How much water does AI consume?

Episode Summary

Executive Summary: The episode investigates claims that AI could consume trillions of litres of water by 2027 and finds the headline figure is wrong: it confuses water withdrawal with consumption, and the underlying academic estimate is itself based on a misread electricity forecast. The discussion then pivots to a more defensible estimate that AI systems were consuming about 750 billion litres annually by end-2025, while stressing uncertainty about local impacts and the importance of both direct and indirect water use.

Main Topics: AI water use and public concern (Priority: 5/5): The episode explains how AI depends on data centres, electricity, and cooling, all of which can require significant water use, raising fears about future scarcity as AI expands. The viral trillion-litre claim is incorrect (Priority: 5/5): Charlotte and Nathan show that Karen Hao’s cited 1.1-1.7 trillion gallons / 4.2-6.6 trillion litres figure was mislabeled: it referred to water withdrawal, not consumption. The academic source was flawed too (Priority: 5/5): The paper behind the claim used an electricity estimate that was misinterpreted; it treated projected 2027 AI server production as total global AI electricity use in 2027, understating the scale. A more careful estimate of AI water consumption (Priority: 4/5): Alex de Vries-Gao describes a supply-chain-based method to estimate AI hardware, electricity demand, and both direct and indirect water consumption, yielding a 2025 annual rate of 750 billion litres. Consumption vs withdrawal as competing metrics (Priority: 4/5): The episode notes that some researchers prioritize consumption for water scarcity, while others argue withdrawal also matters because large temporary demand can still stress local systems. Uncertainty about local and future impacts (Priority: 4/5): The discussion emphasizes that even large global totals are hard to translate into real-world harm without knowing where data centres and power plants are located and whether power and water infrastructure can support growth.

Key Arguments: The headline AI water figure was wrong because it confused water withdrawal with water consumption; consumption is only a subset of withdrawal. The academic paper cited by Karen Hao also relied on an electricity estimate that did not represent total global AI electricity use in 2027, making the downstream water estimate unreliable. A better estimate requires tracing AI hardware production through chips, server modules, and server deployment, then combining that with utilization and water-intensity assumptions. AI’s water footprint is split between direct on-site data-centre cooling and indirect off-site power generation, with most of the estimated consumption occurring off-site. A global figure can be large without proving local harm everywhere; the real risk depends on concentration in water-scarce regions. Withdrawal may matter as much as consumption because even returned water can create stress on local water systems and infrastructure. Future AI growth may be constrained by power availability and the ability to physically deploy all the equipment being produced.

Data Points: Claimed AI water use by 2027: 1.1 trillion to 1.7 trillion gallons per year - Karen Hao’s cited figure from Empire of AI, later shown to be mislabeled as consumption instead of withdrawal Claimed AI water use by 2027 in litres: 4.2 to 6.6 trillion litres per year - Conversion of the above figure; originally presented as fresh water consumption Corrected consumption figure from the paper: 380 to 600 billion litres per year - The paper’s actual estimate for water consumption, not withdrawal Alex de Vries-Gao estimate for end of 2025: 750 billion litres per year - Estimated annual water consumption rate for AI systems using a supply-chain-based method Global bottled water consumption: 446 billion litres per year - Used by Alex de Vries-Gao as a comparison to show the AI estimate is very large Direct on-site water share: About 10% - Portion of Alex de Vries-Gao’s estimated AI water consumption occurring at data centres Indirect off-site water share: About 90% - Portion of estimated AI water consumption occurring at power stations Original paper’s implied figure source: Global AI electricity use in 2027 - The paper extrapolated water use from an electricity estimate that was later shown to be misread Time from racetrack crash to billionaire status: 15 years - Opening teaser about Toto Wolff’s rise from a 2009 crash to billionaire team boss Time from engine start to ambulance: 15 minutes - Opening teaser describing the crash sequence in the Toto Wolff promo

Pivotal Quotes: "it’s time to put my balls on the dashboard" — Toto Wolff: Opening teaser from the promotional segment about Wolff’s 2009 racetrack crash "This is absolutely a big number." — Alex de Vries-Gao: His reaction to the estimate that AI systems were consuming 750 billion litres of water per year "we just don’t know at this time" — Alex de Vries-Gao: He explains why the global water estimate cannot be easily translated into local harm

Implications: The episode warns listeners to scrutinize AI environmental claims carefully: headline numbers can be wrong, but the underlying water demand is still substantial. For industry, the key issue is where AI growth lands and whether power and water infrastructure can support it.