The "24 million cars" data centre stat is a unit error, and citing it costs you the argument

A Cornell-led paper published in Nature Sustainability on 10 November 2025 projects that US AI servers could emit between 24 and 44 million tonnes of CO2-equivalent a year by 2030 ^[1][2]. A widely shared headline recast that as planned US data centres emitting as much carbon dioxide as 24 million cars, which is not what the paper says: the researchers' own vehicle translation is 5 to 10 million cars ^[4][3].

According to The Next Web, the inflated number appears to come from reading "24 million tonnes" as "24 million cars" ^[5]. The arithmetic of the error is worth stating plainly, because it is the first thing an opposing counsel or a sceptical CFO will do: 24 million cars is 2.4 times the authors' upper-bound equivalent and 4.8 times their lower bound ^[6]. Run the authors' own figures and the implied per-vehicle rate is roughly 4.4 to 4.8 tonnes a year, which is what a car-equivalence claim should be built on ^[7]. Anyone putting the 24 million figure into an ESG disclosure, a submission on data-centre siting, or a slide arguing for a moratorium is handing the other side a free win on accuracy, and losing the corrected number in the process.

The corrected number is not small. The study models emissions and water state by state, pairing a hybrid statistical and thermodynamic model of server efficiency with the US government's ReEDS grid model across five demand scenarios ^[8]. The 24-to-44 range is the spread between a restrained build-out and a frenzied one rather than a single forecast ^[9]. Alongside it sits a projected water footprint of 731 to 1,125 million cubic metres a year, roughly the household use of 6 to 10 million Americans ^[10]. Taken together, The Next Web notes, the fleet's resource draw starts to rival that of a mid-sized US state ^[11].

The result is also more sensitive to policy than to physics. Under cheap-renewables assumptions the study finds emissions falling by more than 15 percent; under expensive ones they rise by a fifth ^[12]. Best-practice interventions across siting and procurement could cut emissions by up to 73 percent and water use by up to 86 percent ^[13]. The authors point to the Midwest, and to Texas, Montana, Nebraska and South Dakota in particular, as better locations than water-stressed Northern Virginia ^[14].

Then there is the word doing the most work in the headline version: planned. Announced capacity is not built capacity, and shelved data-centre projects are common when power, permits or demand fail to arrive ^[15]. The study's upper bound assumes expansion close to the industry's most bullish forecasts ^[16]. Hyperscalers are signing power-purchase agreements for wind, solar and nuclear while also hedging with gas; Amazon's planned Texas campus could become one of America's single biggest polluters because it plans to burn its own fuel rather than draw from a cleaner grid ^[17][18]. Environmental advocates warn that on the current trajectory much of the fleet runs on fracked gas well into the 2030s ^[19].

Watch whether the 24 million figure survives into filings and regulatory consultations, particularly in Europe, where reporting rules and efficiency pacts already give officials leverage over data-centre emissions and where The Next Web argues the slip is one regulators can ill afford to repeat ^[20]. Watch, too, how fights over new gas plants resolve, since those are the test of whether the bullish build-out assumption holds ^[16]. Fengqi You's own framing is the usable one: it is not too late to plan for these constraints ^[21].