AI water use and the boundaries of resource accounting

Source: Why is Everyone So Wrong About AI Water Use??, Hank Green, 23:59, uploaded 2025-12-08, playlist index 12.

Hank Green opens with two figures that seem impossible to reconcile. Sam Altman has said that an average ChatGPT query uses about 0.000085 gallons of water, roughly one fifteenth of a teaspoon. A Morgan Stanley projection puts annual water use for AI data-centre cooling and electricity generation at about 1 trillion litres by 2028, an elevenfold increase from its 2024 estimate. Green says both figures can be true as reported because they draw the boundary around the system in different places. The small number describes one visible interaction. The large number tries to account for the infrastructure that makes millions of interactions possible.

The underlying figures remain difficult to inspect. OpenAI has not published the sources or method behind Altman’s number, and the video does not link the Morgan Stanley projection. Green’s point is therefore about the structure of the calculation as much as its result. Water accounting becomes misleading when a single number hides the choices that produced it.

The water inside a data centre

Computer chips heat up during processing and lose efficiency and lifespan when they get too hot. Data centres therefore need cooling, and many use evaporative systems that turn clean water into vapour. The vapour carries heat away from the equipment. Some facilities recycle water or use non-potable water from a sewage-treatment plant, yet many still rely on fresh municipal water delivered through the same civic system that supplies households.

That water does not vanish from the planet. It evaporates and falls elsewhere. Green still counts it as used because the water leaves the local system, and because producing drinkable municipal water and moving it through pipes costs money and consumes treatment capacity. A city cannot send an unlimited volume through the network simply because the water will later return as rain.

The first figure also leaves out what happens inside a query. A complex prompt can cause a reasoning model to run several hidden queries, evaluate intermediate answers, search the internet, and continue until it produces a response. One user-visible query may therefore represent many machine queries. Green guesses that this could multiply the per-query estimate, although OpenAI does not publish the information needed to calculate the factor. Even after that multiplication, he still treats the result as roughly a teaspoon of water. The larger omission lies elsewhere in the lifecycle.

Training as a missing part of the lifecycle

Training continues while people use a model. OpenAI and its competitors build newer versions on large GPU clusters that run for weeks or months and consume electricity and cooling water. A conversation with a model depends on that training, so a serious lifecycle calculation needs to distribute some share of the training footprint across the queries that use the model.

The allocation has several defensible boundaries. One could count the training of the models assembled into GPT-5. One could include earlier training that made those models possible. One could also include models that are still in development because the companies would not be building them without the demand for a larger product. Each choice changes the answer. Green reports a University of California estimate that training can account for about half of an AI model’s total resource use, though the video does not identify the campus, study, or method behind that estimate.

Training water also raises a timing problem. Once a model has been trained, that water has already been consumed. More use spreads the past cost across more queries, while the companies keep training new models and expanding the infrastructure for a future that may or may not arrive. Microsoft, Google, and Amazon together spend more than $100 billion a year on new data centres, according to Green. The scale of that buildout belongs to the water discussion because it represents resource demand for systems beyond the model a person is querying today.

Power plants and the large estimate

The biggest figures often add the water used to generate the electricity that data centres consume. In a thermoelectric power plant, heat turns water into steam, the steam drives a turbine, and a cooling system condenses it again after it passes through the turbine. Coal, natural-gas, and nuclear plants use versions of this process. Green cites the US Geological Survey for the claim that electricity generation accounts for 40 per cent of all freshwater withdrawals in the United States.

Power-plant water has a different path from municipal water. Plants usually draw from a river, lake, or ocean and return much of it. Green adds a later correction that roughly 2 to 3 per cent is lost through evaporation. The water that flows through the plant still carries environmental costs. Waste heat can damage aquatic ecosystems, and even a one- or two-degree change in a river can cause harm. Those effects need regulation even when the water returns to its source.

The distinction matters because a large AI-water estimate may assign the data centre a share of all the water moving through the power plant, including the water that returns. That method makes the footprint look larger than a calculation based on evaporation alone. Green finds the choice misleading when the calculation treats water drawn from a river in the same way as drinking water pulled from a municipal treatment plant. The difference does not make either category harmless. It changes what kind of pressure the use places on a place.

Place, quality, and the local water budget

Every watershed has a finite hydrological budget. A river, aquifer, reservoir, or lake can supply only so much water before its flow, temperature, or ecology suffers. A data centre that draws from a lake may leave household taps untouched while drawing from the same local supply. In a fully allocated watershed, that new demand competes with existing uses. A different place may have room.

