Following AI
Upstream

Most data-center water is upstream

Data-center water comes in two buckets, and the headline figure usually shows only the smaller one.

• Direct water cools the building, through cooling towers or chip-level loops. This is the number operators quote.
• Indirect water cooled the power plant that supplied the electricity, and in most accounting it is the larger bucket, frequently more than 80% of the real total.¹
• The density jump is AI, not a general trend. Ordinary server racks have held near 5 to 9 kilowatts for years; the leap to 120 and beyond is accelerated computing, GPUs packed for AI and run flat out.

Nothing about the water is special to AI; the density is, and a flagship rack is not the fleet average. The per-query figures that circulate are real but not comparable: an early estimate pegged a short ChatGPT session near half a litre, while Google later put a median Gemini prompt at 0.26 mL, five drops.^1,2 One counts the full system, the other only what evaporates on site at the most favorable location.

One more figure routinely gets quoted out of scope.

• Training is one-time. Training GPT-3 used about 700,000 litres on site, the figure that goes viral, but it is paid once and then spread across every query the model serves.¹

Data centers vs golf courses: a comparison that hides the constraint

A comparison of data centers to golf reveals a directional truth: US golf irrigation really does use far more water than data centers do.³ It is still hard to take much comfort from it, because the comfort comes from setting national golf irrigation against the smallest slice of data-center water use.

Count only the direct cooling and golf wins by thirty or forty to one. Add the grid-power water that data centers pull upstream, the larger bucket, and golf still wins, by something like five to nine.⁴ Google’s entire 2023 water use is about 150 of the country’s 16,000 courses.⁵ The comparison is directionally true, and beside the point.

• It averages a local problem into a national one. A gallon spread thin across sixteen thousand courses, many in rainy states, is not a crisis; a five-million-gallon-a-day cluster dropped into Phoenix in August is.
• It sets an annual total against a daily peak. The worry was never how much water AI uses in a year nationwide; it is how much a single cluster draws from one stressed basin on the hottest day.

Consumed water leaves the basin

Withdrawal is pipe demand. Consumption is basin loss. For data centers, the difference is mostly cooling design: evaporative systems send water into the air, while closed loops and dry cooling return most of it as discharge.

The water that returns is not clean. It carries concentrated minerals, treatment chemicals and heat. And the source matters less than people assume: reclaimed or graywater lowers the potable draw, but evaporation is a net loss to the basin no matter where the water started.

• Evaporated water leaves the basin. It rises as vapour and joins the atmosphere, where it stays a mean of eight to ten days before falling again as rain or snow.^12,15
• Almost none returns nearby. Only about one part in sixty, ~1.7%, comes back down within fifty kilometres of where it left.¹³ The rest rides the prevailing winds, often hundreds and sometimes thousands of kilometres,¹⁴ raining out over a different basin, state, or country.
• The water is exported, not destroyed. It rarely comes home, which is why consumption is a real subtraction from the local water budget anywhere it happens.

New cooling tech cuts on-site water, not grid water

Next-generation cooling changes the site-level water story. It can push on-site evaporation toward zero, but the upstream water depends on the extra electricity the system needs and how that electricity is made.

• Evaporative cooling spent water to save energy.
• Air cooling spent energy to save water.
• Chip-level liquid cooling can be both at once, water-light and energy-light, though it only cuts on-site water if the heat-rejection loop is built not to evaporate.

Microsoft says it has applied that standard to everything it has drawn since August 2024, with a fleet rollout from late 2027. The loop is filled once at construction and never evaporates, though the design still draws some water for other uses.⁶

The next kilowatt-hour sets the water bill

Once chip-level cooling zeroes the on-site draw, the footprint simply becomes the water cost of the electricity, and that cost ranges over an order of magnitude depending on the source.⁷

