I mean even disregarding the environmental cost, it’s rapidly accelerating our decline from the “Information Age” where knowledge about anything can be found in an instant to the “Disinformation Age” in which it’s impossible to tell what is fact from fiction because information is overflowing everywhere but most of it is bullshit.
post
Most of those numbers shouldn't be hard to validate, but a reliable source for the 16oz number (how much a single AI prompt uses) would be appreciated. A search turns up a lot of sites I've never heard of (or are unreliable), so I don't know how reliable they are.
Its nonsense.
a) 16oz is 0.5 liters. To evaporate 0.5 litres from 20c to 101c you need 361 Wh of energy. An nvidia B300, top out at 1400W. In order for a single prompt to use 361Wh on a B300, the prompt needs to take 15 minutes to run. Thats clearly not happening.
b) Even when water is used to cool GPUs, it isn't destroyed. Its heated, and then cooled down again. The water just moves heat around the system, it doesnt get destroyed
Okay, while I do suspect the 16oz number is pulled from someone's ass (or just very old, many efficiency improvements ago), there are a couple of things to consider:
a) A single B300 won't even run frontier models. It only has 288 GB of HBM. The DGX B300 server has 2.1 TB, which I suspect is not enough either (running Qwen 3.8 Max lossless (BF16) requires 4.9 TB just to load the model). But these are run in clusters anyway and any single user's prompt only uses a tiny percentage of the compute from tons of servers. OpenAI and Anthropic are going for size with their models so the most expensive ones probably need even more than Qwen.
b) While the GPUs themselves are directly cooled by closed loops, the water inside those closed loops is often cooled by evaporative coolers, or even AC and evaporative cooling in tandem (that allows the evaporation to happen outside the datacenter itself). That's because evaporative cooling is the most energy efficient form of cooling.
There are companies offering very complex heat systems that take advantage of multiple forms of cooling because they're making 100+ kW racks these days and talking of 500-1000 kW racks in the future. Ridiculous numbers, honestly.
This is from some company's website that offers cooling solutions: Notice the outside "cooling tower" - those can be dry or wet. In a hot and dry climate, you'd want wet. It'll cool really well. In a colder and more humid climate, dry might work better. They also optionally let you heat the aquifer apparently, which is bound to cause issues for everyone living nearby over time.

The b300 figure is to give a sense of scale, its clearly more on the order of milliliters per query than half a liter.
Evaporating water into the atmostsphere is really just recirculating it, it hasn't spoiled the water or destroyed it.
The problem with the 16oz number is that its so clearly wrong that is discredits any other valid arguments. If complaints about ai resource usage want to be taken seriously, forget the water "usage" and worry about the gas/coal/diesel generators pumping carbon directly into the atmosphere. Because that is far more concerning that water getting cycled through the water cycle.
I believe the paper was looking at total footprint including building and running data centers with open evaporative cooling, and the water "used" by civilian power stations.
I don't really see more water in the air as the most pressing environmental issue around data centers. That's better than the water being destroyed or contaminated.
The main environmental hazard of data centers is the rapid expansion of them. The running of them for months on end in the training of models... And above all else, the fact billionaires have chosen to power them in the least environmentally friendly ways (gas and disel):
https://duckduckgo.com/?q=dirty+power+ai+data+centers+powered+by
I'm much more concerned about the electricity going in as well. If they want to build these huge power sinks, they should be required to build their own green power for them as well, not just buy from the already constrained grid. And burning gas or any other fossil fuel for them is just madness.
Yep, we're paying for answers we know are wrong, using the currency of environmental health. It's probably the dumbest thing I've seen humanity do.
People who use the technology need to use things like AnythingLLM for running answer machines locally (yes, AnythingLLM can go online, search websites, and do GPT style answering)... and something like Kolbold.cpp (also run locally) for creative writing tasks.
AnythingLLM is very simple to set up, Kolbold.cpp has tutorials on YouTube and is far less restrictive than ChatGPT. Both avoid interacting with AI data centers (now the available models have already been trained).
Self hosted definitely is the way to go. I have a second hand Intel nuc + nvidia p100 setup, and its more than enough for what I need it to be. And it only uses 250w when im using it, I turn it off when I'm done. Plus my power is all my own solar, so effectively free.
