How many cars' worth of CO2 will US data centers really emit?
August 18, 2026 - 10:18 am
A new study out of Cornell has handed the AI industry’s climate critics a fresh statistic to wield, and the internet has duly rounded it up. Planned US data centers, one widely shared headline announced, are set to produce as much carbon dioxide as 24 million cars.
The figure is arresting. It is also a misreading of the research it claims to summarize, and the gap between the two is a neat lesson in how the environmental toll of the AI boom gets reported.
The paper, published in Nature Sustainability on 10 November 2025, comes from Fengqi You’s Process-Energy-Environmental Systems Engineering lab, with lead author Tianqi Xiao and co-authors at KTH in Stockholm, Concordia in Montreal, and the RFF-CMCC institute in Milan.
Its central projection is that US AI servers could emit between 24 and 44 million tonnes of CO2-equivalent annually by 2030. The researchers themselves translate that into vehicles: 5 to 10 million cars, not 24 million.
The scarier number appears to come from mistaking “24 million tonnes” for “24 million cars,” and it is the kind of slip that European regulators leaning on Big Tech over data-centre emissions can ill afford to repeat.
Even the correct figure deserves interrogation rather than reflexive alarm. The team modeled emissions and water use 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.
The 24-to-44 range is not a single forecast but the spread between a restrained build-out and a frenzied one, and it runs alongside a projected water footprint of 731 to 1,125 million cubic metres annually, roughly the household use of 6 to 10 million Americans.
Bundle those together and the resource draw of the fleet starts to rival that of a mid-sized US state, which is why the framing you choose matters as much as the arithmetic behind it.
The grid mix does much of the heavy lifting. Under cheap-renewables assumptions, the study finds emissions falling by more than 15%; under expensive ones, they climb by a fifth. Best-practice interventions, from siting to procurement, could cut emissions by up to 73% and water use by up to 86%.
In other words, the headline number describes a worst-case policy environment rather than an iron law of physics. Where the servers land matters too: the authors point to the Midwest, and Texas, Montana, Nebraska and South Dakota in particular, as better bets than water-stressed Northern Virginia.
That is also where the counterarguments live. Hyperscalers are pouring money into power-purchase agreements for wind, solar and nuclear, and a wave of startups is racing to curb data centre energy use. Yet the same firms are hedging with gas: Amazon’s planned Texas campus could become one of America’s single biggest polluters precisely because it plans to burn its own fuel rather than draw from a cleaner grid. Environmental advocates warn that, on the current trajectory, much of this fleet will run on fracked gas well into the 2030s.
The crucial caveat is buried in a single word: planned. Announced capacity is not built capacity, and history is littered with data centre projects that were quietly shelved when power, permits or demand failed to materialise.
The study’s upper bound assumes the industry expands at close to its most bullish forecasts, which is precisely the assumption that the fight over new gas plants is meant to test.
None of this makes the trend benign.