Did ai use more water this year than every single water bottling company combined?
This fact-check may be outdated. Consider refreshing it to get the most current information.
Executive summary
A peer-reviewed study and multiple major outlets report that AI systems’ water footprint in 2025 is estimated between about 312 billion and 765 billion liters, a range that overlaps and, at its upper end, exceeds the best available estimate for global annual bottled‑water consumption (about 446 billion L), meaning the authors conclude AI may now be using more water than the bottled‑water industry worldwide [1] [2]. That conclusion is headline‑worthy but rests on substantial estimation choices, opaque corporate disclosures and different ways of counting “water use,” so the claim is plausible under the study’s methodology but not a settled, incontrovertible fact [1] [3] [4].
1. What the new research actually measured and why it matters
The Patterns paper that underpins much of the coverage estimated AI’s carbon and water footprints by combining company‑level environmental reports with modelled AI energy demand, producing a 2025 water‑use range of roughly 312.5–764.6 billion liters and noting that range overlaps the global bottled‑water figure (about 446 billion L) used for comparison [1]. The authors framed the finding as isolating “AI systems” rather than data centers generally, arguing that the rapid rise of large‑scale models and inference workloads justifies a focused estimate; that reframing matters because it treats AI as a distinct sectoral driver of resource demand rather than a diffuse component of cloud computing [1] [2].
2. Why many outlets ran with the “AI beats bottled water” headline
Major outlets — The Guardian, The Verge, Tom’s Hardware and others — amplified the paper’s comparison because it offers a vivid, accessible frame: bottled water is something readers buy and see, so equating invisible AI water flows to bottles sold makes an attention‑grabbing story [2] [5] [6]. Those headlines reflect the authors’ own language that AI’s water footprint “could be in the same range” as bottled‑water consumption and therefore can be interpreted, depending on which point in the estimate range one cites, as either comparable or larger [1] [2].
3. Key uncertainties and methodological caveats
The headline claim depends on multiple assumptions: how much of a datacenter’s energy is attributable specifically to AI workloads; whether water used for electricity generation is counted; whether water ‘withdrawn’ versus ‘consumed’ is the metric; and reliance on incomplete corporate disclosures and NDAs that hide location‑specific PUE/WUE values [1] [3] [4]. The Patterns paper is explicit that disclosure gaps and grid‑level variation make the estimate uncertain and that some company designs (Microsoft’s zero‑water AI datacenter design) could materially lower direct water use in some cases [1].
4. Alternative perspectives and implicit agendas
Experts and outlets urging restraint argue the study’s numbers are “conservative” in some ways but could still mislead if readers treat point estimates as precise; critics also warn about sensational framing that serves climate activism or anti‑tech narratives by prioritizing shock value over nuance [5] [7]. Meanwhile, tech firms face regulatory and reputational pressure to appear sustainable, so studies highlighting large footprints create leverage for calls to mandate better reporting — an outcome favored by researchers and some advocacy groups [2] [1].
5. Bottom line: did AI use more water than every bottled‑water company combined?
Under the study’s methodology and the comparison metric used by the authors — global bottled‑water consumption in liters per year — AI’s estimated 2025 water footprint can be equal to or higher than bottled‑water consumption, so the claim that AI “used more water this year than every single water bottling company combined” is supported by the paper’s upper‑range estimates and has been reported as such [1] [2]. That support, however, is conditional: it depends on the chosen accounting rules, incomplete corporate data, and whether one compares “consumed” vs. “withdrawn” water or includes water embedded in electricity generation — factors the authors and multiple analysts explicitly flag as limitations [1] [3] [4].