Nathan Howard/AP Photo
Justin Grant moves his cattle from a dry grazing field, July 24, 2021, in Klamath Falls, Oregon.
Since AI models like ChatGPT and Claude became the latest investment fad in Silicon Valley, outside observers have worried about the broader consequences. They consume tons of electricity, they are trained on trillions of original works without their authors’ consent, and if the most unhinged hype guys are to be believed, they will create mass unemployment in every industry any day now.
But for some reason, the water use of these products has become one of the most common criticisms. A slew of articles and videos argue that the water consumption of all the data centers powering AI systems are a threat to the environment.
It is true that data centers use some water. But there is a great deal of missing context. Even in highly water-stressed areas, all data centers combined are a rounding error compared to the real water wasters: farmers, especially of livestock feed.
Here’s a representative piece in The Washington Post. The authors spoke with a research scientist who estimated how much water it takes to use GPT-4, accounting for both direct cooling and the embedded water in all the facilities. One 100-word email generated with this program, they figure, would use up about 519 milliliters of water.
The first problem here is that in much of the country, water is plentiful. To a first approximation, east of the 98th meridian, which passes roughly from the border between North Dakota and Minnesota down to the southern tip of Texas, there is usually enough rain to farm with just natural precipitation. To be fair, there is some irrigated farming around the Mississippi, and in southern Georgia and Florida. In the latter region, as my colleague Gabrielle Gurley covered back in 2022, climate change has led to drought, and thence to some conflicts between farmers.
But as I’ll explain in a moment, because farming requires so much water, by and large in the eastern half of the country there is easily more than enough for cities and industries, so long as they can build proper purification and treatment facilities, of course.
West of the 98th meridian, water is scarce. But here’s where scale becomes important. A liter is just not a good unit for getting your head around total water use. A better one, in this country at least, is the acre-foot. This is a bizarre and old-fashioned unit, but it’s simple to understand: one acre filled one foot deep with water. It’s handy for farmers who calculate their irrigation in terms of inches of water, just like rainfall is measured.
Diverting just 10 percent of Nebraskan irrigation water would be enough for every single American to use GPT-4 about 24 times per day.
For human-scale stuff like showers or drinking, an acre-foot is a lot of water. Do the math, and one acre-foot comes to 1,233,481 liters. But for agriculture, one of these is barely anything. Specific irrigation needs vary tremendously based on local climate, soil conditions, elevation, the type of irrigation system, and so on, but according to the USDA, on average irrigated farms applied about 1.5 acre-feet of water per acre as of 2017. There are 8.6 million irrigated acres in Nebraska alone, using on the order of 12.9 million acre-feet—or about 16 trillion liters—of water per year. Particularly thirsty crops like rice, oranges, almonds, or alfalfa use even more than this average.
This puts the Post piece in some better context. They calculate that training Microsoft’s GPT-3 required 700,000 liters of water—or just over half of one acre-foot. If 10 percent of working Americans were to use GPT-4 every week, that would require 435,235,476 liters per year—or about 353 acre-feet, or about 0.003 percent what Nebraskan farmers use. Or looking through the other end of the telescope, diverting just 10 percent of Nebraskan irrigation water would be enough for every single American to use GPT-4 about 24 times per day.
Alfalfa and other livestock feed crops are particularly objectionable. Californian desert farmers of nuts, fruits, and vegetables argue that their tremendous water use is justifiable because they provide most of the supply of these important, nutritious foods in American refrigerators and pantries, and indeed in much of the world. Despite being obviously biased, the argument does make some sense. But livestock feed is not eaten by humans. It is for cattle, pigs, sheep, and so forth. This means that most of the water used to grow the feed does not end up in food—rather it is consumed as the animal grows, metabolizes, breathes, excretes, and so on.
To be fair, the Post piece admits this implicitly by noting that training GPT-3 used as much water as just under twice what is required for the average American’s beef consumption. In other words, just two people swearing off beef would more than compensate.
By the same token, while it is quite difficult to grow oranges in Pennsylvania, livestock feed can be grown in most of the country. It is more economical to do so in the Southwest only because of the utterly crack-brained legal regime governing most American water use, which is largely based on “first in time, first in right,”—that is, whoever’s great-great-great-grandfather set up their farm first gets dibs on the water—and “use it or lose it.” In practical terms, this means agricultural water out West is preposterously underpriced.
Now, it is possible for data centers to stress municipal water systems, particularly in Western cities where they are already strained. But this is downstream of the fact that it is legally very hard for cities to buy up water rights from farmers, which they could easily afford in any remotely sane world. Almost half of all the irrigation water in the Colorado River Basin goes to cattle feed alone. A modest diversion from this industry would free up more than enough water for Southwest cities and data centers alike for the rest of the century at least.
