Whether or not you think water use is bad for AI, another aspect hard to ignore with it is that the closed loop systems are mixing PFAs into the water for closed loop systems
It’s always backwards presumption: “No testing for Pfas air or water pollution has yet been done, and companies are not required to report the volume of chemicals they use or discharge.”
Life without batternberg cake is something too horrible to contemplate. Unlike AI, the squares are beautifully and reliably deterministic in their placement and the marzipan surrounding it is truly enhancing to society.
This comparison sounds to me like it is downplaying the significance of the numbers. California's almond-growing is another gross environmental crime, which is regularly controversial in California.
Since California produces about 80% of the world's almonds, it also means that the global water footprint for almonds is about the same as the global water footprint for AI.
More directly, a global analysis does not mean the AI data centers are evenly distributed around the world. From the article:
"A Bloomberg analysis suggests about two-thirds of data centers built or planned since 2022 are water-stressed regions, concentrated in Australia, Chile, Mexico, Spain, Saudi-Arabia, and the US Southwest (Nicoletti et al., 2025). Within the United States, Texas hosts over 400 data centers projected to consume nearly 3% of statewide water supply by 2030 (Martinez, 2025), in a state grappling with recurring drought conditions and aquifer depletion."
"Corporate sustainability reports confirm this concentration: approximately 42% of Microsoft's water consumption occurs in water-stress regions, and 15% of Google's water use is concentrated in high water-scarcity areas (Barratt et al., 2025)."
Comparing water-stressed regions in California for almond production to water-stressed regions used for data centers does seem relevant.
I agree with your link that the unsustainable use of water for almond farms causes problems, and agree that these problems are quite similar to that of water use for AI, so both should have high scrutiny. Quoting your link:
> Almonds use roughly the same amount of water as total indoor water use in California. This is unsustainable.
> There are simply too many almonds being planted in areas with unsustainable water supplies (minimal to no groundwater, uncertain or no surface water). ...
> We’re not opposed to growing almonds; we’re opposed to growing them unsustainably and taking water away from the environment and our cities. Almonds are a luxury and export crop; approximately 70% of the state’s crop is shipped abroad for the express benefit of a handful of industrial-scale growers and hedge funds.
> Water is a public trust resource, and Californians must regain control of it as they face a future certain to be beleaguered by drought, low reservoirs and over-drafted aquifers. We can't continue to subsidize agribusiness plutocrats with the water ratepayers need for basic survival.
Water for AI also uses unsustainable water supplies, directs water away from the environment and our cities, and is subsidized for big tech plutocrats paid for by water ratepayers. It also sure seems to me that most AI use falls in the "luxury" category.
Did we somehow mass build pure lossy evaporatively cooled data-centres in the AI era while I was asleep ? Most of them im familiar with lose at best 3-4% of their water per day due to recovery systems even when they are evaporative or open loop to a lake. Yet nearly everyone I hear talking about this seems to treat it as if they lose 100% of their water loop contents per day as if it’s a matter-antimatter destruction system…
The last datacenter I stepped inside (an FB datacenter in the southwest), they were very proud of how the whole thing operated as a giant swamp cooler. Unclear how they could have been limiting their losses all that much, given the evaporative stacks were open to the outside breeze
I’m familiar with datacentres, I’ve helped build a few, so I kept up with the new systems, they still aren’t that lossy last time I checked their specs.. however obviously I don’t have access to all the designs popping up in this mad rush.
One can criticize ai data centers but the "water consumption"-argument is by far the least-intelligent one and on track on becoming the same kind of populist urban myth that got Germany cancelling (ie destroying) its nuclear plants in order to... checks notes.... save the world.
Can someone explain why the data centers are using this insane amount of water. What's the ratio of closed system, evaporation?, and straight-down-the-drain water?
The use is definitely for cooling, but I'd also be curious to know what happens to the water later. If it's a closed loop circulation, the water should be returned back to the environment, just warmer than it was originally. Surely that wouldn't be so bad?
The water use is for cooling and sometimes includes indirect use for electricity generation. But before you call it "insane amounts of water" consider that the 700.000 liters quoted for training GPT-3 is less than 1 time spraying a 9 acres field, which is the size of a tiny hobby farm not a real farm.
