Narrative chapters
The journey of one gallon
Four chapters, four mental models. Move through the story to feel the system before you analyze it.
A first principles, systems thinking, MDA framework masterclass on what water actually does in the U.S. economy. We compare lawns, golf, almonds, oat milk, Hollywood, and data centers on the same gallon for gallon basis, then ask which uses produce the most human value per drop.
In plain English, here is what the verified data says about U.S. water and data centers in 2026.
About 200 to 250 million gallons per day in 2023. The U.S. as a whole uses about 132 billion gallons per day. Source: Andy Masley analysis cited by Pirate Wires.
California almonds consume about 1.59 trillion gallons per year on roughly 1.38 million acres. About 75% of those almonds are exported. Source: USDA NASS.
National numbers are reassuring. Local stress can still be real (Phoenix, Loudoun County wells, Morrow County OR). Open the map below for accuracy.
Most viral water charts show data centers next to two or three cherry picked categories. Here is the whole board. The two largest users are not on most charts at all, and household plumbing leaks alone waste 3.9 times more water than every U.S. data center combined.
Hover a bar for the source and the accounting caveat. Log scale: each gridline is a 10x jump.
This is the caveat almost every viral chart skips. The bars above are not all measuring the same thing, and honest comparison requires saying so out loud.
Thermoelectric power withdraws the most water in America, but returns almost all of it to the river it came from. Big number, small permanent loss.
Evaporated or embedded in a product, so it does not come back. This is the honest metric for data center cooling and for crop transpiration.
Irrigation and turf figures are usually water applied, part of which percolates back to groundwater. Not directly comparable to consumption.
Because of this, the data center bars use consumption while irrigation and turf bars use withdrawal or applied water. That accounting difference flatters agriculture, not data centers. The gap is far too large for the mismatch to reverse the conclusion, but the comparison is approximate and should be quoted that way.
This line has been going around, and it is one of those claims where the answer flips depending on which building you are standing in. Both halves are worth knowing.
Newer closed loop and air cooled designs recirculate their coolant and evaporate nothing. Microsoft began rolling out a zero water cooling design in 2024. At those sites the bathrooms really are the entire on site water draw, so roughly 60,000 gallons a year of staff sanitary water versus about zero for the servers.
Nationally, U.S. data centers consumed 17.4 billion gallons on site in 2023. All of their workers flushing all year comes to roughly 0.42 billion gallons. So the servers out drink the staff by about 41 to 1.
1.6 gal/flush x 3 flushes x 250 workdays = 1,200 gal per worker per year.
~350,000 U.S. data center workers x 1,200 = 0.42B gal/yr.
17.4B gal cooling / 0.42B gal sanitary = 41x.
To make the claim true nationally you would need 14.5 million on site workers.
Verdict: true and increasingly common at the facility level, false as a statement about the national fleet. The real story is that cooling design, not headcount, decides a data center's water bill.
Dripping U.S. plumbing wastes 900B gallons a year, versus 228.8B for the entire data center footprint including power.
43,070B gallons withdrawn for crops and pasture, against the 228.8B gallon total data center footprint.
Direct on site range for hyperscale sites. Even the high end stays under 0.1 percent of national withdrawals.
Sources: Lawrence Berkeley National Laboratory 2024 U.S. Data Center Energy Usage Report (17.4B gal direct, 2028 hyperscale range), USGS Estimated Use of Water in the United States 2020 (thermoelectric, irrigation, public supply, livestock), EPA WaterSense (residential outdoor, household leaks), GCSAA Golf Course Environmental Profile (golf irrigation, 21 percent recycled), USDA NASS (California almonds), FAO AQUASTAT and the UN World Water Development Report (global figures). Global data center estimates are derived by scaling the U.S. figure by the U.S. share of world capacity and are the least certain numbers on this page.
Strip away the politics. A gallon of water can be evaporated, transpired by a plant, drunk by a person, used to cool a server, or wasted on pavement. The honest question is which outcome creates the most human flourishing per drop, today and across time.
Every gallon spent on one use is a gallon not spent on another. There is no free water.
What does the next gallon create? Not the average gallon, the next one at the margin.
And then what? Almonds export value abroad. AI compounds productivity at home.
Small policy nudges in water rights and efficiency can produce outsized system change.
Comparing farms to data centers on dollars alone is a category error, and this site should say so plainly. Farm water produces food. A data center produces zero calories. So we score the same five gallons two ways: what it earns, and what it feeds. The surprise is how much the answer varies within farming.
Edible energy delivered per five gallons of consumptive water.
Five gallons grows about 140 calories of corn or 120 calories of potatoes, roughly a slice of bread. These staples are the most water efficient calorie machines in American agriculture.
The same five gallons routed through alfalfa into beef yields about 3 calories. That is roughly 45 times less food than corn from identical water, and alfalfa is the single largest water user in the Colorado River basin.
