The Water Question | RLK Consulting

People ask me a version of this question now, and it deserves better than the two reactions it usually gets, which are guilt and eye-rolling. So here is the real answer, with real numbers, and at the end, the tool: the questions a C-suite can put to its vendors and the contract terms that make the answers matter.

The question usually arrives with one of two statistics attached. Either the one where a single ChatGPT question drinks a bottle of water, or the one where it uses a fifteenth of a teaspoon. Both numbers are in circulation. Both come from sources with a stake in your reaction. And the gap between them, roughly six thousand to one, is the most useful thing in this whole debate, because understanding why they differ tells you where the water actually goes.

What a question actually costs

The teaspoon number comes from Sam Altman, who wrote that an average ChatGPT query uses about 0.000085 gallons of water and 0.34 watt-hours of electricity. The bottle number comes from researchers at UC Riverside, whose work found a 100-word AI-generated email accounts for around 519 milliliters of water. Neither is fabricated. They are measuring different pipes.

Altman is counting the water his data center touches directly: the evaporative cooling that carries heat off the servers. The Riverside team also counts the water evaporated at the power plants generating the electricity, which is usually the larger share. One number is the restaurant check for your table. The other includes the farm, the truck, and the kitchen. Vendors quote the first because it is small. Critics quote the second because it is big. An operator should hold both, and notice what they agree on: the per-query cost is tiny, and the per-query cost is also the wrong unit of analysis, the same way nobody evaluates a highway by the gasoline in one car.

The takeaway: the viral per-query numbers differ six-thousand-fold because they draw the accounting boundary in different places, and both are dwarfed by the real question of scale. Distrust anyone whose argument depends on you not asking what got counted.

Where the water actually goes

Scale, then. The Department of Energy asked Berkeley Lab to count, and its report to Congress found US data centers consumed about 66 billion liters of water directly in 2023, on electricity use that hit 4.4 percent of the national total and could reach 6.7 to 12 percent by 2028. Hyperscale facilities alone are projected at 60 to 124 billion liters by 2028. The IEA expects global data center electricity to roughly double to 945 terawatt-hours by 2030, about what Japan uses today. The growth is real, steep, and mostly ahead of us.

Direct water use, United States, per year

The aggregate is small. That is the wrong comfort.

Sources: USGS irrigation estimates (~73B gallons per day); Golf Course Superintendents Association national survey (~500B gallons per year); Berkeley Lab 2024 report to Congress (~66B liters direct data center consumption, 2023). Direct use only; the power-generation share adds more.

So the honest aggregate picture: American golf courses use roughly thirty times more water than American data centers, and agriculture uses more in a single day than data centers use in a year. If the argument stopped there, the industry’s defenders would win it. It should not stop there, for two reasons.

First, data center water is unusually consumptive. A household returns most of what it withdraws to the system. Evaporative cooling loses most of what it takes, up to 70 to 90 percent by some engineering estimates, and what evaporates in a basin is gone from that basin. Second, and this is the real issue, water is not a national number. Water is local, and the industry keeps building where it is scarce. A Bloomberg analysis found roughly two-thirds of new US data centers built or planned since 2022 sit in areas of high water stress, because the same places that have cheap land and cheap power tend to be dry.

Then the receipts get specific. In Newton County, Georgia, a Meta data center draws around 500,000 gallons a day, about a tenth of the county’s water, and after construction began on the newer $750 million facility, nearby families reported wells silting up and failing. In The Dalles, Oregon, the city spent thirteen months in court trying to keep Google’s water use secret before a local newspaper forced the records out. Notice what those two stories have in common. The sin was not computing. The sin was taking scarce water quietly and calling the number a trade secret.

The takeaway: the national numbers are small and the local ones are not, because two-thirds of the buildout is landing in places already short on water. The real offenses are siting and secrecy, and both have names and addresses.

The part I own

Now the reconciliation, because I sell AI strategy for a living and the demand I help create is part of that growth curve. I am not going to hand-wave that away.

Here is what I know from the inside. MIT’s research found that about 95 percent of enterprise AI pilots produce no measurable P&L impact. I have written about that number as a business problem for a year. Read it as an environmental number instead. Every one of those failed pilots ran on real electricity, cooled by real water, evaporated from a real basin, and bought nothing. The water spent on AI that works is at least a trade you can argue about. The water spent on a pilot that was scoped to fail, run past its kill date, and quietly renewed because nobody was measuring it, that water bought a slide deck.

My entire practice is the removal of exactly that waste. Kill the zombie pilots, cap the experiments, make somebody own a number, buy instead of build when a vendor already solved it. I built a career on the argument that most enterprise AI spending is waste, and waste has a footprint. The greenest query is the one you never needed to run, and the second greenest is the one that actually moved a number, because it displaced something. That is the same discipline I sell, pointed at a different bottom line.

