Quick Answer
AI data center emissions are substantial: 403 U.S. hyperscale facilities used an estimated 68 to 99 terawatt-hours of electricity from May 2024 through April 2025, Harvard researchers found. Their midpoint estimate equals about 1.8% of U.S. electricity use and 25.7 million metric tons of carbon dioxide. Consumers should treat AI use as an infrastructure issue, while recognizing usage cannot be assigned a personal emissions total.
Key Takeaways
- Harvard researchers examined 403 U.S. hyperscale data centers operating from May 2024 through April 2025.
- The facilities used an estimated 68 to 99 terawatt-hours of electricity during the study period.
- The midpoint electricity estimate represented about 1.8% of total U.S. electricity consumption.
- The study estimated 25.7 million metric tons of associated carbon dioxide emissions.
- Natural gas and coal supplied roughly 54% of the electricity used by the facilities studied.
What did the Harvard study find about AI data center emissions?
The Harvard study found that 403 U.S. hyperscale data centers used enough electricity to represent about 1.8% of total U.S. consumption under its midpoint scenario. The researchers assembled their facility list using satellite imagery and examined centers operating between May 2024 and April 2025. Harvard T.H. Chan School of Public Health published the findings on November 16, 2026.
Hyperscale data centers are large facilities that support cloud computing workloads at an industrial scale. The research matters because consumer-facing AI tools depend on computing infrastructure that draws electricity away from the screen where a person enters a prompt or receives an answer. The study does not assign a specific electricity or emissions total to a single chatbot request, image generation, or consumer account.
AI data center emissions therefore require a broader interpretation than a personal carbon score. A consumer can understand that AI services operate through energy-intensive infrastructure, but the available research does not support calculating an individual user’s precise share. The most sensible conclusion is to distinguish between the system-wide impact documented by the study and an unsupported estimate for one person’s usage.
How much electricity did the 403 data centers use?
The 403 hyperscale data centers used an estimated 68 to 99 terawatt-hours of electricity during the study period. Harvard researchers placed the midpoint scenario at about 1.8% of total U.S. electricity consumption, while the upper estimate approached the annual electricity use of Indiana, Michigan, and Tennessee. The range matters because electricity demand changes with assumptions about facility operation and power use.
| Study measure | Estimated result | Why the measure matters |
|---|---|---|
| Facilities examined | 403 U.S. hyperscale data centers | The analysis covers a large set of operating facilities, not one company or one AI model. |
| Electricity use | 68 to 99 terawatt-hours | The range shows that the researchers modeled uncertainty rather than presenting one falsely exact number. |
| Midpoint share of U.S. electricity use | About 1.8% | The estimate puts data-center demand in national electricity-system context. |
| Associated carbon dioxide emissions | 25.7 million metric tons | The emissions total depends on the electricity sources serving the facilities. |
AI services are one reason public attention has turned toward hyperscale infrastructure, particularly as major platforms add new consumer features. Products such as AI task automation tools can make artificial intelligence more useful in daily routines, but the convenience of those features does not remove the electricity demand at the data-center level. Consumers do not need to avoid all AI services to understand the tradeoff.
How large were the estimated carbon dioxide emissions?
The study estimated 25.7 million metric tons of associated carbon dioxide emissions from the electricity used by the facilities. The peer-reviewed research was published in Nature Sustainability on November 16, 2026. Harvard said the estimate was two to three times the researchers’ earlier estimate for 2018 data-center emissions.
The carbon result reflects the fuel sources behind the electricity, not just the amount of computing performed inside a building. Roughly 54% of the electricity powering the studied facilities came from fossil fuels, including 38% from natural gas and 15% from coal. This matters because two facilities can use similar amounts of electricity while producing different emissions totals if their regional power supplies rely on different fuels.
AI data center emissions should not be treated as a fixed property of every model or app. The study measures a group of facilities over a defined period and estimates emissions associated with their electricity supply. Consumers evaluating an AI feature should understand that faster or more capable services can still depend on a power system whose emissions vary by location and fuel mix.
Which states accounted for the most data-center electricity use?
Virginia, Oregon, Ohio, and Iowa accounted for more than half of the electricity use examined in the study. The concentration matters because data-center growth does not affect every regional grid equally. Communities near large clusters of facilities can face a more direct connection between cloud computing demand and local electricity planning.
The four-state result also shows why national averages can hide regional differences. A nationwide share of 1.8% describes the scale of the demand across the United States, but it does not show where the electricity is generated or which local systems carry the largest demand. State-level infrastructure decisions can therefore have a greater practical effect than a national percentage suggests.
