The cheapest thing in the AI buildout may be the model that tells you where to put it

An essay in Fast Company makes a narrow, testable claim: it costs about $5 and takes roughly 15 minutes to train an AI model that can help determine where America's next data centers should be built ^[1]. If that is close to right, the expensive item in AI capacity planning is not the analysis but everything the analysis is currently used to justify.

The demand figure doing the work is Bloomberg's projection, cited in the piece, that data centers will consume up to one-fifth of US power by 2035, up to 200 gigawatts, with much of it landing in regions where power is already constrained ^[2]. The supply response is slow: the essay says generation wait times can stretch past five years, on infrastructure that is 50 to 70 years old in much of the country ^[3].

Against that, the argument's load-bearing physical fact: grids are sized for the handful of hours a year when demand peaks, and across most hours only about half of that built and paid-for capacity is in use ^[4]. The author calls locating and routing the rest "capacity mining", meaning models that evaluate where and when the grid has unused power along with the costs and timelines of the infrastructure needed to reroute it ^[5]. The proposed order of operations puts demand matching first, citing locations and operating profiles with available power and operational flexibility rather than a site chosen for other reasons; transmission and battery expansion second; and new generation last, on the grounds that plants are costlier, slower and often polluting ^[6].

The money claim is single-sourced and the source is anonymous. At one large investor-owned utility, which the piece does not name, the author says large loads including data centers can produce roughly $1 million per megawatt per year, about $1 billion per gigawatt ^[7], revenue that, structured correctly, offsets fixed grid costs every other customer currently carries alone ^[8]. Applied to the top of the Bloomberg range, that arithmetic implies on the order of $200 billion a year in new utility revenue ^[9], which is why the essay frames this as the largest revenue growth in a generation for businesses that have been flat for decades ^[10]. No method is shown for the per-megawatt number.

The rest is sequencing politics. The piece notes utilities recently pledged at the White House to shield consumers from rising electric bills tied to data center growth ^[11], and argues that capacity which already exists does not need to be paid for twice, while finding it costs relatively nothing ^[12]. It also rests on a stated belief, not evidence, that hyperscalers will gladly cover more of the cost in exchange for faster access to power ^[13]. The essay is explicit that none of this needs new hardware or a scientific breakthrough, only the right questions and cheap models ^[14]. That framing is also its weak point: asking the right question is the $5 part, and interconnection studies, data access and tariff negotiation are not modelling problems.

Three things would show the argument is being taken up rather than merely published. Whether utilities file flexible or curtailable large-load tariffs that price the operating flexibility the matching step assumes ^[6]. Whether anything resembling $1 million per megawatt per year appears in a rate case or an investor deck with a method attached ^[7]. And whether hyperscalers accept siting and curtailment constraints in exchange for speed, which is the exchange the whole sequence depends on ^[13].