You're talking about closed water cooling systems, which are ethical and more expensive.
Now can you guess how many data centers use those?
If we are playing the guessing game, almost all of them. Its not that expensive to build and run a closed loop water system. You cant just run plain tap water through cooling loops, it needs to be distilled or treated, or youll grow mold inside the loops and get clogs.
Provide your source for these data centres using open loop systems.
Almost none of them run closed loop systems. They’re extremely expensive, and even if they did they actually still use just as much water, it’s just that the water is now being spent at a gas or coal power plant instead of on site due to the increased energy usage.
The systems that most data centers use are called adiabatic cooling systems. They evaporate water after a closed loop because the problem isn’t moving the heat away from the servers, it’s getting rid of them fast enough.
Conveniently most data center providers provide absolutely no statistics about these things because they look very bad to the public, but researchers have calculated usage.
https://www.washingtonpost.com/technology/2024/09/18/energy-ai-use-electricity-water-data-centers/
https://www.eesi.org/articles/view/data-centers-and-water-consumption
They use evaporative cooling on the AC that actually cools the inside of the data center. Evaporating water cools things down significantly more than simply splashing cold water at them.
This significantly reduces their costs on electricity because AC runs much more efficiently when the condenser side is cool.
They do not (usually) use evaporative cooling on the GPUs themselves. Evaporative cooling is used more directly too, but the indirect variant is better because it doesn't raise the humidity inside the data center. This is why so many data centers are being built in deserts now (well this and the cheap land).
When they do use it without the heat pump/AC in tandem, then it means pushing hot air through a water soaked membrane, not using all that water in the coolant pipes for the servers. The closed loop is still closed, the evaporation happens outside of the loop.
An interesting guess in a capitalist dystopia.
The answer is: almost none of them.
Ethical would be if you dumped the heat into space and only used solar power.
Ethical would be if you dumped the Tech Bros into space and switched to solar power.
Make them fly there on one of their own rockets, for the irony.
There's also data centers using closed loop systems, they don't lose water
https://www.aitooldiscovery.com/ai-infra/how-much-water-does-ai-use
Several places report 6.4 Billon Gallons per year by Google alone in 2023. Even if we assume this has icreased across all companies, the 163 Billion Gallons per year seems overblown by quite a bit.
I hate it when people lie to make a point that would have stood well on it's own without lying. Even an estimated 13 Billion Gallons per company and year (so assuming it doubled since 2023) is fuckton of water.
The extremely dumb thing here is that the water use is totally fine. California alone uses ~11 TRILLION gallons of water per year for farming. And so what, that water cycle returns via precipitation (assuming we don't go so hard to fuck up the aquifers, but you know what I mean). The water is fine. Plus that farm water runoff is even kinda hazardous, with pesticides etc. It's still fine
This is a totally recyclable resource. You can probably use it to water plants after it's been through the data center, or it may have just evaporated away as part of the cooling process
Meanwhile, the obvious dart to throw is the fact that the power need is absolutely ridiculous and only growing, leading to every major tech company investing extensively in energy tech. Eventually they will become a new generation of utilities that supply power (and compute) to everyone. So the opportunity is to force them to reinvent energy in a clean way rather than, say, burning an utter shitton of oil to do it
This has been done to death; that's not how water consumption works in a datacenter and most of the water quotes you hear are about construction not operations. Yes there are outliers and cases of running off tap but those are exceedingly rare and usually taken out of context. The water use just flat out isn't an issue, or at the very least isn't a substantial one.
Electrical power, heat, and noise pollution, on the other hand,...
Yeah, it upsets me when I see people posting completely batshit reasonings to make a point that is totally reasonable in the first place. It dilutes the argument against hyperscale datacenters and AI.
I did the math a while back because my workplace was talking about using more AI, and I'd need to run about 50,000 queries to copilot to use the same amount of electricity that it took to drive my car to work. If everyone using AI got one extra work from home day, it would more than offset their AI usage for an entire year.