The type of water can matter as much as the volume. Manufacturing the chips in an AI data centre uses a much smaller amount of water than cooling the facility, according to Green, yet chip fabrication needs ultra-pure water with almost no impurities. Producing it takes more energy than producing drinking water. A water-quality calculation therefore needs to track what the water is fit for and how difficult it is to prepare, rather than treating every gallon as interchangeable.

Cooling water also leaves a treatment problem. Evaporation removes water while minerals remain behind. Those minerals build up in the cooling system and need to be flushed out. The resulting water can become acidic and may need special treatment before it enters a river. Green uses this to show why a single volume figure loses information about the work required to process the water and the place where the waste appears.

The response to a data centre depends on that context. A desert facility might use non-potable water from a treatment plant, recycle its cooling water, or switch to air cooling. Air cooling can consume more electricity, which may be manageable in a place with abundant solar power. The same technical choice carries a different consequence in a desert and in the Pacific Northwest because water is difficult to move between regions.

Corn, ethanol, and the scale of what feels normal

Green then compares AI with an older industrial use that receives less attention. Corn is one of the thirstiest major US crops. He cites the US Department of Agriculture for an estimate of about 20 trillion gallons of water a year for US corn production, against an estimated 260 billion gallons for all AI data centres worldwide. On those figures, US corn uses nearly eighty times as much water as the world’s AI servers.

The comparison becomes stranger when Green asks what the corn is for. About one per cent is eaten by people, he says. Much of the rest feeds livestock, while about 40 per cent becomes fuel. An acre of corn can use up to about one million gallons of irrigation water and produce roughly 500 gallons of ethanol, which gives each gallon of ethanol an irrigation footprint of about 1,500 gallons before processing. The resulting water use reaches tens of trillions of gallons for fuel that moves cars and trucks.

Corn irrigation usually draws from agricultural supplies rather than municipal drinking water. Green still treats it as comparable to the water counted in power-plant estimates because both calculations concern water withdrawn and used for an industrial purpose. Lawn irrigation supplies the clearer municipal example. Cities use trillions of gallons of treated water on grass, even though the water does not need to pass through wastewater treatment afterwards. The comparison does not excuse AI water demand. It shows that public ideas about waste follow habit as much as volume.

Water is local, power is the larger worry

Green’s conclusion on water stays conditional. Some areas sit against their hydrological budgets and cannot add new uses. Building a water-cooled data centre in Tucson, he says, is a bad idea. Other places have room, and resource managers can reduce the pressure with recycled or non-potable water, different cooling systems, and local planning. Water use becomes a serious problem in particular places even if the global total remains smaller than familiar agricultural uses.

His larger concern is electricity. The expected increase in power demand looks much larger relative to existing infrastructure, which would put pressure on carbon budgets and household bills. River health and electricity prices measure different harms, so one cannot replace the other in a single score. Green gives his own priority to power because he expects its political and environmental effects to spread further.

He is also unsure that the planned buildout will happen. The industry is spending as if more compute will produce the future its investors imagine. If that demand fails to appear, companies and markets may absorb a huge investment in Nvidia chips and data centres without getting the promised gains. Green worries that a small group of wealthy decision-makers is placing a large part of the economy on a future that may arrive later than promised or in a different form.

That uncertainty returns the video to its opening problem. A resource estimate already depends on choices about what belongs to the system, and a forecast adds another layer of uncertainty about how much of the system will exist. Green cannot predict the future and does not claim expertise in every part of water accounting. His master’s degree in environmental studies predates the current AI industry, and he expects specialists to find gaps in his explanation. He treats that limit as a reason to take expertise seriously rather than as a reason to accept whichever clean number travels furthest online.

The durable claim is narrower than either side of the AI water argument. AI data centres will use substantial water, with the local consequences determined by the source, the watershed, the cooling system, and the treatment path. Query-only numbers can hide training and infrastructure. Maximal estimates can count water that flows through a power plant and returns to its source. Both choices can produce a persuasive figure while leaving out the part of the lifecycle that changes the judgement.

Sources named in the video

The video names Sam Altman’s per-query estimate, a Morgan Stanley projection, the US Geological Survey, the US Department of Agriculture, and a University of California estimate about training resource use. It does not provide links or full citations for those sources, so the figures above remain Hank Green’s reported claims rather than independently verified results. The video also mentions Ground News in a sponsorship segment, which is omitted here because it does not support the water-use argument.