The eagerness of the corps to switch from "we're all green" to 'fuck Earth with AI' is stunning. I was always suspicious towards the 'corporate left' but this one takes the cake. We must never trust those people again.
lol no, if this is the case then we would have banned beef: a grossly inefficient protein source that takes absurd amount of energy and water, and produce significant greenhouse gas waste. We would also ban disposable cheap goods.
Yet we don't.
If people don't like AI, they should just say so, no need to skirt around.
I agree energy use is the more concerning issue, while water has probably caught more press because it's more tangible. But there's no getting around thermodynamics. That "wasted energy" is dissipated using large amounts of water.
Additionally, using more water allows datacenters to be more energy efficient because evaporative cooling uses less energy than dry cooling. Datacenters outside of water stressed areas should absolutely being using water because doing so is better for the environment.
I think one of the more interesting points is this:
> Two-thirds of post-2022 data centers are located in water-stressed regions.
I don't think it's wrong to ask questions about that _IF_ you also ask questions about all the other usages of water, like the Californian almond stuff.
Just saying "You shouldn't use AI because it uses too much water" is kind of ridiculous.
Misconception of the year. If you can guarantee the steam it produces rains down in areas that actually re-join the water cycle easily, you might be right (even though you still have to cover the just in time demand). You can not guarantee that though.
Best regards, European person where lakes dry out atm.
They are not drying up because of AI datacenters and especially not because of European datacenters.
While man-made climate change is undoubtedly real, the yearly "our lakes dry up"-media cycle is not proof of it at all. Check if the return to normal in November will make the news.
I am aware and never claimed that. Though, the global push towards mega factories, data centers and whatnot does not help the situation. Lots of places here in Germany already struggle with having enough water, so the mindset to "lets just build a Tesla mega factory here" is something I never understood. Can't be that different for data centers. We should use our resources smartly, not for profit margins.
"Our lakes run dry" is a meme at this point. Apparently the doomers entirely forgot their kvetching about Lago di Garda drying out a couple of years ago. Memory-holed doesn't even closely describe this sort of behavior.
I agree that water is an important resource but the wording of "consumption" always rubbed me the wrong way. Is the water contaminated with chemicals? Maybe heat? Can people or wildlife consume it? Can you use rainwater?
> I agree that water is an important resource but the wording of "consumption" always rubbed me the wrong way
While the Earth is (mostly[1]) a closed system when it comes to water, any particular region on Earth is not a closed system with respect to clean drinking water.
Deplete groundwater reserves faster than they refill? No more water for you. Raise the ambient temperature far enough that the rainclouds go somewhere cooler? No more water for you. As long as your usage is below the replacement rate, all seems well. Imbalance the system, and you have indeed "consumed" your local water supply.
> Is the water contaminated with chemicals? Maybe heat?
Typically yes. Definitely yes.
> Can people or wildlife consume it?
With enough water purification infrastructure, sure. I'm not aware of very many datacenters installing such infrastructure.
> Can you use rainwater?
It's all rainwater, whether directly or indirectly. The question is what would have benefited from that rainwater had it not been redirected to a datacenter?
> Imbalance the system, and you have indeed "consumed" your local water supply.
Right, but this is not what is reported by "water consumption". It would be much better to have some kind of index (probably multiple) that estimates how close a particular feature of the water system is to a problem. Is the temperature too high, is there too high a concentration of some contaminant, etc etc.
Whether or not you think water use is bad for AI, another aspect hard to ignore with it is that the closed loop systems are mixing PFAs into the water for closed loop systems
https://www.theguardian.com/environment/2025/oct/04/pfas-pol...
It’s always backwards presumption: “No testing for Pfas air or water pollution has yet been done, and companies are not required to report the volume of chemicals they use or discharge.”
”AI's global water footprint could reach 4.2–6.6 billion cubic meters by 2027”
This is about as much as the annual consumption of the California almond farms.
(An acre-foot is approximately 1200 cubic meters.)
https://www.c-win.org/cwin-water-blog/2022/7/11/california-a...
I'd kinda rather have almonds than data centers, If I could choose.
Life without batternberg cake is something too horrible to contemplate. Unlike AI, the squares are beautifully and reliably deterministic in their placement and the marzipan surrounding it is truly enhancing to society.
This comparison sounds to me like it is downplaying the significance of the numbers. California's almond-growing is another gross environmental crime, which is regularly controversial in California.
That's because the Californian almond industry is exclusively located in California, unlike the global data centre industry.
Since California produces about 80% of the world's almonds, it also means that the global water footprint for almonds is about the same as the global water footprint for AI.