Me: prompts a frontier LLM one time to do 30 minutes of my work in 45 seconds
Randos on the internet: "omg you're killing the planet"
Me, who once left a Minecraft technic modpack game running in the background while I had a 2hr goon session, consuming orders of magnitude more power for nothing productive:

Yeah, there are lots of reasons to hate AI, and resource consumption is absolutely one of them; but people seem to miss the forest for the trees when it comes to exactly what parts are problematic
Shiiiiiiit
I've got this chonker running a couple of game servers, I can hear it humming away through a wall right now. If I scrapped this machine and switched over to AWS or Azure, it would save so much electricity.

Like fuck I'm gonna do that though, cloud machines are expensive.
The water use just flat out isn’t an issue
It's an enormous issue when you're on the hook for inflated municipal water bills.
But I'll concede "you posted an image, so you're responsible for $100B in sunk costs" is just scapegoating.
The resources are wasted whether or not a given individual makes an AI Slop flyer for their local Swingles spouse swap meet up.
Electrical power, heat, and noise pollution, on the other hand,…
Also all big issues.
But it's kinda crazy to tell people who have seen their water bills jump $50+/mo the day after construction breaks ground that "Um, aktuly, water isn't an issue".
Anyone complaining about AI water usage that doesn't also complain about golf courses and water-intense crops grown in deserts is, to me, not actually upset about water usage. They just hate AI and have latched on to the only thing that can justify their hatred without being abstract or complex.
When a single almond uses more water than over 500 AI prompts, maybe we should be focusing on other things.
Anyone complaining about AI water usage that doesn’t also complain about golf courses
Isn't living next to a golf course.
Yeah, people have materialistic views and materialistic concerns. If it's out of sight, it can easily be out of mind.
But the thing about these mega-projects that pop up in your backyard practically overnight and immediately jack up the cost of living in your rural township is that they are highly visible.
When a single almond uses more water than over 500 AI prompts
I don't care, because I don't live next to an almond grove. But I do care when I live next to an AI prompts factory.
Even then, we're rapidly approaching parity. You're weighing 1 trillion almonds/year against 100 trillion tokens consumed per year. And the number of tokens being consumed is rising rapidly, while the volume of almonds being produced are in decline.
The other thing to consider is that token usage is hopefully at its highest cost/token. As the market matures, it should stand to reason that efficiency for both data center structure and hardware should improve.
We're looking at an extremely unregulated, poorly thought out, gold-rush type market that is trying to cash in as much as possible before reality comes in to smash it in the face. As long as we keep pushing, regulations should be able to mitigate a lot of the biggest issues we see with data centers.
The other thing to consider is that token usage is hopefully at its highest cost/token.
With the price of energy skyrocketing and a chip shortage constricting supplies? Doubtful.
We’re looking at an extremely unregulated, poorly thought out, gold-rush type market that is trying to cash in as much as possible before reality comes in to smash it in the face.
But, like with the financial crisis, nobody really expects the Tech Giants to be the ones that feel the pain. Sam Altman and Dario Amodei will die richer than the Pharaohs, no matter what happens next. It's the hillbilly schlubs in Jackson, Mississippi and Bastrop, Texas who are going to be eating shit when the turd pile finally tips over.
The almond thing is actually propaganda put out by the beef industry as part of a widespread campaign to demonize plant based protein btw
Also everyone complains about golf courses using irrigation, but no one complains about baseball fields, soccer fields, football fields, and public parks that all do the exact same thing, so maybe you're the one that isn't actually upset about water usage.
Or maybe, black-and-white purity tests like this are stupid because they leave no room for nuance and productive conversation.
I don't think water argument is real tbh
Yeah, how is the water being consumed? Presumably it's used for cooling the servers but does this result in some kind of contamination where it's no longer useful elsewhere?
Technically it's been used for light industrial purposes, which makes it non-potable after usage. It needs to be run through the treatment and testing cycle again to be fit for human consumption.
That would be a bigger issue if people were expecting to drink from the cooling tap, or if most of them didn't use grey water to begin with.
The water has moved through tubes not necessarily rated for human consumption, and exposed to moderate temperatures.
Most is used in the construction of the datacenters or for air conditioners. It's not used per query at all.
"Chillers and cooling towers are at the heart of most large data center cooling setups. The chiller cools down water, dropping its temperature before it heads to CRAH units or other cooling coils.
