You’re Being Told Not to Use a Hosepipe. Yet AI Is Drinking Millions of Litres.

Britain is hot. Properly hot. I’m sitting in a cafe writing this, and despite their air conditioning, I’m still sweating.

On 24 June 2026, the UK provisionally recorded its hottest June day, with temperatures reaching 36.1°C in Hampshire. South East Water has introduced a temporary hosepipe ban in Kent after demand rose sharply during the heatwave, while water companies elsewhere are asking customers to restrict non-essential use. [see footnote 1]

So, while we are being encouraged not to water the lawn, wash the car or fill the paddling pool, it is worth looking at another rapidly growing user of water.

Artificial intelligence.

Not metaphorically. Not because somebody asked ChatGPT to write a particularly long email.

AI uses real water, in real places, from real water systems.

Why does AI need water?

Artificial intelligence does not live in a cloud.

It runs inside enormous data centres filled with servers containing powerful processors. Training an AI model, searching a company’s document estate or generating an answer requires those processors to perform huge numbers of calculations.

Calculations produce heat.

Too much heat damages equipment, reduces performance and can shut systems down.

Data centres therefore use cooling systems to move that heat away from the servers.

Some facilities use air cooling. Others circulate water through closed-loop systems. Many use evaporative cooling, where water absorbs heat and is then released into the atmosphere as vapour.

There is also an indirect water cost. Electricity generation itself can consume water, particularly where thermal power stations use it for cooling.

This means AI’s water footprint has two parts:

1. Water used at the data centre to cool the equipment, and

2. Water used elsewhere to generate the electricity powering it.

How much water are we talking about?

One widely cited academic study estimated that training GPT-3 in Microsoft’s US data centres could have directly evaporated approximately 700,000 litres of clean freshwater. [see footnote 2]

To put that into terms that feel more tangible, 700,000 litres is approximately:

• 8,750 average 80-litre baths

• 2.8 million glasses of water

• Enough to cover the daily domestic water use of around 4,600 people, assuming approximately 150 litres per person

And that was an estimate for training one generation of one model (ie., GPT-3 for example, not GPT-4, GPT-5, Gemini, Claude, Llama, or any of their later versions and variants).

Training, however, is only part of the story. Once a model has been created, millions of people and businesses begin using it. Every question, document summary, image, search and regenerated answer requires further processing.

The same researchers estimated that, depending on where and when the computation takes place, a conversation involving roughly 20 to 50 AI prompts could consume about 500 millilitres of water.[2]

That does not mean every conversation literally empties a bottle of water into a server. It is an averaged estimate incorporating data-centre cooling and the water associated with electricity generation.

It also varies considerably. A prompt processed in a cool location, at a water-efficient facility, using low-water electricity may have a much smaller footprint than the same prompt handled by a less efficient facility during hot weather.

But averages become significant at scale.

A single half-litre bottle does not sound alarming. Multiply it by hundreds of millions of conversations and it begins to look rather different.

At 500 millilitres per conversation:

• One million conversations would represent approximately 500,000 litres

• One billion conversations would represent approximately 500 million litres

• That is the equivalent of more than 6.2 million 80-litre baths

These are illustrative calculations based on the study’s estimate, not measurements of every AI system. But they demonstrate the basic problem: tiny amounts of resource use become enormous when multiplied across global platforms.

AI’s potential thirst is measured in trillions of litres

The researchers projected that global demand associated with AI could account for 4.2 to 6.6 billion cubic metres of water withdrawal in 2027. [see footnote 2]

That is 4.2 to 6.6 trillion litres.

Their comparison was equally striking: an amount greater than the annual water withdrawal of four to six countries the size of Denmark, and potentially equivalent to around half of the United Kingdom’s total annual water withdrawal.

Water withdrawal is not the same as water consumption. Withdrawn water may be returned to the environment, whereas consumed water is no longer immediately available locally because it has evaporated or been incorporated into another process.

The distinction matters.

But both figures matter to communities when reservoirs are low, treatment facilities are stretched or water is being taken faster than it can be replenished.

The energy number is just as uncomfortable

Water is only one part of AI’s environmental footprint.

The International Energy Agency reported that data centres consumed around 415 terawatt-hours of electricity globally in 2024, approximately 1.5% of worldwide electricity use. [see footnote 3]

By 2030, it expects data-centre electricity demand to more than double to around 945 terawatt-hours - slightly more than Japan currently consumes in an entire year. [see footnote 3]

A typical AI-focused data centre can consume as much electricity as 100,000 households. The largest facilities under construction could consume twenty times that amount. [see footnote 3]

Electricity and water are connected.

More computation means more electricity. More electricity produces more heat. More heat requires more cooling. Depending on the location and infrastructure, that can mean more water.

During a heatwave, the problem becomes particularly awkward: demand for AI services does not fall just because local water supplies are under pressure. In fact, cooling can become more difficult precisely when water is most constrained.

