OGWiseman Reports!

I investigated A.I. water usage.

ONE: I keep hearing very smart people I know say some version of the following claim:

A.I. uses a problematic amount of water.

Having looked into it in some detail, I’m prepared to say that this claim is somewhere between very overblown and completely false.

It’s false in the sense that even if A.I. use goes up by an order of magnitude, it won’t represent a major use of water in the United States.

It’s also false in the sense that, in terms of GDP created per liter of water, it’s quite possibly the single *best* use of water in the United States. Keep in mind that a single pound of beef, which costs under $20 at the store, takes 1,800 *gallons* of water (6800 liters) to produce.

Let’s trace the history of the claim at issue and see why I think it’s basically false:

This claim seems to have started in about 2023 with the publication of “Making A.I. Less Thirsty” by an international team of scientists. After that it became common to hear that “A.I. uses a half-liter of water in an average conversation”. This claim was all over the place.

Here’s an excellent explanation of why that’s not true. In short, the paper’s authors confused “per request” and “per page” in the original research they cited, resulting in a 10x exaggeration of water usage right off the bat. Add to that some data from the earlier paper that had been updated by the time the 2023 paper cited it, and, well, you get the idea.

Despite not being true, this claim was repeated continuously, and in fact exaggerated, until the Washington Post published an article with the headline “A Bottle Of Water Per Email”, which is just obviously crazy when you think about the volume of text A.I. is generating. It doesn’t pass the smell test, and yet, it’s right there in Wapo, wild.

(As an aside, it’s crazy to think about A.I. compared to Beef *even at these fake and exaggerated numbers*. Say it *was* a half-liter of water per conversation—that means you could have 13,600 full conversations with A.I., still more than I’ve had in my life, for the water-price of a single pound of beef. And that’s at the fake number that was >10x too high!)

Fast forward to 2024, and Karen Hao published a very successful book called “Empire of AI”. In that book, she reported that there was an A.I. data center being built in a town outside of Santiago, Chile, that was going to use 1000x as much water as the entire town. This fact went viral when the book was published. If you’ve followed the issue, you’ve probably heard it.

That fact is also totally wrong, as a guy named Andy Masley demonstrated on his own Substack. She made an arithmetical error while converting units of volume, and her estimate of water usage was 1000x too high. The author herself acknowledged the error on Twitter. Whoops. (Though also good on her for not dodging.)

Now it’s fair to say that A.I. can put local stress on water supplies. Even after correcting the error in Hao’s book, that data center was still going to use as much water as an entire small town! Here’s an informative Bloomberg article about it, which unfortunately is paywalled but is the best I can find.

It’s frustrating that so many data centers are being built in water-starved areas like Texas and Arizona. From a pure water perspective, we’d be much better off with the huge data centers in our more northern latitudes that have tons of rain. But, regulatory considerations, especially concerning new electricity generation and grid interconnection, drive the efforts to those states. It’s actually a sneaky signal of how tertiary the water issue is that they’d rather try to get water in Texas than power in Michigan! (Electricity and A.I. will be the subject of another post. That problem is unfortunately not fake, and also multifaceted and complex and deeply political, and I’m still trying to wrap my head around it.)

Local stresses do matter, but it’s also important to look at the aggregate, and from that perspective, A.I. water use is not a serious problem, or even a common phenomenon, actually. Framing this issue by talking only about the amount of water that A.I. uses is actively misleading, because 1) the raw number of liters is a large number, and 2) it’s hard to fathom how much water humanity uses every day.

We use so much water that American farmers don’t even measure their water in liters or gallons, they use “feet-acres”. One foot-acre is the amount of water that it takes to cover an acre of land to a depth of one foot. One foot-acre is over 300,000 gallons of water. Just the almond industry in just the state of California uses about 5,000,000 feet-acres of water every year. That’s 1.5 *trillion* gallons of water, or 5.6 trillion liters of water, or 11.2 trillion conversations with A.I., even if you use the fake and exaggerated “half-a-liter-per-conversation” rule of thumb I mentioned in the beginning. And again, that’s just on almonds, just in California.

Enough obscure farming terminology, back to A.I. Again: All the A.I. searches I (a heavy A.I. user) have ever done have consumed less water than the amount of beef I have eaten in the last week.

Another: A.I. currently uses 5% of the water that America uses just on its golf courses.

My personal opinion is that an agentic, personable, infinitely patient and industrious index of all human knowledge is worth at least 1/20th as much as playing golf, which in the immortal words of Ben Franklin is mostly a way to spoil a perfectly good walk.

To bring some economic rigor to that point: US golf courses have a total annual revenue of about $35 billion. A.I. has revenues more like $150 billion. They use 5% of the water to generate almost 5x as much revenue. A.I. generates over a buck fifty per gallon of water used, golf generates two cents. And that’s just pure revenue—it doesn’t include the downstream productivity-enhancing effects of A.I., nor how much work corporate America misses to play golf!

Of course, it might not be fair to take current numbers as representative, perhaps it’s the trends that are worrying. A.I. use is expected to grow rapidly, after all. Well, conditional on a 3x increase in A.I. water consumption by 2030, A.I. will at that point use… 1.3% of the amount of water then reserved for growing corn used to make ethanol.

We really use a lot of water.

When considering long-term trends, it’s important to think not just about inputs, but outputs. Yes, A.I. use increasing rapidly will make it a larger and more worrying share of total water usage. It has a long way to go before it dominates, but it’s not so far to it being a part of the mix of concerns, not co-equal to ag or ranching, but secondary rather than tertiary at least. I don’t mean to be glib about that.

But remember, as I often say in this space, when it comes to A.I., *this is the worst it is ever going to be*.

If A.I. use does double and then double again, if the megacluster GPU farms they’re now building really do come online, if the cycle of investment and research proceeds until 2030, 2040, beyond, without being strangled by regulation or unforeseen disaster of some kind, then A.I. isn’t going to just be “like it is not but a lot more”. If it stalls out and stays like it is now or slightly better, then we won’t use so much of it that the water will ever become a real concern!

If this ever does become a real problem, then all bets are off, everywhere. A.I. will be the most important thing that’s ever happened, in that scenario. It will be curing diseases and going to space and will likely be able to invent industrial-scale desalinization and solve the water problem absolutely everywhere for all time, including for agriculture and ranching.

(Spoiler alert: This is also my first cut at the electricity question. “If we’re on an A.I. growth path where this is a problem, then A.I. will solve this problem.” The difference in that case is that the actual needs are much higher as a percentage of total U.S. output, which means building a ton of new capacity, and while we have the tech already to do that, the political and regulatory situation is so bad that there may be no way out, just as A.I. cannot invent faster-than-light travel no matter how smart it gets if such a thing is against the laws of physics.)

I don’t have a particularly snappy way to end this report, because in essence this story is a dog that didn’t bark. But it’s still important to know about, and it’s important, in this anxious age, to be able to let go of a particular anxiety if it turns out not to be necessary. There are plenty of problems with A.I and there are plenty of problems with water usage (did you read that stat about almonds? I can’t remember the last time I even *ate* an almond.), and there’s no reason to create a fake problem that’s about both and not let yourself use the most transformational technology since the internet because you’re worried about wasting water. (In a profound way, A.I. is the internet, they are not in fact separate technologies, but that is also the subject of another post.)

END

Thanks as always for reading my work! I hope you found yourself edified by this missive. Have a wonderful week, Merry Christmas and Happy Holidays to everyone, and I will be back in your inbox next Sunday with something fun!