Every prompt your team runs this week has a water footprint. Nobody sends you that invoice, so nobody plans for it.
That arrangement is ending, and not because of a moral argument. It is ending because the cost is being formally assigned.
California has signed a package of seven bills aimed at stopping data centres from passing utility costs to residents. The laws require the state utilities commission to create a new rate classification for data centres, and force operators to pay for upgrades to local power grids and water systems.
Other bills in the package require proposed data centres to disclose estimated water use to local governments, along with energy efficiency information and drought planning. Meeting energy, water and fuel consumption thresholds becomes a condition of the faster approval route.
A cost that was always there, now with an owner
Read that as an accounting change rather than an environmental one and the commercial implication gets obvious.
Grid upgrades were always necessary. Water was always consumed. The open question was who carried it, and the answer was frequently a local utility spreading it across everybody's bill.
Once that cost sits on the operator's balance sheet with a rate classification attached, it flows into the price of compute. Compute is the input cost of every AI tool in your stack.
Not next quarter. These things move through contracts slowly, and hyperscalers have enormous capacity to absorb margin pressure while they are still buying market share.
But the direction is set, and direction is what you budget against. The per-token price you are paying today is the most subsidised it will ever be.
The disclosure gap is the tell
Here is the detail that should interest anyone who reads numbers for a living.
Amazon has committed $20 million to water conservation along the Colorado River, inside a collaborative it expects to raise $100 million over two years. The river supplies 40 million people across seven US states and two Mexican states, and the US Bureau of Reclamation describes the drought of the last 26 years as unprecedented.
The company has a public goal of being water positive by the end of the decade. It has disclosed some information about its water use. The full picture is still incomplete.
A company that measures everything, that knows the power draw of individual racks, has not published a complete water figure. That is not an oversight. Figures get withheld when publishing them would invite a conversation about the trend.
And the conservation commitment is the smaller number. $20 million against a water system serving 40 million people is a rounding error in a way that a mandatory rate classification is not.
Voluntary spending is reputation management. Regulation is cost of goods sold. Only one of them shows up in your renewal quote.
What this does to a marketing budget
Most teams built their AI business case on a ratio: this tool costs X per month and replaces Y hours of work. That ratio has been improving for two years, which trained everyone to treat improvement as the baseline.
Plan instead for a floor. At some point the price of intelligence stops tracking model efficiency and starts tracking the cost of electricity and water in specific regions.
Three consequences follow, and none of them are dramatic in isolation.
First, per-seat pricing becomes less attractive to vendors than usage pricing, because usage pricing passes volatility to you. Expect more plans where the heavy user pays the real cost.
Second, the gap between frontier models and good-enough models becomes a budget line rather than a preference. Running your bulk classification on a smaller model is currently an optimisation. It becomes a cost control.
Third, the agent architectures that burn compute in the background get audited. I looked at the real cost of running fifteen agents earlier this year, and that arithmetic gets worse, not better, as the input price rises.
What the vendors will do about it
Suppliers facing a rising input cost have a predictable set of moves, and they usually arrive in the same order.
Rate limits come first, because they reduce cost without changing a price on a page. If your plan quietly gains a cap it did not have, that is the input price arriving in a form nobody has to announce.
Caching and routing come next. Your request gets served by a smaller model, or from a stored answer, whenever the vendor judges the quality difference tolerable. You will not be told which requests those were.
Then tiering, where the current behaviour becomes the premium plan and the standard plan becomes something slightly worse than what you have today. That is the most common shape of a price rise in software, because it never reads as one.
Watch for all three in renewal terms rather than in announcements. The phrase to look for is any language converting a fixed allowance into a fair use allowance.
The practical move this quarter
Find out what you actually spend. Not the subscriptions, which are easy, but the API usage sitting on three different corporate cards and inside two products you pay for by seat.
Then work out your cost per useful output. Cost per approved asset, per qualified lead, per resolved ticket. Not cost per token, which tells you nothing about whether the spend earned anything.
That number is your actual exposure. If it doubled, which parts of your operation would stop making sense?
When I broke down what a full day of AI television costs, the interesting finding was not the total. It was how much of it was invisible until somebody itemised it.
Do the itemising now, while the numbers are still flattering. Budgets built on a subsidised input price are budgets built on somebody else's patience.
The water was always in the bill. It just had someone else's name on it.