Dairy milk vs AI images: Which uses more water?

Can one glass of dairy milk really use as much water as thousands of AI-generated images?

It is a striking comparison—and exactly the kind of claim that can travel faster than its assumptions. Dairy has a large water footprint. Artificial intelligence also uses water through data-centre cooling and electricity generation. But turning those two facts into one viral ratio is not scientifically straightforward.

The honest answer

A widely cited global assessment estimated the average water footprint of cow’s milk at about 1,000 cubic metres per tonne—roughly 1,000 litres of water per litre of milk. On that basis, a 250 ml glass represents about 250 litres.

However, that total combines three different categories: green water, blue water and grey water. Meanwhile, there is no universal, independently audited “litres of water per AI image” number covering all models, hardware, data centres and electricity grids.

So a precise statement such as “one glass of milk equals 20,000 AI images” should not be presented as a settled fact.

What the dairy number includes

The Water Footprint Network assessment distinguishes:

• Green water: rainwater stored in soil and used to grow pasture and feed.
• Blue water: surface and groundwater consumed through irrigation, drinking and farm operations.
• Grey water: a modelled volume required to dilute pollution to water-quality standards.

The study found a global average of around 1,000 cubic metres per tonne of cow’s milk and reported that 98% of animal production’s water footprint was associated with feed. Read the full report.

This is not the same as saying a dairy farmer pours 250 litres of tap water into the ground for each glass. It is a supply-chain footprint, much of it rainwater, and it varies substantially by country, feed source and production system.

Where AI uses water

AI’s water use can occur directly when water cools servers and indirectly when power stations use water to generate electricity. Manufacturing chips and building data centres add further impacts that per-prompt estimates often omit.

Image generation also requires very different amounts of computation depending on the model, resolution, number of sampling steps, hardware and whether several images are generated before one is kept.

What we know about image-generation energy

Research measuring dozens of machine-learning models found that image generation was among the most energy-intensive inference tasks studied, with large variation between models. The study reported up to about 0.012 kWh per generated image in its test conditions. Read the model-energy study.

Energy is not water. Converting electricity into a water figure requires the local grid’s water intensity, the data centre’s cooling system, climate and operational boundaries. That is why a clean-looking global ratio can hide enormous uncertainty.

Why viral comparisons go wrong

• They compare dairy’s full supply-chain footprint with only the operational water used by an AI request.
• They count rainwater on one side but not infrastructure or manufacturing on the other.
• They treat every model and data centre as equally efficient.
• They present an estimate as a physical measurement.
• They ignore location, even though water scarcity is local.

Does that let dairy off the hook?

No. Dairy production requires land, feed, water and repeated reproduction of cows or buffaloes. Calves are separated because the milk is being sold for human consumption, and the animals eventually leave the dairy system for slaughter, abandonment or other forms of exploitation.

Even if an AI comparison is imprecise, choosing oat, soy, millet, coconut or other plant milk avoids participating in that reproductive system. Environmental comparisons are useful, but animals are not merely units in a footprint calculation.

Is AI harmless?

Also no. Data centres can place pressure on local water and electricity systems, and total demand matters even when one request is efficient. Technology companies need transparent reporting that separates direct and indirect water, training and inference, and operational and supply-chain impacts.

Google’s 2026 environmental reporting, for example, publishes company-level information on energy, resources and AI, but company totals still cannot tell us the exact footprint of every generated image. See Google’s report.

How to communicate the comparison responsibly

A defensible line is: “A global assessment puts the total water footprint of a 250 ml glass of dairy milk at roughly 250 litres, mostly linked to feed. AI image water use varies so widely—and is disclosed so poorly—that no single image count is reliable.”

That is less sensational than a five-digit ratio. It is also more truthful.

The bottom line

Dairy milk carries a substantial water footprint, and plant milks generally offer a way to step away from animal exploitation. AI has a real and growing environmental footprint too. But the claim that one glass of dairy equals a precise number of AI images cannot currently be verified without naming the model, data centre, electricity source and accounting method.

Good advocacy does not need a fragile statistic. The ethical case for choosing plant milk remains strong without pretending uncertain numbers are settled science.