How much water does AI consume?

BBC Sounds Music Radio Podcasts Hello and thanks for downloading the More or Less podcast. With a programme that looks at the numbers in the news and in life and an AI water consumption. I'm Charlotte McDonald. When you sit in front of an AI chatbot and start typing away, the responses appear on your screen like magic. Information apparently springing out of fresh air. The truth is of course, very different. When we send our queries off, these are dealt with by a vast network of data centres which were into action to come up with the answers. These data centres contain servers, which themselves contain processing chips. Those run on electricity from the power grid and as they operate, they generate lots of heat and need to be called to prevent overheating. Both electricity generation and cooling data centres use water. With AI expanding rapidly, some people are worried about how AI's water use might escalate in the future. One striking figure came in a book called Empire of AI, written by the US journalist Karen Howe. According to her, surging AI demand could consume 1.1 trillion to 1.7 trillion gallons of fresh water globally a year by 2027, or half the water annually consumed in the UK. That's between 4.2 and 6.6 trillion litres of fresh water, which sounds like a big number. But should we trust it? Nathan Gower has been looking into this one. Hi Nathan. Hi Charlotte. Nathan, 4.6 trillion litres of fresh water sounds like a lot. It definitely is, but there's all sorts of problems here. So, Howe's claim is about the amount of water that could be consumed. This has a specific meaning. It's water that's taken out of a system or source but not returned. This happens through evaporation when it's used for cooling, whether that's onsite at data centres or offsite at the power plants where electricity is generated. They also evaporate water when they call the generators. OK, so this claim seems to be saying that trillions of litres of water are going to get consumed. You know, they're taken out of the water system and they're not returned. Yes, but the claim is just not right. Howe got her figures from a paper published by academics from the University of California, Riverside. Except those figures of 4 to 6 trillion litres aren't for water consumption. They're for something different called water withdrawal. That is the total amount of water that gets taken out of a water system or source. Now some of this water will be returned, but some of it won't. It will be consumed. So basically the consumption figure is a subset of the wider withdrawal figure. OK, so this author stuck the wrong label on the 4 to 6 trillion litre figure. It should be for withdrawal, not consumption. So what is the actual consumption figure? So the paper says between 380 and 600 billion litres could be consumed in 2027. That's about 10% of the original 4 to 6 trillion figure. This mistake was pointed out by an American substacker, Andy Masley, and Karen Howe has since issued a correction. Right, so that's settled now. It would be really great if it was. So Karen Howe read those figures in the academic paper and misinterpreted them, mistaking the larger figure for withdrawal as the figure for consumption. But I've discovered that those figures in the paper were actually wrong in the first place. How so? The paper says it takes an estimate for global AI electricity use in 2027 and extrapolates from that to a figure for global AI water use in the same year. But that electricity estimate comes from work by another researcher, Alex De Vries-Gow. I spoke to Alex and it turns out that he wasn't estimating total global AI electricity use in 2027. Instead, he made an estimate for global AI electricity use by those AI servers that might be produced in 2027 alone. Right, so it's not counting all the AI servers and data centres built in previous years, most of which will presumably still be running. Exactly. So they've basically underestimated electricity consumption and that throws off the water estimates. Then on top of all this, an author comes along and misinterprets these flawed figures. This sounds like a mess. It's a bit of a mess. So let's try and salvage something. That researcher, Alex De Vries-Gow, has also made estimates for global AI water consumption. To do that, you need to know how much electricity global AI systems might demand. Now the tech giants don't publish these numbers, so Alex came up with a novel strategy. Just to get to the power demand, you actually need to take a deep dive into the supply chain of AI hardware. What I did was looking at how many AI chips could have been produced in the past years, how many AI server modules could have been made with that and ultimately how many AI servers have been made with that. When you get to that point, you still need some assumptions to figure out like how much are those servers going to be utilised. What I did there was I tried to look at the biggest buyers of AI server equipment, which is large tech companies like Microsoft, Google, etc. I examined how do these companies, data centres typically perform in terms of water intensity and I used those numbers ultimately to translate my power demand estimate into an indirect water consumption estimate for AI server hardware. Then on top of that, you still need to include the water that's actually being consumed in the data centre itself, the direct water consumption. Using this method, Alex estimated that AI systems at the end of 2025 were consuming water at a rate of 750 billion litres per year. Now I know what you're going to ask Charlotte, is 750 billion litres a big number? Well here's Alex. This is absolutely a big number. This is exceeding the level of global bottled water consumption, which is at about 446 billion litres. That in itself is a significant amount of consumption. But is it really a problem? Well it could certainly be. It could cause a lot of problems if this consumption is concentrated in a single location where water scarcity is already a potential problem. But we just don't know at this time. So that really makes it hard to make that translation. Is this a problem or not? But at the same time it also makes it impossible to say this is not a problem at all. It's a pretty huge number. It's going to have an effect on local fresh water supply. For sure, we just don't really know where and we just don't really know how much this is going to hurt in which locations. For Alex's estimate for water consumption, only about 10% of that is happening on-site at data centres, which typically use drinking water from local supplies. The other 90% is happening off-site at power stations, which use water from sources like rivers and lakes. Also, Alex has only made estimates for water consumption. He thinks that's the most important metric when thinking about overall water scarcity. But I've spoken to another researcher who thinks that water withdrawal matters equally if not more and argues that even if water is eventually returned to a system or source, increased demand can still have important consequences. What about future predictions? Here's Alex again. One of the big bottlenecks that is starting to appear, can these tech companies even find sufficient power to power their data centres? They're getting a lot of equipment. Where are they going to be able to actually use all that equipment? I can make statements based on, let's say, the current supply chain capacity for producing AI hardware, which I know this year is probably going to be at least similar to the previous year, 2025, which means that the cumulative power demand of AI systems is still going to be rising. This is still going to be adding on top of the production of the power three years. But again, I can't say anything about whether it's going to be possible to find a home for all this equipment. Well, thank you, Nathan. And thanks to Alex DeVries-Gow, as well as Professor Shaolay Ren, who also helped with this episode. That's all we have time for this week. But if we have any more questions or comments, please email us on more or less at bbc.co.uk. We'll be back next week. And until then, goodbye.