When the well runs dry.

Whether AI uses too much water may turn out to be the wrong question. The real question is whether we're putting some of the world's thirstiest digital infrastructure in the places least able to afford it.

If AI doesn't destroy us anytime soon, there's another code-red concern waiting to feed your anxiety levels: water scarcity. This one has been on the doomsday warning list for years, long before AI found its voice.

Researchers project that AI-related water demand could reach as much as 6.6 billion cubic metres annually by 2027. For context, that's around half of the UK's annual water use. Or four times the amount needed to keep New York City running. So, we need to pay attention. Or do we?

The UK government seems to think so. There is no shortage of alarming statistics. For example, fresh water demand is expected to exceed supply by 40% by the end of the decade and 55% of global data centres are in river basins with a high risk of water pollution, heaping pressure on clean water supplies.

The problem is where the data centres are built, not what you type in ChatGPT. Of those proposed in the UK, 84% are planned for areas already short of water, and most of them are in the South East, the driest corner of the country.

A Drop in the Ocean

I was asked, recently, for my views on AI/data centre water use. I skirted around the question and sat on the fence because, as with any debate around resources or climate, the arguments are complex and nuanced. Analyst Benedict Evans has been commenting on this for a while, citing disproportionate moral panic.

In a LinkedIn post two weeks ago, he pointed out that total US freshwater withdrawal is 102tr gallons a year, of which data centres account for around 0.017%. Benedict's point is that this is often a hyperlocal planning issue and critical for communities. But, he writes,

"The idea that AI overall is using lots of water is nonsense."

So, nothing to worry about after all. And that's the problem. Unless you immerse yourself in the science, much of which is opaque and frankly who has the time, then what position should you take?

Will You Turn the Tap Off?

The point is this, though: Will any of the arguments force you to alter those new AI habits? Or stop you from asking ChatGPT to sense check that email to your line manager asking for a raise, because "what about the water?"

In time, you might adjust your behaviour. But there's a lag in environmentalism where it's always someone else's problem until it becomes socially irresponsible not to follow the crowd. Ergo: seatbelts, drink driving etc. For now, we're too captivated by our new friend Claude and his mates to really take notice.

If You're in the C-Suite

But what if you're not a sole-user but sitting in the c-suite. You could try this for starters:

You are a climate scientist specialising in water resource scarcity and its intersection with digital infrastructure, particularly the water footprint of cloud computing and AI/data centre operations. You have deep technical expertise in water accounting methodologies (including Water Usage Effectiveness, indirect water-energy nexus calculations, and Scope 3 emissions.....I run an information services business with 250 employees and approximately £12 million annual turnover. We use Claude Enterprise accounts across our workflows for various business functions. I need to calculate our organisation's water footprint attributable to this AI usage so I can factor it into our carbon and resource offsetting strategy...

(It's a long prompt, DM me for the full version)

Facts on Tap

The real fear, of course, is that by the time regulators, planners and infrastructure providers catch up, it might be too late. Next time the issue comes up, in the boardroom, with clients, or privately - and it will - use these handy facts:

That "one email equals a bottle of water" figure you may have heard is out of date. The researcher who produced it now says so himself. The models got more efficient.

Nobody can tell you with confidence what the average prompt costs. Published estimates vary wildly because companies disclose different data, use different methodologies and rarely submit their figures to independent audit.

Your own usage probably isn't the main issue. Even with some of the higher published estimates, hundreds of prompts a day consume small amounts of water compared with most household activities. You could take the view, of course, that it all adds up.

No UK data centre is required to tell anyone how much water it uses. Ministers rejected a pause on new approvals in August. That, rather than your prompts, is the thing to be concerned about.

Leave It to the Water Companies

Stacked up against the current flow of existential warnings, water scarcity feels like one more thing. And we only have limited headspace for dire warnings. It's far easier to rationalise and hope that "they" will figure something out, desalinate the sea, and build sustainable data centres in regions not already under water stress. Better use of rain water, closed loop cooling systems, responsible planning.

After all, we have such robust well-run water utility companies in the UK. Thank goodness we don't have to worry about the lack of new reservoirs, leakages and polluted rivers. Or droughts. If that were the case, AI's need for freshwater on tap could cause unimaginable problems.

Excuse the cynicism, but once again AI and progress is outpacing the ability of legislators to keep up. But keep up, they must. They need to rehydrate, refocus and act now. We can only hope.

The Piddington Test

Back to the hyperlocal issue. As the UK has relearnt this week, all politics is local. (Just ask the folks in the Principality of Piddington) According to a report in The Guardian, hundreds of datacentres are awaiting planning permission with 315 standing by to connect to the National Grid. Regardless of what big science tells us, it's in places like Piddington where the big data debate will play out.

What the Machine Said

As for that prompt. This was the response from Gemini, summarised:

For an organisation of 250 employees using Claude Enterprise, your annual AI-attributable water consumption is estimated between 3,600 to 24,000 litres, depending on your team's query volume, context length, and the physical location of the cloud data centres handling your requests....Because water scarcity is a hyper-local issue rather than a global one like carbon, any offsetting or ESG strategy should focus on Volumetric Water Benefit (VWB) restoration projects ...

(again DM me for the response)

To put that in context, my family uses 184,000 litres a year for four people. That works out at 126 litres each a day, just under the UK average.

Ask Gemini how we mitigate against this and the answer includes a Volumetric Water Benefit (VWB) project. In plain English, bank the offsetting, restore the wetlands and repair the leaking pipes.

And yes, I hear you on the irony of asking AI to solve the problem of AI data centres using too much water to cool steaming servers. But unless you have a better idea...

Whether AI uses too much water may turn out to be the wrong question. The real question is whether we're putting some of the world's thirstiest digital infrastructure in the places least able to afford it.

Human IntelligenceSubscribe on LinkedIn