The Real Cost of AI Will Shock You

usefulBe Smartvideo2026-10-22read planted ai llmssciencebusiness

Synopsis — AI-drafted from Dan's notes

Joe Hanson, host of PBS’s Be Smart, sets out to find what building AI actually costs in electricity, water, land and the towns next to the data centers, and talks to researchers who study it, among them Jonathan Koomey and Shaolei Ren. He starts with history: late-1990s warnings that the internet would eat half of US power were wildly high, because computing kept getting more efficient (Koomey’s own work found computations per kilowatt-hour doubling about every 1.6 years from the 1940s on). AI breaks that pattern. A rack of AI servers draws many times what a conventional rack does, and US data centers reached about 4.7% of the country’s electricity in 2024 (Lawrence Berkeley National Laboratory’s 2025 update), with forecasts of roughly 7–12% by 2028 that nobody, he stresses, can pin down. Efficiency may not save us either: Jevons’s paradox says cheaper computing means more of it.

His key point is that the costs are local. A data center plugs into one grid, and where the grid can’t keep up, builders now skip it and run their own gas turbines on site. His example is GW Ranch in Pecos County, Texas: 35 turbines and about 7.65 gigawatts of nameplate capacity, which its developer calls the largest power plant in the US. On emissions, he walks through scopes one to three and puts US data centers at about half a percent of national greenhouse gases today, small but growing. The famous “a bottle of water per question” comes from Ren’s 2023 paper, which estimated 500 ml for roughly 10 to 50 GPT-3 responses, most of it water used to generate the electricity; Ren himself warns against applying it to today’s models. Water, like power, is local: around Phoenix, he says, data center cooling demand is forecast to grow severalfold while the Colorado River shrinks.

The last third is about people. xAI’s Colossus in Memphis ran up to 35 turbines against a permit for 15, according to the Southern Environmental Law Center, and a 2024 Caltech and UC Riverside study projected about 1,300 premature US deaths a year from data center air pollution by 2030. Developers work through shell companies and non-disclosure agreements, so in Pine Island, Minnesota, residents say they heard about a Google data center long after officials did. States compete with tax breaks for buildings that employ few people. His conclusion: AI is here to stay, but its full costs aren’t being counted, the companies should pay them, and the data to weigh costs against benefits exists but isn’t public.

Connections