Lettuce earns 20 times more per gallon than corn but delivers 5 times fewer calories. Optimizing water for money and optimizing it for food are different problems with different answers.
| Use of five gallons | Calories | Value | Water per acre | What that actually is |
|---|
Method: crop yield times price divided by seasonal consumptive water (evapotranspiration), converted at 325,851 gallons per acre foot. Calories use USDA food composition values for the edible harvested portion. Beef is modeled at roughly 1,800 gallons per pound, the bulk of it irrigated alfalfa and corn feed. Sources: USDA NASS Crop Production and Prices Received, USDA ERS Irrigation and Water Use, USGS 2020 estimated use, FAO crop water requirement guidance. Figures are national averages and vary widely by region, soil, and irrigation method.
A data center turns five gallons into about $132 and no food. Corn turns the same five gallons into about 0.7 cents and a slice of bread. Neither number makes the other wrong. Food has value that dollars per gallon cannot express, and compute has value that calories cannot express.
The useful question is not farms versus data centers. Farming uses roughly 80 percent of consumed U.S. water and data centers about 0.2 percent, so the decisions that actually move the system happen between crops: alfalfa versus corn, flood versus drip, and which acres get planted at all in a drought year.
of consumed U.S. water
of consumed U.S. water
more calories per gallon
of Colorado River water
Four chapters, four mental models. Move through the story to feel the system before you analyze it.
A claim by claim verification ledger with direct source URLs. Plus first principles tables, systems canvas, MDA explorer, lifecycle timelines, Sankey supply chains, an economic multiplier model, and a full source vault.
Ten viral water claims, fact checked with original quotes, our finding, the math behind it, and links to verified primary sources. Filter by verdict or search.
Toggle the lens. Absolute gallons hide value. Per dollar reveals leverage. Per job reveals opportunity cost. Per future dollar reveals second order effects.
USGS 2020 estimated use, EESI 2024, Brian Potter 2025, USDA NASS 2024. Indirect data center water via power generation is shown as a separate stack.
USDA NASS California Almond Acreage Reports, Potter 2024, LBNL 2024 U.S. Data Center Energy Usage Report.
Andy Masley, profiled in Pirate Wires by Blake Dodge and Harris Sockel in December 2025, shows that in 2023 U.S. data centers consumed about 200 to 250 million gallons per day. That is roughly 0.2 percent of total U.S. daily water consumption, about 132 billion gallons. Even if data center power triples by 2030, the projected draw is only about 8 percent of U.S. golf course water and about 1 percent of irrigated corn farming.
Bottles of water is a misleading unit. A pair of leather shoes carries an embedded footprint of about 16,000 bottles, jeans about 21,600, a smartphone about 25,600. The average American consumes roughly 1,600 bottles per day across food and indirect uses.
U.S. daily water that data centers consume
Texas daily water that data centers add
of golf course water if DC power triples by 2030
of irrigated corn water at 2030 projection
Drag the slider for one glass of water and see what it does in each system.
Move the sliders for a day in your life. We compare your daily consumptive water to the average data center user and the almonds you would replace if your water went into a hyperscale cluster.
Drag the slider for quick adjustment, or type any number on the right. There is no upper limit.
Coefficients per service per day: shower 17 gal, flush 1.6 gal, beef serving 600 gal embedded, almond 1.2 gal each, AI prompt 0.0005 gal direct + 0.002 gal indirect, lawn 0.6 gal/sq ft, streaming 0.04 gal/hour. Sources: USGS, USDA, Masley, Construction Physics.
Donella Meadows reminds us that the biggest gains are in mindset and rules, not in adjusting parameters. We model a small parameter shift to make the lesson visceral.
If we shifted a fraction of California almond water to a mix of desalination capacity, data centers, and aquifer recharge, what would happen?
Aquifers are stocks. Outflows above natural recharge create slow disasters with long delays.
Subsidence shows up 10 to 30 years after overpumping. Policy responds to today, not 2055.
California is great at compute and great at almonds. Where is the next dollar best deployed?
22 real U.S. data center clusters with verified MW capacity, all California almond counties (commercial almond production is concentrated in California; trace amounts in AZ, NM, and TX are shown as faint markers and are not statistically significant), the largest lawn and golf metros, and the 7 most stressed U.S. water basins.
Data center sites: Cushman & Wakefield Global Data Center Market 2024, JLL Data Center Outlook 2024, Construction Physics, Microsoft and Google sustainability reports. Water stress: USGS Aquifer Studies, USBR Colorado River Basin, California DWR SGMA reports.
Modern reverse osmosis takes about 3 kWh per cubic meter. With cheap solar that lands near $0.60 per kilogallon, and at scale, far less. Compared to a residential bill of roughly $5 per kilogallon, it would not just be cost competitive, it would be effectively cheaper than current municipal water in many coastal regions.
After Tomas Pueyo, with first principles overlays. Energy cost and membrane learning curves blended.
