Growing appetite for artificial intelligence tools is placing documented strain on worldwide reserves of water and power. Large-scale data centre construction by cloud infrastructure providers is pushing consumption forecasts for constrained resources well beyond the adjustment capacity of regulatory frameworks.

Power and water consumption at data centres built to support AI workloads has been climbing on a consistent trajectory, with no plateau yet visible in the data. Data centres at the upper end of the scale draw on roughly 5 million gallons of water throughout a single day, the Environmental and Energy Study Institute has noted. Concrete figures remain elusive, given that the technology industry discloses relatively little about its internal workings.

Recent research points to those tallies falling short of the actual water consumption tied to such facilities. A study out of the University of Michigan argues that calculations of water footprints ought to factor in the water consumed by the energy sources powering those data centres. Researchers at the University of California, Riverside found that factoring in offsite water consumption pushes the combined water footprint of three planned Michigan AI data centres up by 14 billion gallons. Shaolei Ren, a professor there, noted that this form of indirect water draw tends to exceed direct on-site usage yet receives far less scrutiny, despite its consequences for regional watersheds and water supplies.

Petroleum industry wastewater has been identified by University of Texas at Austin researchers as a potential resource for meeting data centres' water needs without depleting freshwater supplies. A team at the university assessed the energy demands of several established desalination methods applied to wastewater, framing the analysis around their potential role in cooling data centres and, by extension, easing the technology sector's draw on limited freshwater supplies. The findings were published in the journal Water Research.

Reverse osmosis emerged from the research as the approach delivering the greatest efficiency, according to the team's findings. Wastewater streams with salinity levels above that of seawater fall outside what the technology handles reliably, a gap that may limit its usefulness for certain treatment applications. Distillation — heating water until it vaporises and then capturing the cooled condensate — remains a viable treatment route in settings where reverse osmosis cannot be installed, though its energy demands run notably higher than those of membrane-based filtration.

Bahadur noted that even the most energy-demanding approach would account for only a modest slice of a data centre's overall power budget, though some added expenditure remains unavoidable. Recycled water from petroleum and gas extraction sites has the potential to reduce how much freshwater data centres draw on, while also offering the drilling sector a practical route for handling a disposal problem it has contended with for years.

Water consumption tied to oil and gas operations outpaces that of the artificial intelligence industry — both in volume drawn and in the amount ultimately lost or wasted. Bahadur's figures put daily wastewater output from the Permian Basin at around 790 million gallons — a volume exceeding what a large data centre uses in a year by a factor of roughly 150 — and, as Bahadur noted, that flow currently travels considerable distances before being pumped into the ground rather than being put to further use.

Bahadur noted, in remarks carried by a news outlet, that the approach raises multiple concerns — among them the prospect of seismic activity — and that available capacity for storing the water underground is projected to be exhausted by around 2030. Bahadur identified putting the resource to work through data-centre purification as a practical path forward.