🤖 AI

Assessing the Carbon Emissions and Energy Consumption of U.S. Hyperscale Data Centers

This study estimates that 403 U.S. hyperscale data centers operating between May 2024 and April 2025 consumed 68–99 TWh of electricity, contributing 37–54 million metric tons of CO2 emissions, with a carbon intensity approximately 48% higher than the national grid average due to a heavy reliance on fossil fuels.

Original authors: Gianluca Guidi, Francesca Dominici, Tiziano Squartini, Callaway Sprinkle, Jonathan Gilmour, Kevin Butler, Eric Bell, Scott Delaney, Falco J. Bargagli-Stoffi

Published 2026-06-05

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Gianluca Guidi, Francesca Dominici, Tiziano Squartini, Callaway Sprinkle, Jonathan Gilmour, Kevin Butler, Eric Bell, Scott Delaney, Falco J. Bargagli-Stoffi

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the United States is a giant city, and scattered throughout it are massive, super-powered factories called Hyperscale Data Centers (HDCs). These aren't your average server rooms; they are the colossal engines driving our modern digital life, especially the new wave of Artificial Intelligence. They are so big they could power tens of thousands of homes at once.

This paper is like a team of detectives (researchers from Harvard and other institutions) who decided to take a census of these factories to answer a simple but urgent question: How much electricity are they eating, and how dirty is that energy?

Here is what they found, broken down into simple concepts:

1. The "Energy Appetite"

The researchers tracked 403 of these giant data centers operating between May 2024 and April 2025.

The Analogy: Think of these 403 factories as a single, massive power-hungry beast.
The Finding: This beast ate between 68 and 99 Terawatt-hours (TWh) of electricity. To put that in perspective, that's roughly the same amount of electricity consumed by entire states like Indiana, Michigan, or Tennessee for a whole year. It represents about 1.8% of all the electricity used in the entire United States.

2. The "Carbon Footprint"

Electricity doesn't just appear; it has to be generated, often by burning things like coal or natural gas.

The Analogy: If electricity is the food the data centers eat, then Carbon Dioxide (CO2) is the "exhaust" they leave behind.
The Finding: Because these centers are so hungry, they are responsible for creating between 37 and 54 million metric tons of CO2. That is a massive cloud of exhaust. Compared to a 2018 study, this number has jumped up by about 3 to 5 times, showing how fast the appetite for AI is growing.

3. The "Dirty vs. Clean" Mix

Not all electricity is created equal. Some comes from clean wind and solar, while some comes from burning fossil fuels.

The Analogy: Imagine the power grid is a giant smoothie. Sometimes the smoothie is mostly fruit (clean energy), and sometimes it's mostly coffee (fossil fuels).
The Finding: The data centers in this study are mostly drinking the "coffee" version of the smoothie.
- About 54% of their power comes from fossil fuels (mostly natural gas and coal).
- Only about 25% comes from renewables (wind, solar, hydro).
- The rest is nuclear.
The Result: Because they are located in areas where the local power grid relies heavily on fossil fuels, their "exhaust" is much dirtier than the average American household's. Their carbon footprint is about 48% higher than the national average.

4. The "Where" Matters

The researchers noticed that these data centers aren't spread out evenly like a sprinkler; they are clustered in specific "hotspots."

The Analogy: It's like if all the biggest restaurants in a city were located in just three neighborhoods. Those three neighborhoods would suddenly have to deal with all the traffic and waste.
The Finding: Virginia is the biggest hotspot, hosting 142 of these centers and using about a quarter of the total electricity. Oregon, Ohio, and Iowa are also major hubs. Together, these four states account for more than half of the total energy use in the study.

5. How They Did the Math (The Detective Work)

Since data center companies don't always publish their exact energy bills, the researchers had to be clever detectives:

Satellite Sleuthing: They used satellite images to find the buildings and measure their size, ensuring they were actually data centers and not just office parks.
The "Load" Guess: They didn't know exactly how hard the computers were working every second, so they created four different "scenarios" (like guessing a car's speed: slow, average, fast, or racing). They settled on a "central" guess that seemed most realistic based on how these machines usually run.
The Grid Map: They mapped every data center to the local power company (called a "Balancing Authority") and looked at the "menu" of power plants that company uses. If a data center is in a region where the power plant burns coal, the data center gets "charged" for that coal pollution, even if the data center itself didn't burn the coal.

The Bottom Line

This paper provides a clear, data-driven snapshot of the environmental cost of the AI boom. It shows that while these data centers are the brain of our digital future, they are currently running on a diet that is heavy on fossil fuels, resulting in a significant amount of carbon emissions. The researchers built a public tool (a website) so anyone can see exactly how much energy and pollution is happening in their local area.

In short: The AI revolution is real, but it's currently running on a lot of fossil-fuel power, and it's leaving a very large carbon footprint, especially in a few key states.