Not every data centre is equally thirsty

It would be wrong to suggest that every AI facility consumes water in the same way.

Technology companies are redesigning data centres to reduce their dependence on evaporative cooling. Microsoft says more than 90% of its data-centre capacity now uses closed-loop liquid cooling, where water is introduced during construction and continually recirculated without routine evaporation losses. [see footnote 4]

Its newer AI-focused data-centre design is intended to use no water for cooling during normal operation, avoiding an estimated 125 million litres of water per facility each year. [see footnote 5]

That is equivalent to approximately:

• 1.56 million baths

• Around 833,000 people’s average daily domestic water use

• Fifty Olympic-sized swimming pools

These improvements are important. They demonstrate that AI does not inevitably have to consume ever-increasing volumes of freshwater.

But more efficient infrastructure does not automatically solve the problem if demand grows faster than efficiency improves.

A car that uses 20% less fuel does not reduce total fuel consumption if we begin driving it ten times as far.

The waste nobody is talking about

Most discussion focuses on making chips, servers and cooling systems more efficient.

Far less attention is paid to the work we are asking those systems to perform.

Businesses are connecting AI to millions of documents containing duplicates, contradictions, obsolete policies, broken links, poorly formatted files and multiple versions of the same information.

The AI then searches all of it.

It indexes duplicated content. It embeds obsolete documents. It retrieves conflicting answers. It processes unnecessarily large context windows. It produces an unreliable response, so the user rephrases the question, tries again or asks another system.

Then somebody checks the answer manually because they don’t trust it.

The same task may be processed two, three or ten times - not because the AI is incapable, but because the information underneath it is disorganised.

Each unnecessary search, retrieval, embedding, prompt and regenerated response consumes computing power.

1. Computing power consumes electricity.

2. Electricity produces heat.

3. And heat must be removed.

4. Poor-quality information therefore has a physical cost. It is not simply a knowledge-management inconvenience or an AI accuracy issue. It creates avoidable demand for processing, power and cooling.

Better data means less waste

This is where the conversation needs to change.

Sustainable AI is not only about building greener data centres. It is also about reducing the amount of unnecessary work taking place inside them.

Before organisations add another chatbot, copilot or large language model, they should understand the condition of the knowledge being supplied to it.

• How much is duplicated?

• How much is obsolete?

• How many policies contradict one another?

• Which documents are unreadable by machines?

• How much irrelevant information is being repeatedly indexed, stored and searched?

• How often are employees regenerating answers because the first result was incomplete or wrong?

This is precisely the problem index addresses.

index scans enterprise knowledge environments to identify outdated, duplicated, contradictory, poorly structured and unreliable content. It helps organisations remediate those problems and maintain a healthier information estate over time.

The immediate benefits are better answers, improved governance, greater trust and more successful AI deployments.

But there is a wider consequence.

Cleaner knowledge means fewer irrelevant documents to process. Fewer duplicates to embed. Smaller and more accurate retrieval sets. Fewer failed searches. Fewer regenerated answers. Less human checking and repetition.

In practical terms: less unnecessary computation, less wasted electricity, less heat and potentially less water used to remove it.

index can’t solve the global water demands of artificial intelligence on its own (we wish!). No knowledge-management platform can.

But organisations can stop wasting resources asking expensive infrastructure to repeatedly process information that should have been corrected before the AI ever saw it.

When households are being told to put down the hosepipe, that feels like a responsibility worth taking seriously.

The future of AI will not only be judged by what it can do.

It will also be judged by how intelligently we choose to use it.

Better data. Better AI. Less waste.

Find out more at index-ai.net or contact contact@index-ai.net.

Sources and verification

[1] On 25 June 2026, South East Water’s Kent hosepipe ban was reported alongside a Met Office red extreme-heat warning. The provisional UK temperature of 36.1°C on 24 June exceeded the previous June record of 35.6°C. (ITVX)

[2] The 700,000-litre GPT-3 training estimate, the 500ml-per-20-to-50-prompts estimate and the 2027 global water-withdrawal projection come from Making AI Less “Thirsty”, by Li, Yang, Islam and Ren. These are modelled estimates rather than direct measurements supplied by OpenAI or Microsoft, so the article deliberately labels them as estimates. (arXiv)

[3] The International Energy Agency reports that data centres used around 415 TWh in 2024 and projects approximately 945 TWh by 2030. It also compares a typical AI-focused data centre with the electricity demand of 100,000 households. (IEA)

[4] Microsoft states that more than 90% of its data-centre capacity uses closed-loop liquid cooling, requiring water initially but recirculating it without evaporation losses. (Microsoft)

[5] Microsoft says its newer zero-water cooling design can avoid an estimated 125,000 cubic metres—125 million litres—of water use per facility annually. (blogs.microsoft.com)

by Paul Tucker