This was my rabbit hole for today - a fun and fact filled romp through AI datacentre (& other) water usage discussion from Hank Green:
Why is Everyone So Wrong About AI Water Use??
https://www.youtube[.]com/watch?v=H_c6MWk7PQc
As always - Hank takes a complex topic and breaks it down into small enough, saccharine-and-sarcasm flavoured bites that even someone as woefully under-educated and attention span deficient as I can feel smart about stuff like this.
That being said - the episode is about 23 minutes and change long - which is roughly 20 minute longer than my normal attention span lasts for web based thingies. But certainly well worth the watch.
Not gonna lie though - he did indicate that this was a hard subject to talk about accurately, as there are a number of intertwined factors that the majority of people simply can't (nor should be expected to) understand.
Dear readers - I am happy to report that I am in the majority in this case. But on to the content of the make-you-feel-smart video:
Sam Altman says that the average ChatGPT query uses around 0.000085 gallons of water, or roughly 1 15th of a teaspoon. But then, at the same time, somehow a Morgan Stanley projection predicted annual water use for cooling and electricity generation by AI data centers could reach around 1,000 billion liters by 2028. That's a trillion liters, an 11-fold increase from 2024 estimates.
Given that Morgan Stanley does appear to release the data and methodology for their calculations, and OpenAI, does not - I am apt to find Morgan Stanley more credulous, and that's phrase that I've personally never used before.
So - OpenAI First
First, Sam is talking about the water use per query. But importantly, different queries work different ways with AI. And many queries will actually result in multiple queries you never even see.
This kind of like the folks who make Fig Newtons™ list the caloric count of a serving size to be that of, say, 2 Fig Newtons™, rather than say - a whole sleeve. [1]
However . . .
This is something Sam Altman knows, but it's not something that most people know. Behind the scenes, when you ask GPT-5 a question, it frequently "thinks". They call this reasoning models.
And it "thinks" by, like, preparing and sending out other queries and then reading the results of those queries and then sending out more queries. And then maybe, like, it might spur a search of the internet. So if you ask it a somewhat complex question, it will run an initial query and then it will take that response.
It will evaluate it using another query. It sometimes runs follow-ups until it's happy with the final answer. All those extra queries are additional queries.
So one query might not be one query. Sometimes it is, but sometimes it's a bunch. So this in itself might multiply this 1/15th of a teaspoon by, like, 15.
Most LLM queries are at least 3 queries disguised in a trench-coat.
And then there's the more in-depth analysis:
Even while we're using one model like GPT-5, which is actually a bunch of models all stuck together, OpenAI and its competitors are constantly training newer, bigger versions that no one can use yet. And to create these models, like the system runs for weeks or months on enormous clusters of GPUs burning through electricity and water for cooling. It's not really fair to treat that training footprint as separate from every conversation you have with the model.
The conversation could not happen without the training. So if you wanted to be honest, you've got to make some choices. So probably you would want to spread the water used to train all of the models in GPT-5 and spread it across every query people make.
Problem here is no one knows how to do that accurately because OpenAI doesn't share this information, which is part of why it is so easy to get numbers that are both fairly correct and very different from each other. And part of why it's so easy to lie about this from either direction.
So - how does one get to these truly massive estimates of water usage?
We know that data centers use lots of water, but they also use a lot of electricity. And you know what else uses a lot of water? Power plants, specifically thermoelectric power plants. So, a lot of power plants work in the following way.
First, you make heat, then you expose water to that heat, it expands into steam, and that expansion drives past a turbine, and that turbine then spins and that creates the electricity. But then on the other side of this, no one ever thinks about what happens. It doesn't just vent out into the atmosphere.
And according to the US Geological Survey, electricity generation accounts for, get this, 40% of all freshwater withdrawals in the United States. Now, this is confusing though, because the power plants then just put a lot, not all, but a lot of that water back. So, a lot of this water is intake and then return.
So it's not apples to apples in terms of comparing water usage of datacentres to that of powerplants, but at the same time - none of this occurs in a vacuum, and water is a finite resource - whether it's processed for municipal use or not.
Every place has a finite hydrological budget. A certain amount of water that can be pulled from rivers, lakes, reservoirs, or aquifers without causing real harm. You can shift where the strain shows up, because maybe it's in municipal treatment capacity, but maybe it's in an overdrawn aquifer, or maybe it's in a river whose temperature or flow is already stressed.
But you cannot escape the fact that water is locally limited. A data center drawing from a lake is not competing with households for tap water, but it is drawing from the same watershed. And in a lot of places, that watershed is already fully allocated.
Guess where (cough Texas) a lot of these datacentre proposals are being submitted where local aquifers are likely already oversubscribed. But I'm sure that the local folks are putting their Very Best People™ on solving this and won't be wooed by intangible promises of many monies and much jobs as a result of a potential build-out.
But in the grand scheme of things - datacentre water usage is a drop in the bucket (pun like so totally intended) compared to some other uses - specifically corn farming in the states, which brings with it it's own set of peccadilloes, peculiarities and pork barreling.
On average, it takes between 600,000 and 1 million gallons of irrigation water to grow an acre of corn, depending on rainfall and region. Corn uses orders of magnitude more water than AI. According to the US Department of Agriculture, US corn production requires around 20 trillion gallons of water per year, compared to the total estimated global AI data center water use of around 260 billion gallons.
In other words, American corn alone uses nearly 80 times more water annually than all of the world's AI servers combine. And I totally forgive you if you are thinking right now, okay, Hank, yes, but corn is food. We eat it.
Food is very important for people. But that's the thing. We don't eat it.
Maybe 1% of corn is eaten by humans. A lot of it is eaten by livestock. But 40% of it is burned in our cars and trucks.
That acre of corn that evaporated a million gallons of irrigation water will get you roughly 500 gallons of ethanol. So before we even talk about processing, every gallon of ethanol already carries an irrigation footprint of around 1500 gallons of water. Extend that to 40% of the US corn crop.
I mean that may seem like whataboutism, but I see it as perspective setting.
When we talk about water use, it makes sense that you and I don't have a deep understanding of all of this complexity. You do not need to have the level of complexity that you now have having watched this I don't really need to have it either. The reality is some areas are right up against their hydrological budgets.
They can't have new uses. Others have room. Some uses, like irrigating the entire corn belt, involve staggering amounts of water that we've just learned to see as normal.
And I get why people jump on AI water use. Wasting water feels immoral. We are told our whole lives to turn off that sink while we brush.
I'll leave you all with some of my favorites from the conclusion, which I will undoubtedly shamelessly steal and quote in some form or another in the future:
I think that our entire economy is being wagered by not very many people making very strange choices based on an imagining of the future that is, honestly, I don't think likely to occur. Which is not the topic of the video, but I ended up here anyway because I started talking about what I'm most worried about. Like, I can't predict the future.
There seems to be a great deal of debate over whether these tools are actually that useful at all, which I can't find a place in. Like, I just simply don't know. But we cannot predict the future.
We cannot even, apparently, agree upon the present. But yes, in conclusion, resource analysis is complex, the incentives are weird, and we have a very long history of underestimating how dumb corn ethanol is. And all of that combined means that it is very easy to lie about AI water use.
And that's why I drink. [2]
[1]: Shamelessly stolen from the brilliant stand up comedy of Brian Regan.
[2]: Shamelessly stolen from the brilliant stand up comedy of Doug Stanhope
#AI #AISlop #AIDataCenters #WaterUsage #RabbitHole #CornSubsidies #UsPol
