Browse
Log in / Register
Share
Published: November 12, 2025
A widely shared headline last week claimed that planned US data centers would emit as much carbon dioxide as 24 million cars. The number felt like a sledgehammer. It is also wrong.
The real figure, from a peer-reviewed study in Nature Sustainability (November 10, 2025), is 5 to 10 million cars’ worth of CO2 equivalent — not 24 million. The 24 million car-equivalent appears to be a simple arithmetic error: someone mistook “24 million tonnes” for “24 million cars.” That slip has since ricocheted across social media and news outlets, and it neatly illustrates how the environmental cost of the AI boom gets distorted in translation.
The paper, led by Tianqi Xiao and Fengqi You at Cornell’s Process-Energy-Environmental Systems Engineering lab, with co-authors from KTH Stockholm, Concordia University, and the RFF-CMCC European Institute on Economics and the Environment, models the fleet of US AI servers through 2030. Its central projection: 24 to 44 million tonnes of CO2 equivalent per year. The researchers themselves convert that into vehicle equivalents: 5 to 10 million passenger cars, depending on the fleet’s composition and grid mix.
The misreading matters beyond academic accuracy. European regulators, who are already leaning on Big Tech to disclose data-center emissions, risk basing policy on inflated figures if they rely on the viral headline. The correct number is still sobering — but it is not a catastrophe locked in by physics.
The team paired a hybrid statistical-thermodynamic model of server efficiency with the US government’s ReEDS grid model, simulating five demand scenarios across all 50 states. The 24-to-44 million tonne range is not a single forecast; it is the spread between a restrained build-out and a frenzied one. Alongside carbon, the projected water footprint reaches 731 to 1,125 million cubic metres annually — equivalent to the household use of 6 to 10 million Americans.
Taken together, the resource draw of the AI server fleet would rival that of a mid-sized US state. But the framing matters. The study shows that emissions swing by more than 15% under cheap-renewables assumptions and climb by a fifth under expensive ones. Best-practice interventions — siting decisions, renewable procurement, efficiency upgrades — could cut emissions by up to 73% and water use by 86%.
In other words, the headline number describes a worst-case policy environment, not an iron law. Where the servers land is critical: the authors point to the Midwest, especially Texas, Montana, Nebraska, and South Dakota, as better bets than water-stressed Northern Virginia.
Hyperscalers are pouring money into power-purchase agreements for wind, solar, and nuclear. A wave of startups, such as Crusoe Energy and Heirloom Carbon, is racing to curb data-center energy use and capture remaining emissions. But the same firms are hedging with gas. Amazon’s planned Texas campus, for example, could become one of America’s single biggest polluters precisely because it plans to burn its own natural gas rather than draw from a cleaner grid.
Environmental advocates warn that on the current trajectory, much of the fleet will run on fracked gas well into the 2030s. The fight over new gas plants — whether to approve them as backup or as primary power — is the central battleground for data-center emissions.
Exclusive perspective: Since the study’s publication, the US Energy Information Administration (EIA) released its November 2025 Short-Term Energy Outlook, which shows that natural gas-fired electricity generation in the US data-center-heavy regions (PJM, ERCOT, MISO) is projected to rise 4.2% in 2026, even as wind and solar capacity additions hit record highs. This suggests that the “expensive renewables” scenario in the Cornell model may be materializing sooner than expected, pushing emissions toward the upper end of the 24–44 million tonne range.
The study’s upper bound assumes the industry expands at close to its most bullish forecasts. But “planned” capacity is not built capacity. History is littered with data-center projects that were quietly shelved when power, permits, or demand failed to materialise. The fight over new gas plants is precisely the mechanism that will test whether the upper bound is realistic.
None of this makes the trend benign. Even the lower estimate of 24 million tonnes is a meaningful addition to US emissions at a time when the grid is supposed to be decarbonising. The number to fear is not fixed. It is a policy choice.
The Cornell team’s message is clear: it is not too late to plan for these constraints. The US can bend the curve by accelerating renewable deployment, tightening siting rules, and ensuring that data centers do not lock in fossil-fuel dependence. Europe, with its data-center reporting rules and efficiency pacts, is quietly betting it can be bent downwards. The US now has the same opportunity — if it can look past the scary headline and read the actual math.
