The Water Footprint of Food, Compared to AI

A question that keeps coming up: how much water does it take to ask a chatbot something? The worry is reasonable. Data centers are real, they use real water for cooling, and the technology arrived fast enough that most of us never got a chance to form an intuition about its costs. So it is worth answering the question carefully, and then setting the answer next to a cost we rarely think about at all, which is the water embedded in food.

The short version: a single chat session with an AI assistant consumes water on the order of half a litre, while a single beef burger carries with it something closer to 1,500 to 1,800 litres. Both of those figures need unpacking, and one of them rests on a preprint, so neither should be quoted without its caveats. But the size of the gap survives almost any reasonable adjustment, and that is the part worth holding onto.

What "the water footprint of a burger" actually counts

When researchers talk about the water footprint of beef, they are not talking about what comes out of a tap at a slaughterhouse. They are counting everything: the rain and irrigation that grew the feed, the water the animal drank, and the water used to process the meat, summed across the animal's whole life. The standard global accounting puts the average at roughly 15,400 litres per kilogram of beef, and the great majority of that is water embedded in growing feed (Mekonnen & Hoekstra 2012).

That same work splits the total into three colors, and the distinction matters because not all water costs the same. Green water is rain stored in soil that a crop draws on as it grows. Blue water is surface and groundwater, the kind drawn from rivers, lakes, and aquifers that people and ecosystems also depend on. Grey water is the volume needed to dilute the pollution that runs off a farm back to a safe standard. For beef, the average footprint is dominated by green water, with blue and grey making up smaller shares (Mekonnen & Hoekstra 2012). A litre of rain falling on a pasture is not interchangeable with a litre pumped from a depleting aquifer, and a fair comparison should keep that in mind rather than treat all 15,400 litres as equally scarce.

To get from a per-kilogram figure to a burger, you need an assumption about portion size. A common quarter-pound patty is about 113 grams of raw beef, a little over a tenth of a kilogram. Applying the average footprint of about 15,400 litres per kilogram gives very roughly 1,500 to 1,800 litres for the beef in one burger, depending on the exact patty weight and where the cattle were raised. That range is a derivation from the Mekonnen & Hoekstra 2012 average, not a separately measured number, and beef from different systems and regions varies widely around the mean. Treat it as an order-of-magnitude figure, which is all the comparison really needs.

Beef is the high end. Most foods sit far below it. The point of using beef here is not to pick the worst case for effect; it is that beef is the food people most often eat without any sense of its water cost, and it happens to be the one that dwarfs the AI figure most clearly.

What AI actually uses, and why the numbers are soft

Now the other side. The most-cited estimate of AI water use comes from a preprint, which means it has been posted publicly but not yet been through formal peer review. It is worth being explicit about that, because the figure travels through headlines as if it were settled, and it is not.

The preprint reports that training GPT-3 in Microsoft's US data centers directly evaporated about 700,000 litres of freshwater (Li, Yang, Islam & Ren 2023). That number sounds enormous, and in isolation it is a large volume. It is also a one-time cost spread across a model used by enormous numbers of people, and it counts on-site evaporative cooling rather than the full supply chain, so it is not directly comparable to the cradle-to-plate accounting used for food.

For everyday use, the same preprint estimates that a short exchange with a model like GPT-3 consumes roughly 500 millilitres of water, about a standard water bottle, across something like 10 to 50 medium-length responses (Li, Yang, Islam & Ren 2023). The wide range is honest: the figure depends heavily on which data center serves the request, the local climate, the time of year, and how the cooling and electricity are accounted for. A request routed through a cool, efficient facility costs far less water than the same request in a hot region drawing on water-intensive power. So "about half a litre per chat session" is a reasonable rule of thumb from this source, with the understanding that the real number for any given query could be meaningfully higher or lower, and that the underlying study has not been peer reviewed.

It is also fair to AI to note what these numbers do not capture. AI water use is concentrated, growing quickly, and sometimes located in places already short of water, which is a real local concern even when the per-query figure looks small. A half-litre per session is not nothing when multiplied across billions of queries. The case here is not that AI's water use is fake or trivial in aggregate. It is that for an individual deciding where their own water footprint comes from, the per-use comparison is lopsided.

Putting the two side by side

Here is the comparison in one place, with the source and the main caveat for each figure.

A single burger embeds roughly three thousand times the water of a single chat session, give or take, on these figures. You could run a chat session every day for years and not approach the water cost of one beef dinner. Even the full one-time cost of training GPT-3, the headline-grabbing 700,000 litres, is on the order of the water embedded in a few hundred kilograms of beef.

Where this leaves the worry

The diet point is the durable one, and it does not rest on the soft AI preprint at all. The largest assessment of food's environmental impacts, a meta-analysis covering about 38,700 farms, found that moving away from animal products cuts food's land use, emissions, and water demand substantially, with animal products carrying far heavier footprints per unit of nutrition than plant foods (Poore & Nemecek 2018). For most people, the water that flows through their food choices is a much larger lever than the water behind their software.

None of this is an argument to stop worrying about data centers, which deserve scrutiny on water, energy, and where they are built. It is an argument about proportion. If you want to spend less of the world's water, the most effective place to look is not your chat history. It is the part of the menu nobody prices in litres. For the broader picture of how food stacks up on land and carbon, see why animal farming needs so much land, and the full science library collects the studies behind these numbers.