If a number here changed your mind, the person who did the work deserves a follow. Direct profiles, no algorithmic intermediation.
Based on the sliders you moved in the Personal Water Footprint section above. Updates as you change them.
Pick a claim you have argued about. We will pre-fill a short, sourced message you can send by text, email, or X.
Direct answers to the questions that brought you here. Click any to expand.
At the national scale, no. About 0.2% of U.S. daily water. At the local scale, sometimes yes. Specific places where data centers added measurable stress include parts of Loudoun County VA, Morrow County OR, and some Phoenix wells during droughts. The mainstream framing of an "AI water crisis" does not match the aggregate data.
No. The viral framing requires you to do 1,000 plus prompts to use a 500ml bottle. A single email is more like 1 to 3 prompts. Andy Masley walked through the math in detail at andymasley.substack.com.
Withdrawn water is taken from a source. Consumptive water is not returned (it evaporated, became part of a product, or was contaminated). Power plant cooling withdraws huge volumes but returns most of it. Almonds consumptively use most of their irrigation. This site uses consumptive water for fair comparisons.
Almond trees transpire about 3 to 4 acre-feet per acre per year. California has roughly 1.38 million bearing acres. About 75% of the harvest is exported, so the embedded water leaves the U.S. as nuts. Drip irrigation has improved efficiency but total acreage growth offset most gains.
For water, no. Your personal AI water footprint is negligible (a few drops per day at a few hundred prompts). For electricity and carbon, the answer depends on the grid your provider uses. Push for clean energy procurement rather than limiting use.
Use the Personal Footprint calculator above for personalized actions. The highest leverage items in most U.S. cities are turf rebates (replace lawn), low-flow shower heads, and dual-flush toilet retrofits. At the policy level, support SGMA enforcement in California and aquifer recharge funding.
For coastal regions and certain inland reuse scenarios, yes. Modern reverse osmosis at scale is cheaper than many municipal water systems. The unsolved engineering problem is brine discharge, not energy or cost. See Tomas Pueyo's deep dive.
Because the public conversation right now is "AI is destroying water." That framing is testable. We picked almonds because the comparison is the most surprising at the per-gallon-of-value level. Total U.S. agriculture uses about 80 trillion gallons a year, with corn, alfalfa, and rice being the largest users.
If a term in any chart, claim, or chapter was unclear, the definition is here.
Water removed from a source and not returned. Lost to evaporation, transpiration, or embedded in a product.
Total water taken from a source. Much of it is returned. Larger than consumptive.
One acre covered one foot deep. About 325,851 U.S. gallons. The standard ag water unit.
Total facility energy divided by IT energy. 1.0 is perfect. Modern hyperscale runs 1.1 to 1.4.
Liters of water per kilowatt-hour of IT load. Lower is better.
Cooling towers spray water that evaporates, removing heat. Effective but consumptive.
Coolant in direct contact with chips. Closed loop, low consumptive water. Standard for new AI builds.
California's Sustainable Groundwater Management Act, 2014. Requires basins to halt overdraft by 2042.
Ground level dropping due to aquifer overdraft. Visible in parts of California's Central Valley.
Data centers above ~100,000 sq ft and 25 MW IT load. Examples: AWS, Azure, Google Cloud campuses.
Pressure forces seawater through a membrane that blocks salt. Modern energy intensity ~3 kWh/m³.
Concentrated salt waste from desalination. The unsolved environmental constraint.
Every number on this page is cited. Click to expand each source. Raw data tables are downloadable below.
Architecture, capex, and cooling for hyperscale AI facilities. Pages on power density, liquid cooling, and water reuse informed our value per gallon framing.
The author corrects earlier estimates. Direct site water is much smaller than thermoelectric power water that supplies the grid. We use his corrected figures.
Defines compute density for modern hyperscale sites and grounds our value per gallon math.
Cost curves for reverse osmosis, energy intensity, and the marginal economics of large scale desalination on the U.S. coasts.
Authoritative baseline for U.S. water withdrawals by category.
Acreage, yield, and farm gate value used for the almond gallons math.
Energy demand projections and PUE trends.
Sector value added used for data center GDP attribution.
Residential outdoor water use baseline.
Average irrigation per course and aggregate U.S. consumption.
Profile of Andy Masley's research showing U.S. data centers consumed about 200 to 250 million gallons per day in 2023, which is roughly 0.2 percent of total U.S. daily water consumption. Even if data center power triples by 2030, the projected draw is only 8 percent of U.S. golf course water and about 1 percent of irrigated corn farming water. Texas data centers add about 0.005 percent to state daily water demand. The article also dismantles the "bottles of water" framing, comparing it to a pair of leather shoes at 16,000 bottles, jeans at 21,600 bottles, and a smartphone at 25,600 bottles. We use these figures to ground the data center side of our value per gallon analysis.
| Category | Annual gallons | Annual value | $/gallon | Export share |
|---|
Built for learning. Verify the numbers, then teach someone else.