• Carbon-clean is not water-clean. Nuclear is among the lowest-carbon options and among the thirstiest.
• Only wind and solar score low on both axes, at least operationally, which means it matters which number you are optimizing.
• The clean option is the intermittent one. Wind and solar are water-light but not firm, so serving must-run AI load on them takes storage and flexible operation; skip that, and the round-the-clock gap gets filled by thirsty gas or thirsty nuclear.
• One firm source is also water-light: the ground. Next-generation geothermal runs around the clock at low carbon, and closed-loop designs recirculate their water; it is early but scaling fast, and some analyses suggest that, sited strategically, it could power the entire US data-center build-out.^30,31

Bitcoin shows why per-unit math breaks

Bitcoin is the useful baseline, the same “is this compute worth the resources” debate a few years early. US data centers have now passed it, with AI driving the climb. US Bitcoin mining draws around 53 terawatt-hours a year; all US data centers draw about 176, roughly three to three and a half times as much^9,10, and climbing far faster.

The two loads behave differently, and that matters more than the totals.

• Bitcoin can switch off in seconds. In principle it chases stranded, flared or curtailed power, which holds its marginal footprint down, though in practice many miners run as baseload.
• AI is far less curtailable. Inference-heavy load is firm and commercially must-run, so it is the thing that forces new capacity to be built.
• Mining is mostly air-cooled, so its direct water is small even where its indirect water is large; data centers lean on water cooling, so both buckets fill.

The per-transaction trap

The sharpest-looking number in this debate is also the most misleading. A single Bitcoin transaction carries roughly 1,335 kilowatt-hours and something like sixteen thousand litres of water, a swimming pool.^10,11 A single AI query runs about a third of a watt-hour and a fraction of a millilitre.¹ Side by side, the transaction looks four to five million times heavier on energy.

Visa and per-query water figures carry the most dispute; treat them as order-of-magnitude, not precise, and the Bitcoin per-transaction figures are a snapshot of a moving index, not a constant. The ranking is what holds.

Drawn to scale, the gap is hard to even fit on a page. Each step to the right below is ten times the energy of the step before it.

That comparison is an argument wearing the costume of a measurement.¹⁷

• Bitcoin’s energy is a fixed security budget, not a per-transaction cost. Miners hash continuously for the block reward whether a block carries one transaction or three thousand.
• So the per-transaction figure is a constant over a capped throughput of about seven transactions a second, which means adding one more transaction is nearly free.
• An AI query is closer to a true marginal cost, because inference scales with use. Hold that distinction straight and the ranking can invert: at the margin, one more AI query can cost more energy than one more Bitcoin transaction.
• The trajectories point opposite ways. Bitcoin’s per-unit cost is the system working as designed, and it rises with the coin price. AI’s is incidental, and it falls with every efficiency gain.

The same trap runs on energy. The widely repeated line that a ChatGPT query uses about three watt-hours, ten times a Google search,²⁷ traces to an early estimate and a 2009 search figure; a typical query today is closer to 0.3 Wh, comparable to a search, and falling.²⁸

Ethereum settled the point by example. The same ledger, moved from proof of work to proof of stake, cut its per-transaction energy by 99.96% in a single night.¹⁶

Seven misleading claims come from the same few swaps

Most contradictory headlines about AI and water come from the same handful of swaps, each one debunked in a section above. None of the underlying numbers is fake; each claim quietly changes the boundary, the baseline, or the metric, and the disagreement is manufactured downstream.

1. 01“Every AI prompt uses a bottle of water.”

   Seen in: the Washington Post’s “a bottle of water per email” and the wave of 2024 coverage and videos that ran with it.²¹ Swap: a worst-case, high-energy estimate for one prompt, plus the off-site power-plant water, reported as the routine on-site cost. Correct read: the on-site water for a typical query is a fraction of a millilitre; the bottle figure folds in grid-power water and an energy assumption far above a normal prompt.^2,1 Addressed in: §01.

2. 02“Training one model boils a swimming pool.”

   Seen in: 2023 coverage that training GPT-3 used 700,000 litres, “enough to fill a nuclear reactor’s cooling tower.”²⁶ Swap: a one-time training cost, paid once for the whole model, read as a recurring or per-use cost. Correct read: training water is spent once and spread across the billions of queries the model then serves, so the share behind any single query is tiny.¹ Addressed in: §01.

