What is the actual CO2 emissions from US data centres in terms of the equivalent of cars?

      A recent study from Cornell has provided the climate critics of the AI industry with a new statistic to champion, which has quickly garnered attention online. A widely circulated headline claimed that planned data centers in the US will generate as much carbon dioxide as 24 million cars.

      This striking figure is misleading. It misinterprets the original research it purports to summarize, highlighting a significant gap in how the environmental impact of the AI boom is communicated. Published in *Nature Sustainability* on November 10, 2025, the paper originates from Fengqi You's Process-Energy-Environmental Systems Engineering lab, with Tianqi Xiao as the lead author alongside co-authors from KTH in Stockholm, Concordia in Montreal, and the RFF-CMCC institute in Milan.

      The paper's main projection indicates that US AI servers could emit between 24 and 44 million tonnes of CO2-equivalent annually by 2030. The researchers themselves relate this to the equivalent of 5 to 10 million cars, not 24 million. The alarming number appears to stem from mistakenly interpreting "24 million tonnes" as "24 million cars," a slip that European regulators, who are scrutinizing Big Tech over data center emissions, cannot afford to repeat.

      Even the accurate figure warrants careful examination rather than immediate alarm. The research team modeled emissions and water usage state by state, combining a hybrid statistical and thermodynamic model of server efficiency with the US government's ReEDS grid model across five demand scenarios. The range of 24 to 44 million tonnes is not a single prediction but rather the difference between a conservative build-out and an aggressive one. This number coincides with a projected water footprint of 731 to 1,125 million cubic meters per year, which is comparable to the household usage of 6 to 10 million Americans.

      When combined, these figures reveal that the resource demands of the fleet may rival those of a mid-sized US state, making the way the information is presented as crucial as the numbers themselves. The grid mix plays a significant role in determining outcomes. Under scenarios assuming low renewable energy costs, emissions could reduce by over 15%; conversely, under high-cost scenarios, emissions might increase by a fifth. Implementing best-practice measures, from site selection to procurement, could decrease emissions by up to 73% and water usage by as much as 86%.

      Essentially, the headline figure reflects a worst-case policy scenario rather than a fixed scientific principle. The locations of the servers also matter: the authors highlight the Midwest, specifically Texas, Montana, Nebraska, and South Dakota, as more favorable than water-scarce Northern Virginia.

      However, counterarguments are also present. Major tech companies are investing heavily in power purchase agreements for wind, solar, and nuclear energy, while a wave of startups is striving to reduce data center energy consumption. Still, many of these companies are also relying on gas. For instance, Amazon's planned campus in Texas could potentially become one of the largest polluters in the country due to its plan to use its own fuel instead of connecting to a cleaner energy grid. Environmentalists warn that, if current trends continue, much of this new infrastructure will operate on fracked gas well into the 2030s.

      A crucial caveat lies in the word "planned." Announced capacity does not equal built capacity, and history shows numerous data center projects have been quietly abandoned when power, permits, or demand failed to meet expectations. The study's upper limit assumes the industry will grow in line with its most optimistic forecasts—an assumption that will be tested in the ongoing debates over new gas plants.

      This does not imply that the trend is harmless. Even the study's lower estimate represents a significant increase in US emissions at a time when efforts to decarbonize the grid are underway. However, the responsible interpretation is one that You himself presents: it is still possible to plan for these limitations.

      The concerning number is not immutable; it is a matter of policy choice. Europe, with its regulations on data center reporting and efficiency agreements, is quietly betting that it can reduce this figure.