AI data centers are also part of a wider discussion about grid reliability during periods of high demand. Programs focused on reducing power use during grid stress show why timing and location matter alongside total annual consumption. The Harvard study does not evaluate those programs, so readers should not assume that a single corporate initiative changes the reported national estimates.
Why does the study matter for people who use AI tools?
The study matters for AI users because cloud-based tools have physical electricity and emissions costs that are usually invisible in an app. A prompt, generated image, search result, or automated task appears on a phone or laptop, while the computation occurs in remote facilities that require electricity. The research gives consumers a concrete national-scale estimate rather than a vague claim that AI uses energy.
AI data center emissions also matter when companies market new AI capabilities as routine product improvements. A feature such as new AI work and coding tools may provide useful functions, but a product announcement alone does not explain the energy system supporting those functions. Consumers can ask whether a feature solves a meaningful problem before using it repeatedly or enabling it by default.
The practical response is not to treat ordinary AI use as an individual moral failure. The study does not calculate per-user emissions, and consumers cannot choose the power source for a remote facility. Consumers can, however, use AI deliberately, avoid generating unnecessary duplicate outputs, and expect companies to provide clearer information about infrastructure demand and emissions.
What are the main limits of the Harvard data-center estimates?
The Harvard estimates describe 403 U.S. hyperscale facilities operating between May 2024 and April 2025, not every data center in the country and not every workload running inside the facilities. The researchers used satellite imagery to assemble the facility list and reported an electricity-use range of 68 to 99 terawatt-hours. The range is important because it communicates uncertainty instead of implying that data-center consumption can be measured perfectly from outside the facilities.
The study also does not separate AI workloads from other cloud activity. Hyperscale data centers can support many services, including storage, search, video, business software, online retail, and artificial intelligence. Readers should therefore avoid describing the 25.7 million metric ton estimate as emissions caused exclusively by consumer AI tools.
Reuters described the 25.7 million metric ton figure as roughly comparable to Denmark’s annual emissions and quoted lead author Gianluca Guidi saying the work was intended to improve transparency in the AI data-center debate. Reuters’ report provides that comparison. The comparison offers useful scale, but it does not establish that data centers and a country have the same economic activity or emissions sources.
What should consumers and companies do with these findings?
Consumers should use the findings as a reason to evaluate AI features by their practical value rather than assume every automated feature needs to run continuously. AI data center emissions are primarily an infrastructure and corporate planning issue, because individual users do not control facility design, electricity procurement, or regional grid policy. Purposeful use is sensible, but it cannot substitute for transparent reporting from operators and technology companies.
Companies should disclose enough information for customers and communities to understand electricity demand, regional concentration, and progress on reducing emissions. The Harvard study shows why aggregate reporting is valuable: a large national estimate gives the public a clearer basis for discussion than isolated claims about efficiency. At the same time, a company’s statement about cleaner energy should be evaluated against the location and timing of its actual electricity use.
Policymakers and utilities face the separate task of planning for concentrated demand without reducing the question to one technology category. Consumers who use AI in sensitive areas, including health, should also remember that utility does not prove accuracy or safety. An AI health chatbot study may explore potential benefits, but health guidance still requires appropriate evidence and professional care when symptoms or urgent decisions are involved.
FAQ
Did AI data centers use 1.8% of all U.S. electricity?
Harvard researchers estimated that the midpoint scenario for 403 U.S. hyperscale data centers equaled about 1.8% of total U.S. electricity consumption. The estimate covers the study period from May 2024 through April 2025, not every data center or AI workload nationwide.
How much carbon dioxide did the data centers produce?
The study estimated 25.7 million metric tons of associated carbon dioxide emissions. The estimate reflects electricity use and the fuels supplying that electricity, including natural gas and coal.
Did the study measure only artificial intelligence workloads?
The study did not measure only artificial intelligence workloads. The research covered hyperscale data centers, which can support AI alongside cloud storage, search, video, software, and other computing services.
Which states had the highest electricity use in the study?
Virginia, Oregon, Ohio, and Iowa accounted for more than half of the electricity use examined in the study. The result shows that hyperscale data-center demand is concentrated in a limited number of states.
Can a person calculate the emissions from one AI prompt?
A person cannot calculate the emissions from one AI prompt using the Harvard study alone. The research provides facility-level and national-scale estimates, not a verified emissions score for an individual request or account.
