# The AI water issue is fake [[andymasley-blog|Andy Masley]]'s long piece argues that AI water use is barely a problem at any scale — national, local, or personal — and that the public belief otherwise rests on three things: discomfort with physical resources spent on digital products, failure to internalize how many people use AI, and decontextualized big-gallon numbers reported without comparison. The post explicitly excludes AI's electricity story, which Masley considers a real issue. The framing matters: Masley isn't saying data centers never cause water problems, only that the actual scale of impact does not justify the rhetorical scale of the public conversation. Where impact is real, it's almost always construction sediment rather than operational water use. ## Definitions that change the conversation - **Consumptive** vs **non-consumptive** use — evaporated and gone vs taken out, returned unchanged. - **Direct** vs **indirect** — water in the data center vs water at the power plants feeding it. - **Potable** vs raw freshwater — drinking-quality vs untreated. Two facts that follow: - About 80% of "AI water use" in reports is at the power plant, not the data center. Almost all of the power-plant water is non-consumptive — withdrawn, used, returned. - The average kWh uses 4.35 L of water. A digital clock has 0 direct water cost but 0.2 L/day of indirect cost — every three days the clock accounts for a bottle of water at a nearby power plant. This is just how electricity works. Breakdown of where AI water goes: - ~90% — non-consumptive, indirect, non-potable. Power plants returning water to the source. - ~7% — consumptive, indirect, non-potable. Evaporated at power plants. - ~3% — consumptive, direct, potable. Inside the data center. ## National scale All US data centers (mostly serving the internet, not AI) consumed 200–250 million gallons of freshwater daily in 2023. The US consumes ~132 billion gallons of freshwater daily. Data centers = 0.2% of national freshwater consumption. Onsite water alone was 50M gal/day — 0.04% of national freshwater. AI is about 20% of data-center power use, so AI's national freshwater share is roughly 0.04% including offsite, 0.008% if only onsite. The "0.008% of America's freshwater" comparison Masley returns to: that's about 10.6M gal/day, eight times the water of his hometown (Webster, MA, pop. 16,000). All AI in all US data centers uses as much water as eight small towns. By 2030, with the standard 3× growth forecast for data-center power, the projection is ~150M gal/day onsite for all data centers (0.12% of national freshwater), of which AI is roughly 0.08%. That's 5% of US golf-course water, ~5% of US steel production, or 173 sq mi of irrigated corn farms. ## Local scale Only one US county has documented a data center contributing to rising water costs: Newton County, Georgia, where Meta is one of several large industrial users alongside a pharmaceutical plant and a Rivian factory, with inflation and housing pressures also in the mix. The famous NYT "taps ran dry" headline from a Georgia couple traces to *construction sediment* in groundwater, not data-center operations — the center hadn't started running and doesn't draw from local groundwater. The piece flags this kind of misleading framing repeatedly: the actual problem in those stories is the construction phase of building any large building, not the operating phase of a data center. In Maricopa County (Arizona, the highest-stress water region with the most new data centers), data centers use 0.12% of county water. Golf courses use 3.8%. Data centers generate ~50× more tax revenue per gallon than golf courses. Replacing all Arizona golf courses with data centers at the same water usage would add ~$42B/year in tax revenue, roughly double the state's current total. Masley's local-impact argument: data centers behave like other normal industries, are subject to the same regulations, and where they're built they often *fund* improvements to local water systems via development agreements (The Dalles OR, Council Bluffs IA, Quincy WA, Goodyear AZ, Umatilla OR — the latter returning ~96% of cooling water to farmers). ## Personal scale Average American daily water footprint: 422 gallons. Each AI prompt costs ~2 mL including offsite (~0.3 mL onsite only). So one daily personal water footprint = 800,000 prompts. What other things cost in prompt-equivalents (from public water-footprint data): - Leather shoes: 4,000,000 prompts - Smartphone: 6,400,000 prompts - Jeans: 5,400,000 prompts - T-shirt: 1,300,000 prompts - A single piece of paper: 2,550 prompts - A 400-page book: 1,000,000 prompts In electricity terms (~2 L/kWh at the power plant ≈ 1000 prompts of water per kWh): - PS5 for an hour: 200 prompts - Laptop for an hour: 50 prompts - LED bulb for an hour: 6 prompts - A bath: 5,000 prompts (the bathtub holds 80,000 prompts of water on its own) 10,000 prompts/year = 1/300,000 of personal water footprint. If your annual footprint were a mile, 10,000 prompts is 0.2 inches. ## Potable-water specifics Common objection: AI uses *potable* water, which is more constrained than raw freshwater. Masley's counter: turning raw freshwater into potable is cheap (~$1 per 1,000 gallons in treatment-only cost), water utilities have well-documented economies of scale, and adding a large stable buyer to a public water system in a freshwater-abundant area *funds* the infrastructure that makes treated water cheaper for everyone. The case where this breaks is when raw freshwater itself is scarce — but that's the *freshwater* problem, not the potable-water problem. The "data centers using precious drinking water" framing is inverted: in freshwater-rich regions, more buyers makes drinking water cheaper. In an absurd-but-thorough scenario where 10× AI growth concentrates entirely in the top four data-center counties, with all potable-treatment costs somehow falling exclusively on households (rather than the AI companies or the commercial rate class), per-household water spending in those counties would rise about 3% by 2030. ## Pollution The piece walks through the three mechanisms by which data centers could affect water quality — construction sediment, concentrating existing contaminants by evaporation, treatment chemicals in blowdown. None has produced documented widespread harm. The two prominent stories (Georgia couple with dry taps, Oregon "Amazon causing cancer") were both about other causes: construction sediment in Georgia, and decades of agricultural nitrate pollution and a failing wastewater treatment plant in Oregon. Amazon settled the Oregon suit specifically as one of 17 defendants and explicitly denied wrongdoing in the settlement documents. ## Trade-offs Water cooling is *more* energy-efficient than air cooling (10% lower total power), which means lower CO₂ emissions. In water-abundant regions, water cooling is usually the right environmental choice. Air cooling shows up in deserts because that's where the cost ratio flips. ## Coverage criticism The piece spends substantial space on misleading news coverage: - The 2024 Washington Post "every email = one bottle of water" claim required stacking six worst-case assumptions, including that the data center is in Washington state, that hydroelectric water is "wasted," and that 2024 LLMs were no more efficient than 2020 LLMs. - The Economic Times "Texans showering less because of AI" article was 463M gal over two years = 640K gal/day = 0.005% of Texas water use = equivalent to 1,600 additional residents. - Multiple iterations of the same scary framing across regional outlets. ## Where this lands Sits as a counterweight to [[hold-on-to-your-hardware]] (which is about consumer hardware scarcity, a real downstream effect of the AI build-out) and to the anti-LLM cluster in [[anti-llm-discourse]] more broadly. It's also adjacent to [[ai-bubble-pale-horses]] and [[subprime-ai-crisis]] but argues a narrower point — *if* the buildout proceeds, water specifically isn't the binding constraint or the social-license blocker; if the buildout collapses on its own economics, it does so for unrelated reasons. The piece's stance fits a category not well represented elsewhere in the wiki — pro-AI on a specific narrow technical critique. Masley flags he's not consistently pro-AI overall, but objects to misleading numerical claims diluting more serious AI concerns.