3. 03“Golf uses more water, so the panic is overblown.”

   Seen in: the golf comparison repeated in the June 2026 Atlantic piece.³ Swap: direct data-center cooling, the smallest bucket, set against golf’s largest, total national irrigation, then averaged across the country. Correct read: a national annual total does not answer a local, seasonal constraint, and the gap closes once the grid-power water is added back.⁴ Addressed in: §02.

4. 04“A data center withdrawing millions of gallons is consuming them.”

   Seen in: reporting that quotes a facility’s peak or permitted draw, or its withdrawal, as steady consumption.²² Swap: peak infrastructure capacity, or water withdrawn and largely returned, treated as water consumed and gone. Correct read: consumption is the real basin subtraction; withdrawal and peak capacity can overstate it several-fold.¹ Addressed in: §03.

5. 05“Recycled water and ‘water positive’ pledges fix it.”

   Seen in: hyperscaler commitments to be “water positive” by 2030, replenishing more water than they consume.²³ Swap: replenishment in other basins, or a switch to reclaimed water, substituted for consumption in the basin the data center actually sits in. Correct read: source substitution lowers the potable draw, but evaporation is still a net loss to the local basin, and the pledges sit alongside fast-rising absolute use.²⁴ Addressed in: §03–§04.

6. 06Assumption: “Clean power is water-clean power.”

   Seen in: the push to run AI on low-carbon “firm” power, especially the wave of nuclear deals.²⁵ Swap: the carbon ledger stands in for the water ledger. Correct read: nuclear is among the lowest-carbon sources and among the thirstiest; only wind and solar run low on both, and only operationally.⁷ Addressed in: §05.

7. 07“An AI query uses ten times the energy of a Google search.”

   Seen in: the widely repeated figure of about three watt-hours per query.²⁷ Swap: an early, high estimate set against a 2009 search figure, with neither updated. Correct read: a typical query today is around 0.3 watt-hours, comparable to a search and falling; ten times a tiny number is still tiny.²⁸ Addressed in: §06.

Underneath all five is a data problem. Most public figures are best-available estimates with active methodological debate, and there is still no standardized water disclosure across operators.¹ That vacuum is what lets each side reach for the number that suits it.

The future of AI is a power decision

Follow the water far enough and it stops being a water story. It is the first visible sign of what the AI build-out really is: an energy build-out, bounded by the grid and the basin, not by chips or algorithms.

• The constraint is physical and local. The next data center is built where there is power to run it and water to cool it, so interconnection queues, siting, and basin permits now gate AI more than model size does.
• The footprint is chosen, not given. The same model is thirsty on a gas grid and nearly water-free on wind and solar, and a utility makes that choice once, for decades.
• AI has become a forcing function on the grid. It is large enough to pull new power plants into existence, and what gets built, gas or wind and solar or nuclear, is being decided now, in procurement and siting, not in the labs.

Run the simple version. Power per rack is exploding, from 5 to 9 kilowatts a few years ago to 120-132 today, with roadmaps to 240-600, as much as a hundred times a standard server rack.¹⁹ Stack that density into a build-out this size and the totals move fast: US data-center electricity, about 176 terawatt-hours in 2023, is projected at 325 to 580 by 2028, double to triple, much of it AI.²⁹ The water that growth drags upstream depends entirely on what gets built to supply it.

That is the conclusion under all the numbers. The future of AI is not really a question about chatbots or water bottles. It is a question about what we plug them into. Point the marginal megawatt at wind and solar, firmed with storage, and AI grows cleaner and less thirsty as it scales; let the round-the-clock gap fall to gas or thirsty nuclear, the easy default, and it does the reverse. There is no free megawatt, but there is a choice, and a newer one worth watching: firm, water-light heat from the ground, if next-generation geothermal scales in time. The technology is not making the call. We are, and we are deciding it now.