Rain Aside, Animal Agriculture Still Uses 38 to 250 Times More Water Than AI

Comparisons of AI and animal agriculture’s water use are often dismissed on a technicality: most of livestock’s water is rain that would have fallen on the land anyway. So we set rain aside and counted only the water drawn from rivers, lakes and aquifers. AI water use is still a drop in the bucket compared to animal agriculture.

Note. Each drop is AI’s highest projected water use in 2030. Animal agriculture counts river and groundwater only; rainwater is excluded.

Last year, we compared the water footprint of AI with that of the cattle industry. The gap was enormous, but the comparison lacked a key point of nuance. Most of animal agriculture’s water footprint is rain naturally falling on fields and pasture, while data centres’ water is piped in. Critics of these comparisons are right that the two are not the same kind of water, and our previous analyses lumped them together.

We are not the first to take this objection seriously. Björn Ólafsson re-ran his food-versus-AI comparison using only surface and groundwater and found that a year’s worth of AI prompts (10,000) uses, at most, about as much as 2 kilos of tofu. Meanwhile, over three years of AI prompts — using the highest estimate of AI water use — still would not amount to the water needed for a single kilogram of beef. Similarly, FoodFacts.org compared livestock’s river and groundwater use against a 2028 projection for AI and found a gap on the order of 200x. Against the same projection, our three livestock estimates give gaps of roughly 140 to 250 times.

Below, we split animal agriculture’s water footprint into green, blue and grey; compare three independent estimates of its river and groundwater use; and test them against the most demanding AI projection available. We deliberately set the lowest estimate of animal agriculture’s blue water use against the highest projection of AI’s water use we could find, to give the smallest gap the evidence allows. The true gap is likely much larger because the lowest animal agriculture estimate is also the oldest, and the most demanding AI figure seemingly includes non-blue water.

Researchers divide agricultural water into three types. Green water is rain stored in the soil and taken up by crops and pasture, while blue water is drawn from rivers, lakes and aquifers for irrigation. Grey water, in the context of agriculture, measures pollution. It is the volume of freshwater needed to dilute pollutants (e.g. fertiliser runoff, pesticides) to acceptable levels. Blue water is the kind that competes with other uses, and it is the kind that data centres often draw from the same supplies as local communities.

Most of animal agriculture’s water footprint is green (rainwater). A widely cited global estimate puts the water footprint of animal farming at 2,422 billion cubic metres a year, of which 87.2% is green, 6.2% blue and 6.6% grey, with 98% of the total going into animal feed. An independent model reached a similar split, finding that 94% of the 4,387 billion cubic metres used to produce animal feed each year is green. 

Since 1 billion cubic metres equals 1 trillion litres, the two models put animal agriculture’s green water use at roughly 2,100 to 4,100 trillion litres a year.

Cut out rain and we are left with three global estimates for animal agriculture’s water use. Mekonnen and Hoekstra imply about 150 trillion litres of blue water a year (our calculation from their published total and shares), and Heinke and colleagues estimate 264 trillion litres for feed production alone. The most recent estimate, from researchers at the UN Food and Agriculture Organization and the University of New Hampshire, finds that livestock systems consume 228 trillion litres (range: 200–280 trillion litres) of blue water a year. The same study found that feed accounts for more than 90% of animal agriculture’s water withdrawals. 

The IEA estimates that data centres worldwide consume around 0.56 trillion litres of water a year, which could rise to around 1.2 trillion litres by 2030. That covers every data centre, not only AI. In 2023, about two-thirds of it was linked to energy supply and electricity generation, about a quarter to on-site cooling, and the rest to manufacturing chips. Morgan Stanley projects that AI data centres alone could use about 1.1 trillion litres a year by 2028 (range: 0.6–1.5), counting cooling, electricity and chips.

The highest projection we have found comes from the United Nations University, which puts the water footprint of data centres’ electricity at 9.3 trillion litres a year by 2030. The same report includes a projection where AI’s share of data centre electricity could double from about 20% in 2025 to 40% by 2030, giving AI a water footprint of about 3.7 trillion litres. Unlike the IEA figure, this covers only the water used to generate data centres’ electricity, not water used on site for cooling or in manufacturing chips. The IEA’s projections put those at roughly 0.5 trillion litres across all data centres in 2030. Applying the same 40% AI share gives about 0.19 trillion litres, bringing AI’s total to about 3.9 trillion litres. 

The UN figure likely attributes an outsized footprint to AI. It is a broad footprint of electricity generation, and the report notes that hydropower-dominated grids have water intensities more than double the global average and that bioenergy’s water footprint is more than 30 times that of coal. That suggests it counts water such as evaporation from reservoirs and the water used to grow energy crops, not only water drawn from rivers and aquifers.

The gap between animal agriculture and AI water use varies significantly in scale, depending on the estimates or projections chosen. But the pattern always holds: animal agriculture demands much, much more water. 

Against the IEA’s estimate for all data centres today, animal agriculture’s blue water use is roughly 270 to 470 times larger. The UN University’s projection for AI alone in 2030 is higher than the all-data-centres IEA figure, despite AI covering only a portion of the sector, because it looks further ahead and counts water more broadly. Even against that figure, animal agriculture still uses roughly 38 to 67 times more blue water, once on-site cooling and chip manufacturing are added.

Animal agriculture uses 38–250 times more water than AI

Sources: Mekonnen & Hoekstra (2012); Heinke et al. (2020); Wisser et al. (2024); IEA (2025); UNU-INWEH (2026); Morgan Stanley (2025), via YourStory.

The strongest remaining argument is that AI’s water use, though small globally, is concentrated in already water-scarce places. There is something to this. In the US, 32% of existing data centres sit in areas of high or extremely high water stress, and so do about two-thirds of those built or in development since 2022.

Animal feed has the same problem. Of 393 highly stressed small river basins, over a third (140) draw more than a fifth of all their water for animal feed, and in 23 of them, feed accounts for more than half of water draws.

The same study adds an important caveat. Removing all livestock water use would lower the stress category of 135 of the 673 stressed basins it analysed, mostly in North America, Europe and East Asia, but would change little in the most stressed regions of Central Asia, South Asia, the Middle East, North Africa and Central America. There, domestic, industrial and food-crop irrigation also drive the stress. This means that animal agriculture is a major contributor to local water stress, but not the only one.

Setting aside rainwater narrows the gap between animal agriculture and AI substantially. But even generous estimates do not come close to closing it. On every major estimate, animal agriculture consumes dozens of times more river and groundwater than even the most demanding projection of AI’s footprint at the end of the decade, and hundreds of times more than all data centres today.

Scrutiny of where data centres are built and how they are cooled is legitimate. But if the concern is freshwater, AI is a drop in the bucket compared to animal farming.