's total annual human water use, which stands at approximately 40 million acre-feet. The analysis notes that even the higher estimate is sufficient to irrigate merely 10,000 to 100,000 acres of farmland, a fraction of the state's 7 million acres of irrigated agriculture.
The calculations are based on the physical footprint of about 1.4 million square metres of data centre space and assumptions regarding cooling efficiency. While the author acknowledges a lack of explicit transparency from the AI industry regarding specific water usage data, he argues that too many journalists and advocates have wallowed in speculation arising from this silence. The study utilises fundamental physics to convert energy use into water use for cooling, noting that major industrial cooling systems typically operate with efficiencies between 60 and 90 per cent.
Lund contextualises the current alarmism by comparing it to historical fears surrounding new technologies, ranging from the illusory flying cars of the Jetsons to the beneficial introduction of vaccines and water chlorination. He observes that while data centres are described as factories of data and computation requiring substantial energy and water, their impact in California is modest. However, he cautions that water use may be more significant in other states with less developed water infrastructure.
In a self-deprecating comparison regarding the magnitude of the study's findings, Lund suggests that his own breathing while writing the post may have evaporated more water than the incremental impact of all four AI models used in the estimation. He emphasises that all water problems are local and that while the impact is negligible in California, the narrative often overlooks the specific infrastructure realities of different regions.
The analysis concludes that the broad range of estimates, from 2,300 to 400,000 acre-feet per year, supports a reasonable figure of 32,000 to 290,000 acre-feet. Despite the volume seeming large to an individual, it pales in comparison to the state's total consumption, suggesting that current AI water use is among the more economically effective uses of water in the region.
