TL;DR: AI data centre pollution makes for scary headlines, but most of them aim at the wrong thing. On-site cooling is only a small slice of the footprint, with the electricity behind the building and the chip factories taking the lion’s share.
Evaporated cooling water is not destroyed and the rising vapour is essentially clean, but the leftover water does pick up treatment chemicals and trace metals and has to be handled properly.
The real issue is where and when data centres are built. For your own business website, the green wins are a renewable host and a lean, fast build.
You have probably seen the headlines; AI is draining the rivers, every question you type into a chatbot pours a bottle of water down the drain, and data centres are quietly emptying reservoirs while the rest of us are told to take shorter showers. And some of that is true, however, a lot of it is conflated nonsense. And almost none of it is explained in a way that helps you make a better, greener choice for your own website.
This is a plain English guide to what data centres really do with water, which bits of the panic are fair, which bits are overcooked, and what any of it means when you are picking hosting for a small business website.
What actually uses the water in data centres
The first thing to clear up is that the picture in most people’s heads is wrong. When folk imagine an AI using water, they picture a server somewhere sipping from a tap to cool itself down. That happens, but it is the smallest piece of the puzzle.
A data centre’s water footprint really comes in three parts:
- On-site cooling. Water used at the building itself to stop the servers overheating. This is the part everyone pictures.
- Electricity generation. The power stations that supply the building use water too, often a lot of it, to generate the electricity. This happens miles away and rarely gets counted in the scary headlines.
- Making the chips. Manufacturing the processors in the first place is extremely water hungry. This is hidden away in the supply chain and almost never mentioned.
Here is the bit that should change how you read the next headline. A water demand analysis reported in June 2026 put on-site cooling at only around 4 percent of the extra water AI will need, with power generation taking roughly 54 percent and chip manufacturing around 42 percent. In other words, the thing everyone argues about is the tiny slice. The big slices are the ones nobody is shouting about.
The two ways data centres cool servers
Cooling is still worth understanding, because it is where most of the confusion lives, and because the difference between the two main methods explains why “data centres use water” is true in one breath and misleading in the next.
There are two basic approaches, and two everyday things they are just like.
A closed loop works like the radiator in your car. The same fluid goes round and round in a sealed circuit, picking up heat from the servers and dumping it into the air through a radiator. The fluid is reused over and over, so almost no water is actually used up. The catch is that it can take more electricity to shift the heat this way.
An open loop works like the water cycle you were taught at primary school. Water soaks up the heat, then some of it evaporates away to carry that heat off. It rises, forms clouds and falls again as rain, so it is not destroyed. The catch is that it rejoins the wider water cycle, not the site’s own tank, so fresh water has to be topped up constantly. That evaporation, and the top-up it forces, is exactly the bit that gets quoted as water being “consumed”. The diagram below follows the water all the way round, from the data centre to the cooling tower, up to the rain cloud, and back through the reservoir and treatment works.
So when someone says a data centre guzzles water, what they usually mean is an open loop, evaporative system on a hot day. A modern closed loop or air cooled facility might use very little on site at all. The cooling method matters far more than the fact that cooling exists.
Closed loop
Like a car radiator
The same fluid circulates in a sealed circuit. Heat is released to the air through the radiator, so almost no water is used up. The trade off is that it can use more electricity.
Open loop
Like the water cycle you learned at school
Instead of sealing the water in, an open loop lets some of it evaporate to carry the heat away. That water is not destroyed. It rises, forms clouds and falls again as rain. But it rejoins the wider water cycle, not the site's own tank, so fresh mains water has to be added constantly to top it back up.
That constant top-up is the bit quoted as water "consumption". The full journey is shown below.
The open loop, told as the whole water cycle
Follow the water round: it always comes back, just not to the same place
The water always returns as rain somewhere, which is why people get confused about whether it is really "used up". The catch is that it rarely returns to the same place, so a site in a dry region still has to keep drawing fresh water from the local supply. That is why where a data centre is built matters far more than the fact that it uses water at all.
QED Web Design, weareqed.com
The bottle of water per question claim
This is the stat that went everywhere, so it is worth getting right. The research it comes from, from the University of California, Riverside, did not say one question equals one bottle. It estimated that an older model like GPT-3 used roughly a 500ml bottle of water for somewhere between 10 and 50 medium length answers, and crucially that the figure swings wildly depending on where and when the work is done.
That “where and when” part is the whole game. The same task in a cool climate with an efficient building uses a fraction of the water it would use in a hot, dry region in the middle of summer. There is no single honest number, which is exactly why a single scary number should make you suspicious.
Withdrawal versus consumption
This is the single most useful idea in the whole debate, and it takes about a minute to learn. There are two different things being measured, and headlines love to mix them up.
| Term | What it means | Why it matters |
|---|---|---|
| Water withdrawal | Water taken from a river, lake or supply, used, and much of it then returned. | Often a scary number |
| Water consumption | Water that is gone for good from that supply, usually because it has evaporated away. | The figure that bites |
When a headline quotes an enormous “water use” figure, it is almost always quoting withdrawal, because withdrawal numbers are bigger and read more dramatically.
Consumption is the figure that actually tells you whether a local supply is under strain. Knowing which one you are looking at turns a frightening number into a meaningful one.
AI data centre pollution: is the cooling water full of poison?