The takeaway: the 95 percent failure rate is an environmental statistic wearing a business costume, and killing wasted AI spend is the most direct water policy a company controls. The footprint that bought nothing is the one nobody can defend.

The harder questions

The buyer side has more leverage than it thinks. Microsoft is piloting closed-loop data centers that evaporate no water at all, starting in Phoenix and Wisconsin, saving a claimed 33 million gallons per site per year. That engineering did not happen out of kindness. It happened because customers, regulators, and neighbors made water a cost. Pressure works, and the pressure arrives as questions your vendors have to answer in writing. Here are the five I would put in the next QBR or RFP, with what each answer tells you.

The Water Questions

Which regions do our workloads run in, and what is the water-stress rating of each? Every major provider maps this internally. A vendor who cannot answer has not looked, which is also an answer.
What is the water usage effectiveness of the facilities serving us, and which direction is it trending? Microsoft publishes a fleet WUE and a 40 percent improvement target. A vendor who will not name theirs is telling you the number is worse.
Is the cooling evaporative, closed-loop, or air, and what share of withdrawn water is consumed? Evaporated water leaves the basin for good. The cooling design is the water policy.
Do you disclose site-level water use to host communities, or has disclosure ever required a lawsuit? Call this the Dalles test. There is a right answer.
Can our workloads be placed in lower-stress regions, and at what price difference? The moment water becomes a routable, priceable dimension of the service, it starts getting managed like one.

Terms that hold water

Questions establish the facts. Contracts give them consequences, and here the ground has shifted in the buyer’s favor: Amazon, Google, and Microsoft have all publicly pledged to be water positive by 2030, and Microsoft reports that in fiscal 2025 it replenished more water than it withdrew across its global operations, which it describes as a milestone toward that 2030 commitment. The promises exist. Your contract’s job is to borrow them.

Terms That Hold Water

Borrow the pledge. Incorporate the vendor’s own public water commitments by reference, with missed milestones triggering a service review or credits. They wrote the promise in a press release; move it somewhere with consequences.
A disclosure covenant. Annual, site-level water reporting for the regions serving you, WUE included, with no trade-secret carve-out for aggregate numbers. If a small-town newspaper can win this in court, you can ask for it at renewal.
A siting ceiling. Notice, and a no-cost migration right, when your workload moves into a high-water-stress region. You already cap price escalators. Cap basin risk the same way.
An efficiency floor. New capacity serving you meets a closed-loop or air-cooled standard, or a named WUE threshold, by a named year. Zero-evaporation sites ship in 2026, so “industry standard” now has a number.
Score it even when you cannot redline it. A $60M company will not mark up a hyperscaler’s master agreement, and does not need to. RFP scoring, region selection, and renewal leverage are where this bites. Ask in writing, grade the answers, choose accordingly.

One more thing I tell clients, for balance-sheet reasons that double as water reasons: model your AI workloads at two to three times today’s prices, because current pricing is subsidized by other people’s capital. What survives honest pricing is the work worth doing. Honest prices are coming either way, and they will do more for the aquifers than any pledge.

And if your instinct says the clean answer is abstinence, I will give that argument its due. It is the one position with no accounting games in it. It is also the position with no leverage. A mid-market firm that walks away changes no aquifer in Georgia; a mid-market firm that stays a customer and puts water questions in its contracts is one of thousands of small pressures that produced the closed-loop design. Meanwhile the growth projections are real, efficiency gains have a long history of getting eaten by new demand, and the CEO of OpenAI went on stage this year and called the water concerns “fake.” They are not fake. They are concentrated, in specific dry counties, on specific failing wells, behind specific nondisclosure fights, and dismissing them is how an industry earns the moratoriums now being drafted against it.

The takeaway: your procurement questions and your program discipline are the two water levers a normal company actually holds. Pull both and you have done more than a year of guilt.

So, the answer

How do I reconcile it? I don’t reconcile it by calling the water fake. The vendor selling the tokens does that, and he is wrong, and he is also selling something. I reconcile it by knowing the numbers cold, respecting the difference between a teaspoon and a watershed, and spending my working hours killing the one part of the footprint that buys nothing: the 95 percent of programs that evaporate water to produce a status update. The technology is not going back in the box. The water math will be decided by siting rules, disclosure fights, cooling engineering, and buyers who ask harder questions. The list above is the harder questions. I intend to keep being one of them.

How RLK Can Help

The Vendor and Spend Strategy engagement puts the questions and terms above into your actual RFPs and renewals, alongside the price and exit terms that belong next to them. My free AI Diagnostic shows in about eight minutes which parts of your AI program are earning their footprint and which parts are the 95 percent. And the AI Business Case engagement stress-tests the numbers at the prices that are coming. Start the conversation.