More directly, a global analysis does not mean the AI data centers are evenly distributed around the world. From the article:
"A Bloomberg analysis suggests about two-thirds of data centers built or planned since 2022 are water-stressed regions, concentrated in Australia, Chile, Mexico, Spain, Saudi-Arabia, and the US Southwest (Nicoletti et al., 2025). Within the United States, Texas hosts over 400 data centers projected to consume nearly 3% of statewide water supply by 2030 (Martinez, 2025), in a state grappling with recurring drought conditions and aquifer depletion."
"Corporate sustainability reports confirm this concentration: approximately 42% of Microsoft's water consumption occurs in water-stress regions, and 15% of Google's water use is concentrated in high water-scarcity areas (Barratt et al., 2025)."
Comparing water-stressed regions in California for almond production to water-stressed regions used for data centers does seem relevant.
The Californian almond industry has had over 100 years to develop. The AI data centre industry has had 4.
And imagine, it's already at the water usage of an industry that had a 100 years to develop. Great strides!
I agree with your link that the unsustainable use of water for almond farms causes problems, and agree that these problems are quite similar to that of water use for AI, so both should have high scrutiny. Quoting your link:
> Almonds use roughly the same amount of water as total indoor water use in California. This is unsustainable.
> There are simply too many almonds being planted in areas with unsustainable water supplies (minimal to no groundwater, uncertain or no surface water). ...
> We’re not opposed to growing almonds; we’re opposed to growing them unsustainably and taking water away from the environment and our cities. Almonds are a luxury and export crop; approximately 70% of the state’s crop is shipped abroad for the express benefit of a handful of industrial-scale growers and hedge funds.
> Water is a public trust resource, and Californians must regain control of it as they face a future certain to be beleaguered by drought, low reservoirs and over-drafted aquifers. We can't continue to subsidize agribusiness plutocrats with the water ratepayers need for basic survival.
Water for AI also uses unsustainable water supplies, directs water away from the environment and our cities, and is subsidized for big tech plutocrats paid for by water ratepayers. It also sure seems to me that most AI use falls in the "luxury" category.
I'd be willing to bet almonds bring more joy to more people in a single year than AI data centres ever will.
You must never meet vibe coders then
That seems closer to addiction symptoms than joy.
let's go to Rob from the valley, a vibe coder with an almond allergy for his take on the subject!
Rob will solve his almond allergy given enough tokens!
Did we somehow mass build pure lossy evaporatively cooled data-centres in the AI era while I was asleep ? Most of them im familiar with lose at best 3-4% of their water per day due to recovery systems even when they are evaporative or open loop to a lake. Yet nearly everyone I hear talking about this seems to treat it as if they lose 100% of their water loop contents per day as if it’s a matter-antimatter destruction system…
The last datacenter I stepped inside (an FB datacenter in the southwest), they were very proud of how the whole thing operated as a giant swamp cooler. Unclear how they could have been limiting their losses all that much, given the evaporative stacks were open to the outside breeze
The systems used for LLMs are several times more power-hungry than traditional servers. Like, 50-100kW per rack of B200s.
I’m familiar with datacentres, I’ve helped build a few, so I kept up with the new systems, they still aren’t that lossy last time I checked their specs.. however obviously I don’t have access to all the designs popping up in this mad rush.
One can criticize ai data centers but the "water consumption"-argument is by far the least-intelligent one and on track on becoming the same kind of populist urban myth that got Germany cancelling (ie destroying) its nuclear plants in order to... checks notes.... save the world.
Can someone explain why the data centers are using this insane amount of water. What's the ratio of closed system, evaporation?, and straight-down-the-drain water?
The use is definitely for cooling, but I'd also be curious to know what happens to the water later. If it's a closed loop circulation, the water should be returned back to the environment, just warmer than it was originally. Surely that wouldn't be so bad?
The water use is for cooling and sometimes includes indirect use for electricity generation. But before you call it "insane amounts of water" consider that the 700.000 liters quoted for training GPT-3 is less than 1 time spraying a 9 acres field, which is the size of a tiny hobby farm not a real farm.
The issue with AI datacenters was never about water. Water is an easily dismissable strawman that sloppers love to bring on.
The real issue with AI datacenters is the huge amounts of wasted energy, after decades of energy efficiency efforts.