One claim doing the rounds is that the real problem is not the volume of water but that data centre cooling water is laced with heavy metals and toxic chemicals. Like a lot of this debate, it is part true and part overcooked, so it is worth separating the two.
The true part. Water sitting in an open cooling system does not stay pure. To stop bacteria, algae and corrosion building up in warm, damp pipework, operators add treatment chemicals such as biocides like chlorine or bromine, and corrosion inhibitors like phosphates. Over time the circulating water also picks up trace metals such as copper and zinc, mostly leached from the system’s own pipes and fittings.
The concentrated leftover, known as blowdown, genuinely has to be dealt with, and where lots of facilities cluster together it can pile pressure on the local treatment works.
The overcooked part. The water that evaporates and rises to form clouds is essentially clean. The chemicals and metals stay behind in the liquid that does not evaporate, so the rain in the water cycle is not carrying poison back down.
And in the UK and most regulated places, that leftover blowdown cannot simply be tipped into a river or the drinking supply. It goes to the sewer and a treatment works, or it needs a discharge permit with monitored limits on metals, pH and dissolved solids. The metals are largely the system corroding itself, not a toxic brew added on purpose.
Notice where that lands. The treatment works is exactly the stage in the water cycle that handles this, which is why it sits in the diagram above. The honest worry is about capacity and location, not about secretly poisoned water.
So what is the real problem with AI data centre pollution?
None of this means the concern is invented. There is a genuine issue here, it is just a more specific one than the headlines suggest. The problem is not that AI uses an unimaginable amount of water everywhere. The problem is where and when it is being used.
A Guardian analysis in June 2026 found that around two thirds of the data centres planned across the United States are slated for areas that have recently been in drought. Building water hungry, evaporative cooled facilities in places that are already short of water, and running them hardest on the hottest days when supplies are tightest, is a real and fair thing to be worried about.
A United Nations University report around the same time reached a similar split on the footprint, so this is not a fringe view.
That is a much sharper and more defensible worry than “AI is drinking the rivers”. It points at siting, timing and planning, which are things that can actually be fixed, rather than at the existence of cooling, which cannot.
What this means for your business’ website
Here is where it lands for you, because every website, not just AI, runs on a physical building somewhere that uses power and, sometimes, water. The honest takeaway is not to panic about AI specifically. It is to remember that everything digital has a physical footprint, and the genuinely green move is to build lighter and choose better.
A few practical things that actually move the needle for a small business site:
- Choose a green host. Look for hosting powered by renewable electricity, since the power behind the building is the biggest part of the footprint, not the cooling.
- Build a lean, fast site. Bloated pages stuffed with huge images and unused code make every visit do more work, which means more energy. A tidy, fast site is a greener site, and it ranks better and converts better too (See our post on how to build a Wordpress website that ranks).
- Do not over-engineer. Most small business sites do not need heavy frameworks and endless plugins. Less code means less compute, which means a smaller footprint.
- Keep what you have working well. A well maintained site that loads quickly will always be greener than a sluggish one that has been left to rot.
In other words, the same things that make a website sustainable are the things that make it good. Fast, lean and well hosted is the whole recipe, and it happens to be kinder to the planet at the same time. If you want a hand building a website that is quick, lean and hosted on renewable energy, that is exactly the sort of work we do at QED, so feel free to get in touch for a friendly chat about your site.
Common questions
Does every AI question really use a bottle of water?
No. The often quoted research suggested an older model used roughly one 500ml bottle for between 10 and 50 answers, and that the figure changes a lot depending on location, season and how the building is cooled. There is no single honest “per question” number.
Is data centre cooling the main water problem?
No. On-site cooling is a small slice, recently estimated at around 4 percent of AI’s extra water demand. The bigger shares come from generating the electricity and manufacturing the chips.
Is the cooling water toxic?
The leftover water does contain treatment chemicals and trace metals and must be handled properly, but the evaporated vapour that forms rain is essentially clean, and regulated sites cannot legally dump the leftover into rivers or drinking supplies. The fair concern is the load on local treatment works, not poisoned rainfall.
What is the difference between water withdrawal and consumption?
Withdrawal is water taken and largely returned. Consumption is water that is gone for good from that supply, usually evaporated. Consumption is the figure that actually affects local supplies, and the one worth paying attention to.
What can I do to make my own website greener?
Pick a host powered by renewable energy, keep the site lean and fast, avoid unnecessary plugins and heavy code, and maintain it properly. The same choices that shrink the footprint also make the site quicker and better for visitors.
Sources
- The Guardian, analysis of planned US data centres on drought-hit land, June 2026.
- Tom’s Hardware, summary of the water demand split between cooling, power generation and chip fabrication, June 2026.
- Li, Yang, Islam and Ren, “Making AI Less Thirsty: Uncovering and Addressing the Secret Water Footprint of AI Models“, University of California, Riverside, updated 2025.
- United Nations University, report on data centre water demand, 2026.
- Data Center Knowledge and STL Partners, on cooling system chemicals, blowdown and water quality, 2025 and 2026.
- US Environmental Protection Agency, National Pollutant Discharge Elimination System guidance on cooling tower blowdown.
If you want a hand building a website that is quick, lean and hosted on renewable energy!
That is exactly the sort of work we do at QED web design, so feel free to get in touch for a friendly chat about your site.