Cooling towers push this heat out to the open air, usually by evaporating water. Pairing these together lets chilled water systems handle much bigger loads than just using CRAC units alone.
Key components include:
- Chillers: Make chilled water for cooling coils
- Cooling towers: Get rid of heat from the chiller’s condenser water loop
- Pumps: Move water around between units and throughout the system
Chilled water systems are efficient and can scale up, but they do need a big upfront investment and regular maintenance. They’re the go-to for hyperscale and enterprise data centers where uptime and energy use really matter."
In industrial cooling applications, chilled water is run over massive radiators to dissipate heat more efficiently. This usually causes a good portion of the water to be evaporated. The problems with these massive data centers is that they can pump millions of gallons of groundwater a day only to be evaporated to the atmosphere. They end up pumping more groundwater than can be restored naturally, causing groundwater levels to fall which leads to all sorts of other issues.
Yeah anything that's flushed through dirty industrial cooling pumps is going to pick up lubricants, dust, pesticide, microplastics from parts rubbing against each other, basically all the industrial stuff that works it's way into the system which will get worse with age, and they'll have to put nasty chemicals in it to keep slime from growing.
Basically it's not drinkable and you wouldn't want to shower in it and nothing except slime can live it until its diluted enough.
Will they clean it before dumping it back into the environment every time they flush the system? The answer is no. They won't.
And the idea that it won't use that much water because they'll "recycle" it in a closed system is also bullshit. They will have to flush it all the time because water that's kept like that will foul no matter what you to do it and it will start to clog filters and cause issues. So yeah lots and lots of clean water in and lots and lots of nasty water coming out.
Oh yeah, and when the dumped dirty water evaporates, all the contaminates are left on the ground and will start to concentrate until it starts to be a bigger and bigger problem. It will get blown around into people's yards and crops and into food, pets, etc.
short answer: many datacenters have evaporative coolers that shove the water into the air
Aaaaand it's nothing compared to the rest of the economy.
You'll never see percentage of water consumption only flashy gallons liters or olimpic pools
How did they go from 16oz per prompt to 163 000 000 000 gallons of water?
1 prompt = 16 fl oz = 0.125 gallon^[1]^
1 year = 163 000 000 000 gallons^[1]^
163 000 000 000 (gallons) / 0.125 (gallons/prompt) = 1 304 000 000 000 prompts
Number of customer = 900 000 000^[2]^
1 304 000 000 000 (prompts) / 900 000 000 (customers) = 1449 prompt per customer (rounded up from 1448.888888..).
If all customers did 1 prompt/s it would take 24 minutes 9 second to reach 1 304 000 000 000 prompts (9 hours if they did it once per minute). That is by no means unfeasible but the number feels very arbitrary.
1: From the image
2: https://www.getpanto.ai/blog/openai-statistics
That paper never passed peer review, and I believe it was trying to find the total footprint. Training probably does take a lot (considering it involves building and running a data center for 3 or 4 months), but once a model is trained you can run it locally on a regular mid level PC.
It's unlikely running models locally on your own PC uses a glass of water per question. But there is a large environmental cost to building and running so many data centers or asking the latest LLMs running that way.
A lot of that environmental cost comes from the fact Governments have asked data centers to not interrupt the power grid, and billionaires have subsequently responded by choosing the dirtiest ways to power them so they're off grid.
https://duckduckgo.com/?q=dirty+power+ai+data+centers+powered+by
This was a choice made by governments, billionaires, and their investors.
But yeah, if a model was trained in an environmentally powered data center then run locally it wouldn't be such a problem.
I don't understand the point the post is trying to make.
Accelerating climate change due to increased water vapour in the atmosphere I guess? The post seems to imply that the water disappears but it's not like fracking where the water is no longer retrievable or contaminated afterwards. The pollution to the water from data centres would be heat pollution (also bad, but again, that doesn't seem to be what the post is talking about).
Let's just take a second and ignore the massive lapses of information in your post and I will ask you this: do you think the water cycle is instant?
As in, when they are mass pumping out the underground aquifers and draining the man-made reservoirs, do you think that the "magically clean water" discharge from the data centers instantly goes back into said aquifers and reservoirs?
It's simply no longer available to the plants, animals, and humans that would have consumed it otherwise. It's missing downstream.
top 67 comments