A 1,667x spread is not a metric: stop governing AI water use per query

The 500ml estimate and the 0.3ml estimate are not two measurements of one thing. According to the Forbes account, the gap opens up depending on whether upstream cooling is counted, what the task actually is, and where the demand originates ^[4]. Each of those is a scoping decision taken before anything is measured, which is how two people working honestly end up roughly 1,667 times apart on the same prompt ^[12].

That makes the per-query figure unusable as a governance object. Put one in a sustainability report and you have published a number whose opposite is equally citable from the same accounting menu ^[3][4]. The Forbes piece reports that a fair number of experts have already written the exercise off as futile and would rather push on efficiency ^[5]. Fine, but efficiency needs a denominator someone owns.

The macro side is not automatically that denominator. Yankai Jiang's estimate of total data center consumption at around 4.5 terawatts, glossed as "7 New York Cities" of power, came from a talk at Planet Action ^[6], an event the article's author says he helps run ^[7]. It is stated as a power figure with no reporting period attached, and it is an estimate from a stage rather than a company filing, which is precisely the distinction Jordan Hale of Tech Journal draws when he argues aggregate disclosures are more reliable because companies report them directly ^[11][14]. The auditable quantity is not global consumption. It is what your named suppliers report, what share of it you contracted for, and whether the ratio of output to input moves year over year.

The design recommendations in the same piece are more useful to a leadership team than any per-prompt figure, because they are contract language rather than disclosure prose. Jiang's Arizona-versus-Finland contrast turns on siting and cooling method: the Finnish case uses cold climate and nearby seawater and runs jobs when renewable energy peaks, without fresh water ^[8]. He treats carbon-aware scheduling as a partial but worthwhile start ^[9]. He also argues against overprovisioning hardware, in his words, "You do not always need a Ferrari. Sometimes a reliable Toyota will do just fine" ^[10]. A region, a cooling type, a scheduling window and a machine class are all things a procurement team can specify and later verify.

There is a cost line hiding in his rhetoric too. Jiang notes that when data centers burn out, "we just retire the old ones and build new ones" ^[13]. Replacement cadence is capital expenditure, and capital expenditure is already reviewed by the people being asked to defend millilitres per prompt.

The per-query number will keep circulating because it is vivid and free to quote. It is also the only figure in this debate that cannot survive a second question about its boundary.