The eagerness of the corps to switch from "we're all green" to 'fuck Earth with AI' is stunning. I was always suspicious towards the 'corporate left' but this one takes the cake. We must never trust those people again.
What influence do you have over these companies where your trust or distrust matters?
lol no, if this is the case then we would have banned beef: a grossly inefficient protein source that takes absurd amount of energy and water, and produce significant greenhouse gas waste. We would also ban disposable cheap goods.
Yet we don't.
If people don't like AI, they should just say so, no need to skirt around.
It's not AI, anymore than the Microsoft setup wizard is a sorcerer.
I don't like the snake oil being sold as AI.
I agree energy use is the more concerning issue, while water has probably caught more press because it's more tangible. But there's no getting around thermodynamics. That "wasted energy" is dissipated using large amounts of water.
Additionally, using more water allows datacenters to be more energy efficient because evaporative cooling uses less energy than dry cooling. Datacenters outside of water stressed areas should absolutely being using water because doing so is better for the environment.
Water consumption of AI is a non-issue. Why even spend research and policymaking resources on this?
I think one of the more interesting points is this:
> Two-thirds of post-2022 data centers are located in water-stressed regions.
I don't think it's wrong to ask questions about that _IF_ you also ask questions about all the other usages of water, like the Californian almond stuff.
Just saying "You shouldn't use AI because it uses too much water" is kind of ridiculous.
...I don't think you can declare that something is a non-issue without having spent at least some resources to determine whether it is?
For who?
Misconception of the year. If you can guarantee the steam it produces rains down in areas that actually re-join the water cycle easily, you might be right (even though you still have to cover the just in time demand). You can not guarantee that though.
Best regards, European person where lakes dry out atm.
They are not drying up because of AI datacenters and especially not because of European datacenters.
While man-made climate change is undoubtedly real, the yearly "our lakes dry up"-media cycle is not proof of it at all. Check if the return to normal in November will make the news.
>They are not drying up because of AI datacenters
I am aware and never claimed that. Though, the global push towards mega factories, data centers and whatnot does not help the situation. Lots of places here in Germany already struggle with having enough water, so the mindset to "lets just build a Tesla mega factory here" is something I never understood. Can't be that different for data centers. We should use our resources smartly, not for profit margins.
"Our lakes run dry" is a meme at this point. Apparently the doomers entirely forgot their kvetching about Lago di Garda drying out a couple of years ago. Memory-holed doesn't even closely describe this sort of behavior.
I think you mean "re-join the FRESHWATER cycle"? Because obviously it rejoins the global water cycle.
0.16% (high estimate) of global water usage by the way.
This statistic doesn't show the disproportionate effect it may have on local communities close to these AI datacenters.
I agree that water is an important resource but the wording of "consumption" always rubbed me the wrong way. Is the water contaminated with chemicals? Maybe heat? Can people or wildlife consume it? Can you use rainwater?
> I agree that water is an important resource but the wording of "consumption" always rubbed me the wrong way
While the Earth is (mostly[1]) a closed system when it comes to water, any particular region on Earth is not a closed system with respect to clean drinking water.
Deplete groundwater reserves faster than they refill? No more water for you. Raise the ambient temperature far enough that the rainclouds go somewhere cooler? No more water for you. As long as your usage is below the replacement rate, all seems well. Imbalance the system, and you have indeed "consumed" your local water supply.
> Is the water contaminated with chemicals? Maybe heat?
Typically yes. Definitely yes.
> Can people or wildlife consume it?
With enough water purification infrastructure, sure. I'm not aware of very many datacenters installing such infrastructure.
> Can you use rainwater?
It's all rainwater, whether directly or indirectly. The question is what would have benefited from that rainwater had it not been redirected to a datacenter?
[1]: https://en.wikipedia.org/wiki/Atmospheric_escape
> Imbalance the system, and you have indeed "consumed" your local water supply.
Right, but this is not what is reported by "water consumption". It would be much better to have some kind of index (probably multiple) that estimates how close a particular feature of the water system is to a problem. Is the temperature too high, is there too high a concentration of some contaminant, etc etc.
> this is not what is reported by "water consumption"
You are make a semantic distinction here, but "consumptive use" is a term of art in modelling water usage: https://en.wikipedia.org/wiki/Consumptive_water_use
Yes but: https://en.wikipedia.org/wiki/Consumptive_water_use#:~:text=...
You can't drink it once it's gone through a bunch of mouldy pipes, no.
cf