The claim that ChatGPT consumes a 500ml bottle of water for every 10 to 50 responses came from one modelled GPT-3 scenario, not a universal measurement; location, cooling and electricity supply can change the result

A striking number has followed generative AI around the internet: ChatGPT supposedly consumes a 500ml bottle of water for every 10 to 50 responses.

It is easy to see why the claim stuck. A bottle is tangible in a way that litres per kilowatt-hour are not. But the neatness of the comparison hides what the original research actually did.

The bottle figure was not read from a meter attached to an individual ChatGPT conversation. It came from a model of one GPT-3-era deployment scenario, built from assumptions about energy demand, data-centre cooling and the water used to generate electricity. Change the model, hardware, location, season, cooling system or power supply, and the answer changes too.

That does not make AI’s water demand imaginary. It means the honest answer is less memorable and more useful: there is no universal amount of water consumed by a prompt.

Where the bottle estimate 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 set out to estimate both the direct water used for cooling data centres and the indirect water associated with producing their electricity.

For GPT-3 training in Microsoft’s US data centres, they estimated direct freshwater consumption of about 700,000 litres. Their paper also offered the now-famous inference that GPT-3 or ChatGPT could consume a 500ml bottle of water for roughly 10 to 50 medium-length responses, depending on where and when the system was deployed.

Those qualifications matter. The number was a scenario estimate, not a permanent property of ChatGPT, and it was based on the infrastructure and assumptions available to the researchers. A short answer and a long answer do not require the same computation. Newer chips, denser workloads and different models do not necessarily behave like the GPT-3 system they modelled.

I do not think the research is the problem. The problem begins when an estimate with an explicit range and several dependencies gets repeated as if every chatbot request triggers a tiny, fixed water bill.

One footprint contains two different kinds of water

Data-centre water figures often combine concepts that should be kept separate.

Direct water use happens at the facility. Some cooling systems evaporate water to remove heat from servers. Operators often describe this through water usage effectiveness, or WUE, which measures the water used by a data centre relative to the energy consumed by its computing equipment.

Indirect water use occurs elsewhere, mainly when power stations consume water to generate the electricity that reaches the facility. A data centre that uses little or no water for on-site cooling can therefore still have a water footprint through its grid supply.

There is another distinction between withdrawal and consumption. Withdrawal is water taken from a river, reservoir, aquifer or municipal supply. Consumption is the share not returned to the same water system in a readily available form, often because it has evaporated. Articles that switch between these measures can make comparisons look cleaner than they are.

A 2025 review led by Lawrence Berkeley National Laboratory researcher Nuoa Lei found that workload water use depends on many linked factors, including location, cooling technology, operational WUE, server efficiency, grid water intensity and utilisation. That is why a single per-prompt number cannot travel intact from one facility to another.

The large national totals are not AI-only totals

The scale of the broader data-centre system is still substantial. The 2024 United States Data Center Energy Usage Report estimated that US data centres directly consumed about 66 billion litres of water in 2023, up from 21.2 billion litres in 2014. Hyperscale and colocation facilities accounted for 84 per cent of that 2023 direct total.

The report also estimated roughly 800 billion litres of indirect water consumption through electricity generation in 2023. Yet neither figure belongs to AI alone. The totals include many kinds of computing, from cloud storage and video streaming to enterprise software and scientific work.

Dividing those national numbers by a guessed count of AI prompts would produce false precision. Researchers would need to know which workloads ran, on what hardware, for how long, at what utilisation, in which facilities and against which electricity mix. Much of that information is not publicly available.

Company disclosures show improvement, but not a prompt-level answer

Technology companies now publish more water information, although it generally appears as an annual corporate or fleet-wide total.

Google says in its 2025 environmental report that it replenished 4.5 billion gallons of water in 2024, equivalent to 64 per cent of its freshwater consumption. That offers a view of Google’s overall operations and water programme, but it does not reveal how much water was attributable to Gemini, Search, YouTube or any individual request.

Microsoft said in June 2026 that its average data-centre WUE fell from 2.3 litres per kilowatt-hour in the early 2000s to 0.27 litres in 2025. The company also says its new AI data-centre design uses zero water for cooling during operations through a closed-loop system.

That is a specific and meaningful claim, but “zero water for cooling” is not the same as a zero-water AI service. It does not include water associated with electricity generation, semiconductor manufacturing or construction. It also describes a new design, not every facility serving every Microsoft AI request.

Location is part of the impact

A litre consumed in a water-stressed catchment during a hot, dry month is not equivalent to a litre consumed where water is plentiful. The source matters too. Potable municipal water, reclaimed wastewater and seawater place different pressures on communities and ecosystems.

This is the piece that global averages tend to erase. 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 needed to assess water sources and community impacts.

Better disclosure would connect the workload to its context. It would identify the model and approximate output length, hardware, energy demand, facility efficiency, direct WUE, grid water intensity, location, time of operation, water source, and whether a figure describes withdrawal or consumption.

The memorable number is not the useful one

The bottle comparison did something valuable: it made the physical infrastructure behind apparently weightless software visible.

But its value disappears when a conditional estimate becomes a universal fact. The most defensible conclusion is not that every 10 to 50 ChatGPT answers consume exactly 500ml of water. It is that AI services have a real water footprint, that the footprint can vary sharply, and that outsiders still lack enough workload-level and site-level data to calculate it cleanly.

One bottle is easy to picture. Accountability will require the less photogenic details printed underneath it.