10 AI Water Facts

@distilledscience

562.2K followers · usually 16.7K views a post

Views

85.1K

Likes

11.3K

Comments

262

1.0K

Why it worked

The video addresses a highly debated and polarizing topic (AI's environmental impact) with a structured, fact-based approach, using clear visuals and citations to build credibility. By debunking common myths and providing nuanced information, it appeals to a broad audience interested in AI and environmental sustainability.

Summary

The video debunks common misconceptions about AI's water footprint by presenting 10 nuanced facts with citations. It clarifies the difference between water withdrawal and consumption, explains various types of water footprints (green, blue, gray), and highlights the significant water usage of data centers, including both direct consumption and indirect use for power generation. The video also discusses future projections and potential solutions like closed-loop cooling systems.

Structure

1Introduction to AI water use nuance
2Distinguishing water withdrawal vs. consumption
3Defining green, blue, and gray water
4Data center water consumption statistics
5AI prompt water usage estimates
6Projected growth in AI data center water use
7The role of hydropower and accounting methods
8

Original caption

AI Water use is weirdly polarizing: The people who hate AI like to talk about how it's destroying the environment, often quoting the one bottle per prompt statistic, and the AI supporters tend to write the whole thing off as nothing compared to real water sucks, like beef or corn. Or golf. But, as per usual, the issue NEEDS NUANCE! A normal text prompt can have a tiny direct cooling footprint while billions of requests, expanding infrastructure and electricity generation create a much larger aggregate footprint. National totals can also look small while one facility creates serious pressure on a particular town, aquifer or water system. The numbers change depending on whether we count withdrawal or consumption; direct cooling or electricity; blue water or rain; averages or peaks; measurements or projections. It’s what lets each side with apparently contradictory claims each sound credible. Pay attention: Google's use of median vs. mean is tricky. Two precision notes for this video: The 0.04% national comparison is derived by dividing LBNL’s modeled 2023 data-center consumption by an older national blue-water-footprint estimate Microsoft’s paper says 31–52% less blue water in its abstract, while its main results table reports 31–48% under grid electricity. The table is the safer figure. Morgan Stanley’s 1.068-trillion-liter number is a 2028 base-case projection for global AI-data-center cooling and electricity—not a measurement, withdrawal estimate or all-data-center total. The point isn’t that AI is harmless or uniquely catastrophic. It’s that we need comparable, facility-level disclosure with honest boundaries. Peer-reviewed sources and DOIs: Li et al.: DOI 10.1145/3724499 Mekonnen & Hoekstra: DOI 10.5194/hess-15-1577-2011 Rushforth & Ruddell: DOI 10.5194/hess-22-3007-2018 Lei et al.: DOI 10.1016/j.resconrec.2025.108310 Alissa et al.: DOI 10.1038/s41586-025-08832-3 Additional sources: LBNL-2001637; arXiv:2607.02531; Google’s 2025 Gemini environmental analysis; DOE FEMP; USDA EIB-99; Morgan Stanley projection reported by *The Economic Times*. #AI #Water #DataCenters #Sustainability

More from @distilledscience

Their other posts in the index, biggest breakout first.

Transcript

Here's 10 facts about AI water use to give the issue the nuance it deserves, with citations. One, there's a big difference between water withdrawal and water consumption. Water can be then returned to the water system and able to be reused like most power plants do. About 3/4th of data center water is consumed, but consumption usually evaporates, and can no longer be drinking or for irrigation and other purposes. Headlines will often confuse this. Two, Green water is rain stored in usable by a lot of agriculture and grazing. Blue water is from reservoirs, or ground water, and that's what people fight about, used for drinking water, irrigation, and data centers and a lot more. Gray water is a term for how much water you would need to dilute out a bunch of pollutants. Three, A 2025 Google report estimates that the median prompt just 0.26 milliliters of water. About five drops. Four, A viral study estimated that 10 to 50 prompts around a bottle of water, or 500 milliliters, including the power production for it. But that's not one modern prompt, and a single prompt often leads to confusion. Five, US data centers consumed roughly 66 billion liters of water in 2023. But that's directly, models, electricity they consumed had the power consuming another 800 billion liters of water. Six, Hydropower matters. Counting or not counting water from reservoirs is almost a doubling factor in numbers, but most people don't calculate it. Seven, data centers account for roughly 0.04% of the United States' freshwater compared to 80% for agriculture. It's still error, but not where they're being. Eight, Growth matters, predicted that by 2028, total water consumption of data centers plus the power that they consume is going to grow by a factor of 11 since 2024, reaching over a trillion liters annually. Nine, Cooling is necessary, but evaporative cooling isn't. There are now closed-loop systems, don't use water for the cooling beyond the initial filling of the tanks. But also take power to operate, which itself takes water back at the plant, a whole calculus. and climate-dependent. Ten, Better designs can win twice. Microsoft put out a study of a closed-loop water cooling system modeled the whole thing at 15-20% less energy consumption and 31-52% less blue water. So hopefully we'll move more in that direction. As always, needs Stay curious.