Cornell Study Projects AI Server Emissions by 2030
A Cornell University study projects that US AI servers could emit 24-44 million metric tons of CO2-equivalent annually by 2030, equal to emissions from 5-10 million cars. The Next Web reports that a widely circulated comparison of the lower-bound 24 million tonnes with 24 million cars misstates the study's vehicle-equivalency calculation. Cornell also estimates annual water use of 731-1,125 million cubic meters.
A Cornell University research team projects that US AI servers could emit 24-44 million metric tons of CO2-equivalent per year by 2030, an amount the researchers equate to emissions from 5-10 million cars. The estimate does not support a comparison to 24 million cars, according to an August 18 article by The Next Web that examined how the study has been summarized.
The underlying paper, published in *Nature Sustainability* on November 10, 2025, models US AI-server infrastructure at the state level. Cornell's announcement identifies Tianqi Xiao as first author and Fengqi You as the project lead. The study also estimates annual water consumption of 731-1,125 million cubic meters by 2030, which Cornell compares with annual household water use by 6-10 million Americans.
What the emissions comparison means
The central numerical distinction is between mass emissions and vehicle-equivalent emissions. The study's 24-44 million-tonne range is a CO2-equivalent estimate, while its researchers convert that range to 5-10 million cars, Cornell reports. Treating 24 million tonnes as 24 million vehicles conflates two different units.
The Next Web reports that the range spans five modeled demand scenarios rather than representing one fixed forecast. Its account describes a hybrid statistical and thermodynamic model of server efficiency combined with the US government's ReEDS grid model. This type of modeling makes the assumed pace and location of build-out, server efficiency, and regional electricity mix material inputs to the result.
Grid mix, siting, and water use
Cornell reports that the researchers compiled financial, marketing, and manufacturing data on infrastructure growth, then combined those inputs with location-specific power-system, resource-consumption, and climate data. You said incomplete industrial reporting complicated the work, because companies do not disclose every relevant data point.
The study's results vary with electricity assumptions. The Next Web reports that emissions fall by more than 15% under cheap-renewables assumptions and rise by about one-fifth under expensive-renewables assumptions. That dependence is consistent with a basic infrastructure accounting constraint: an AI workload's operational carbon footprint depends not only on compute demand, but also on the carbon intensity of the electricity serving it.
Cornell reports that smart siting, faster grid decarbonization, and operational efficiency could reduce carbon impacts by approximately 73% and water impacts by 86% relative to worst-case scenarios. The PEESE research group describes these measures as a state-by-state roadmap for decisions about where and how AI data centers are built.
NPR reported that Google, Microsoft, and Meta had set 2030 net-zero and water-positive targets, while Amazon's net-zero deadline is 2040. NPR further reported that the paper concluded AI growth could put technology-sector climate goals out of reach; Google did not respond to NPR's request for comment, while the other companies declined to comment.
For ML infrastructure teams, the study provides a reminder that model efficiency metrics alone do not describe deployment externalities. Comparable capacity expansions are commonly evaluated through a combination of workload efficiency, regional grid emissions, cooling design, and local water constraints.
Key Points
- 1Cornell projects 24-44 million tonnes of annual US AI-server emissions by 2030, equivalent to 5-10 million cars.
- 2The widely shared 24-million-car framing confuses tonnes of CO2-equivalent with the study's vehicle-equivalent emissions calculation.
- 3Infrastructure emissions models commonly depend on compute growth, regional electricity mix, cooling demand, and data-center location.
Scoring Rationale
The study offers state-level estimates of carbon and water impacts from a major expansion of AI compute infrastructure. Its correction of a widely repeated unit error is relevant to practitioners who communicate infrastructure sustainability metrics.
Sources
Public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
