The AI water issue is fake

"Suppose I take a cup of water from a lake, and then immediately dump it back in."

Three distinctions arrive in the first few paragraphs, and most confusing water headlines dissolve into one of them. Withdrawal versus consumption: water borrowed and returned versus water evaporated away. Direct versus indirect: water in the cooling loop versus water at the power plant supplying the electricity. And potable versus ordinary freshwater. On his accounting, most of what gets reported as AI's water is withdrawn non-consumptively at power plants, and only a thin slice is drinking water evaporated inside a data center — which is why two honest numbers about one facility can differ tenfold. These are the standard terms of water management, not rhetorical inventions — but a distinction can also be used as a shield.

"It's a miracle that something we spend 50% of our time using only consumes 0.2% of our water."

A count of gallons is not information until it is divided by something. He performs that division repeatedly — against irrigated corn, golf courses, steel, the daily water footprint of an average American life. Read slowly enough to notice that choosing a denominator is itself an argument. Comparing data centers to golf carries the premise that golf is a tolerated use; comparing them to agriculture treats agriculture as fixed background. Both are fair; both persuade beyond the arithmetic. The habit worth stealing: ask what a statistic is a share of, then ask what an opponent would have divided by instead.

"Data centers have an impact on local water systems, just like any other private industry."

National aggregates cannot answer local objections, because concentration is the complaint: a small share of a national total can be a large share of one aquifer. This is where he makes his real concessions: a Georgia county whose utility documents name a data center campus among the drivers of rising costs, and construction cases where sediment ruined nearby wells. His argument is that those are problems of large buildings, not server cooling, and once separated the operational record is nearly empty.

Read fairly in both directions. The strongest local objection is not about gallons. It is that a burden lands on a specific place, decided largely by people who do not live there — and "it behaves like any normal industry" reassures only someone who already trusts industrial siting. His reply — these facilities generate far more tax revenue per gallon than alternatives — answers the resource question and not the procedural one.

"There's nothing metaphysically special about this water."

This section takes the most potent version of the complaint — they're taking our drinking water — and inverts it. Treating freshwater to drinking standards costs a couple of dollars for every thousand gallons; water utilities have real economies of scale; so a large, steady, paying customer tends to fund treatment capacity rather than exhaust it. He presses further with a "crisis that doesn't happen" test, hunting for cases where commercial demand in a freshwater-rich region drove residential prices up, and finding almost none.

One link is worth marking: it depends on utilities allocating commercial costs to the commercial rate class, not to households — an empirical claim the whole inversion turns on.

"If water usage isn't an issue, it seems like the main effect of water cooling is preventing a significant amount of CO2 emissions"

Short section, disproportionate importance. Water carries heat far better than air, so a facility cooling without water spends electricity doing mechanically what water does thermally. Pressure on operators to go water-free can therefore push them toward higher energy use and higher emissions — which should complicate any policy view built from the essay. The general lesson is its most useful: optimizing hard on a single metric tends to relocate a harm rather than remove it. Note that one efficiency figure carries the section.

"This would leave you with an incorrect understanding of banks."

The most enjoyable part of the essay, and the most easily read badly. Three stories get examined: the Washington Post piece behind "every email costs a bottle of water," a syndicated story about Texans asked to shower less, and a New York Times feature about a family whose well failed near a Meta site.

Try naming the failure in each case before reading his diagnosis, because they are three different species. One is an arithmetic error propagated at scale. One is a real number presented without the denominator that would have made it unremarkable. The third is an article whose body is broadly accurate while its subtitle asserts what the body never claims — harm from cooling water, when the harm came from construction. That last case generalizes: the story was not fabricated, it was framed, and the frame is what readers retained.

"This is the great singular sin of bad climate communication. The second you see it, you should assume it's misleading."

The closing taxonomy is the most portable page here: measuring against household use when homes are a tiny share of a water footprint; gesturing at "hidden true costs" without stating them; vague verbs like straining and exacerbating that are technically true of any use at all; large raw numbers with no comparison; and permit ceilings reported as actual consumption.

Give the hidden-costs item an extra minute. He builds a parable with two invented data centers showing that the ratio of unreported to reported water ranks facilities backwards: the more efficient a facility's own cooling, the larger its offsite water looks beside it, and the more it appears to be concealing. A ratio that inverts the ordering of what you care about is a failure mode you will meet far outside this topic.