The ChatGPT Water Usage Myth: Why AI's Environmental Footprint Is Far More Complex Than One Bottle

The Viral Statistic That Oversimplified AI Water Consumption

A single, memorable figure has been circulating across tech forums, sustainability reports, and regulatory discussions for the past couple of years: that ChatGPT consumes the equivalent of a 500ml bottle of water for every 10 to 50 responses it generates. It is the kind of concrete, tangible comparison that sticks — easy to visualise, easy to share, and easy to cite in policy arguments about the environmental cost of generative AI. There is just one significant problem: it was never a universal measurement. It was a conditional estimate derived from one modelled deployment scenario for GPT-3, and treating it as a fixed property of every AI interaction fundamentally distorts the conversation around AI water consumption.

For developers choosing cloud infrastructure, IT decision-makers evaluating AI toolchains, and policy professionals shaping digital regulation, this distinction matters enormously. A figure borrowed from a 2023 academic preprint and stripped of its qualifications is not a reliable basis for procurement decisions, infrastructure planning, or regulatory frameworks. Understanding where the number came from — and what it cannot tell us — is the starting point for more rigorous thinking about AI's true resource footprint.

Where the 500ml Estimate Actually Came From

The claim traces back to research by Pengfei Li, Jianyi Yang, Mohammad A. Islam, and Shaolei Ren, first released as a preprint in 2023 and later published in Communications of the ACM. The researchers aimed to estimate both the direct water used for cooling data centres and the indirect water embedded in the electricity those facilities consume. For GPT-3 training in Microsoft's US data centres, they estimated direct freshwater consumption of approximately 700,000 litres — a significant number in its own right.

The now-famous inference — that GPT-3 or ChatGPT could consume a 500ml bottle of water for roughly 10 to 50 medium-length responses — was a scenario-based estimate, not a metered measurement. It depended explicitly on where and when the system was deployed, the cooling infrastructure in use, the prevailing electricity grid's water intensity, and the hardware generation running the workload. A different data centre in a different region on a different grid using different cooling technology would produce a different number. As Shaolei Ren noted in subsequent commentary on the research, "We are not trying to give an exact number — the goal is to make the hidden resource cost of AI visible."

The research itself is sound. The problem arises when a conditional estimate with an explicit range gets laundered through repetition into an immutable fact — one that policy documents and sustainability pledges then treat as if it describes every ChatGPT request everywhere, regardless of infrastructure.

Direct vs. Indirect Water Use: The Split That Most Coverage Gets Wrong

Any serious analysis of AI water consumption needs to separate two fundamentally different categories of water use. Conflating them is one of the most common sources of confusion in both media coverage and corporate disclosures.

Direct water use occurs at the data centre facility itself. Many cooling systems — particularly evaporative cooling towers — remove heat from servers by evaporating water. Operators typically track this through Water Usage Effectiveness (WUE), a metric that measures the volume of water consumed per kilowatt-hour of energy used by computing equipment. A facility with a low WUE is more water-efficient in its direct operations.

Indirect water use occurs upstream, primarily at power stations that consume water to generate the electricity delivered to the data centre. A facility that uses zero water for on-site cooling can still carry a substantial water footprint through its grid supply — particularly if it draws from coal or nuclear generation, both of which are water-intensive processes.

A further distinction — withdrawal versus consumption — adds another layer of complexity. Withdrawal refers to total water drawn from a river, reservoir, or municipal supply. Consumption refers only to the portion not returned to the same water system, typically because it has evaporated. Articles and reports that slide between these two measures without signalling the switch make comparisons appear cleaner and more consistent than they actually are.

A 2025 review led by Lawrence Berkeley National Laboratory researcher Nuoa Lei confirmed that workload-level water use depends on a web of linked variables: location, cooling technology, operational WUE, server efficiency, grid water intensity, and utilisation rates. This is precisely why a per-prompt figure cannot be transplanted intact from one facility to another.

Why National Data Centre Totals Cannot Be Assigned to AI Alone

The aggregate scale of data centre water demand is substantial and growing. The 2024 United States Data Center Energy Usage Report estimated that US data centres directly consumed approximately 66 billion litres of water in 2023, up from 21.2 billion litres in 2014. Hyperscale and colocation facilities — the type that run AI workloads, but also cloud storage, video streaming, enterprise software, and scientific computing — accounted for 84 per cent of that 2023 direct total. The same report estimated roughly 800 billion litres of indirect water consumption through electricity generation in 2023.

These are significant numbers. But they belong to the entire data centre ecosystem, not to AI workloads specifically. Dividing a national total by an estimated count of AI prompts would manufacture false precision. Reaching a defensible per-prompt number would require knowing which workloads ran, on what hardware generations, at what utilisation rates, in which specific facilities, and against which electricity mixes — data that is largely not publicly available, according to reporting by Wired and analysis from the Lawrence Berkeley National Laboratory.

This gap between aggregate disclosure and workload-level accountability is particularly relevant for IT decision-makers and procurement teams. When a vendor's sustainability report cites a corporate-level water figure, it provides no actionable signal about the footprint of a specific AI service being evaluated.

What Corporate Disclosures Actually Tell Us — and What They Do Not

Technology companies have expanded their environmental reporting in recent years, and some disclosures contain genuinely useful data. The challenge is that they operate at the wrong level of granularity for most practical purposes.

Google's 2025 environmental report states that it replenished 4.5 billion gallons of water in 2024, equivalent to 64 per cent of its freshwater consumption. That gives a view of Google's overall water stewardship programme, but it does not attribute any share of water use to Gemini, Search, YouTube, or any individual service or request type.

Microsoft reported that its average data centre WUE fell from 2.3 litres per kilowatt-hour in the early 2000s to 0.27 litres in 2025 — a meaningful efficiency gain. The company has also stated that its new AI data-centre design uses zero water for cooling during operations through a closed-loop system. That is a specific and verifiable engineering claim, but it requires careful interpretation. "Zero water for cooling" does not mean zero water footprint: it excludes water embedded in electricity generation, semiconductor manufacturing, and construction. It also describes a new facility design, not every existing Microsoft data centre currently serving AI inference requests.

For developers and enterprises integrating AI tools into their infrastructure, these disclosures are a starting point, not a conclusion. The Communications of the ACM paper that originated the bottle estimate itself called for more granular, workload-level reporting — a recommendation that remains largely unmet across the industry.

Why Location Is Not a Side Note — It Is the Core Variable

One of the most significant factors that aggregate AI water consumption figures obscure is geography. A litre of water evaporated at a data centre in a water-stressed region during a dry season carries a fundamentally different environmental impact than a litre consumed at a facility situated where water is abundant and rainfall is consistent. The source of the water matters too: potable municipal supply, reclaimed wastewater, and seawater place very different pressures on local communities and ecosystems.

A 2026 UC Berkeley Law report on California data-centre water use found that existing corporate and research reports rarely provide the locally relevant detail required to assess water sources and their community impacts. This is a particular concern as hyperscale AI infrastructure expands into regions that may face increasing water stress under changing climate conditions — a dynamic that European digital sovereignty discussions have begun to incorporate as data localisation requirements intersect with sustainability obligations under frameworks like the EU Data Act.

For IT decision-makers and policy professionals evaluating AI cloud infrastructure, this means that a vendor's

Originally reported by Silicon Canals. Summarised and curated by European Purpose.