Technical Summary: Assessing the Carbon Emissions and Energy Consumption of U.S. Hyperscale Data Centers

Problem Statement
The rapid proliferation of hyperscale data centers (HDCs) in the United States, driven largely by the adoption of artificial intelligence (AI), has raised significant concerns regarding the industry's environmental footprint. While data centers are the backbone of modern digital infrastructure, their energy-intensive nature—stemming from computational power and cooling requirements—has led to increased electricity demand and associated greenhouse gas emissions. Existing literature presents a range of estimates for data center energy use, but peer-reviewed studies specifically examining the emissions of hyperscale facilities are scarce. Previous estimates, such as those by Siddik et al. (17) for 2018, are now outdated given the surge in demand from AI workloads and cryptocurrency mining (which is excluded from this analysis). Furthermore, there is a lack of facility-level data that integrates heterogeneous sources, validates infrastructure through satellite imagery, and attributes emissions to specific power plants using the most recent grid data.

Methodology
The authors developed a scalable, data-driven pipeline to estimate electricity consumption and attributable CO2 emissions for 403 U.S. hyperscale data centers operating between May 2024 and April 2025. The methodology consists of five key steps:

Data Collection and Validation: The study compiled a dataset of 675 potential HDCs from private providers (Baxtel). This dataset was rigorously validated using high-resolution satellite imagery (Esri World Imagery) and building footprint data from OpenStreetMap (OSM). Through manual verification and geocoding, the authors confirmed the location and identity of 403 facilities, obtaining square footage data for all of them.
Power Capacity Estimation: For the 6 facilities lacking reported power capacity, a Gradient Boosted Regression Tree (GBRT) model was trained to impute missing values. The model utilized building square footage, climate type (Köppen-Geiger classification), and balancing authority region as predictors. The model achieved an $R^2$ of 0.807 on a random 85/15 hold-out split, with grouped validation schemes confirming its utility as an imputation tool rather than a general predictive model for new siting.
Electricity Consumption Modeling: Annual electricity consumption was estimated by multiplying the total facility electrical capacity (including IT, cooling, and overhead) by annual operating hours (8,760) and a facility-load coefficient ( $u$ $u$ ). Recognizing the uncertainty in translating nameplate capacity to actual operational load, the authors reported results across four physically motivated scenarios:
- Low-load ( $u=0.48$ ): Conventional servers, rated-max design.
- Central ( $u=0.58$ ): Conventional servers, measured-max design, consistent with bottom-up power-flow modeling and independent PUE-implied checks.
- Intermediate-high ( $u=0.663$ ): Server-level effective utilization applied to facility capacity (retained for continuity with prior studies).
- AI-weighted high ( $u=0.70$ ): Reflecting higher utilization for AI-specialized facilities.
Attributional Emissions Modeling: Each HDC was geolocated to its corresponding balancing authority (BA). An energy-generation-weighted (EGW) attributional model was applied, assuming that within each BA, the total HDC electricity demand is distributed across power plants in proportion to their annual net electricity generation. This approach utilizes EPA eGRID2023 plant-level data (the most recent complete release at the time of revision) to assign fuel mixes and emission factors.
Carbon Intensity Calculation: The study calculated the weighted average carbon intensity (gCO2/kWh) for HDCs by aggregating the emissions and load attributed to the supplying power plants, comparing these metrics against the U.S. national grid average.

Key Contributions

Comprehensive Dataset: The creation of a validated, facility-level dataset of 403 U.S. HDCs, representing an estimated 56–82% of the national hyperscale electricity footprint.
Attributional Framework: The application of a transparent, attributional accounting method that links specific data centers to the physical generation mix of their balancing authorities, distinguishing this work from consequential (marginal) approaches.
Scenario-Based Estimation: The provision of a range of estimates (68–99 TWh electricity; 37–54 Mt CO2) rather than a single point estimate, explicitly accounting for uncertainties in facility load factors and incomplete coverage.
Public Tool: The development of a public-facing web platform allowing users to explore carbon footprints, electricity loads, and fuel mixes at the state and balancing authority levels.

Results

Electricity Consumption: Under the central scenario ( $u=0.58$ ), the 403 HDCs consumed approximately 82 TWh of electricity, representing about 1.8% of total U.S. electricity consumption. Across all scenarios, consumption ranged from 68 to 99 TWh (1.5–2.2% of the national total).
Carbon Emissions: The associated CO2 emissions ranged from 37 to 54 million metric tons (Mt), with a central estimate of 45 Mt. This is approximately 3.5 to 5 times higher than the 10.5 Mt estimated for HDCs in 2018.
Geographic Distribution: Virginia hosted the most HDCs (142) and the highest aggregate consumption (~21 TWh), followed by Oregon, Ohio, and Iowa. Together, these four states accounted for over 50% of the total HDC electricity consumption in the sample.
Fuel Mix and Carbon Intensity: Approximately 53.9% of the electricity attributed to HDCs was supplied by fossil fuels (38.2% natural gas, 15.2% coal), 20.9% by nuclear, and 25.3% by renewables. The HDC electricity-weighted average carbon intensity was approximately 545 gCO2/kWh, which is about 48% higher than the contemporaneous U.S. national grid average of 370 gCO2/kWh. Notably, 90% of HDC electricity demand is located in balancing authorities with carbon intensities above the national average.

Significance and Claims
The paper claims to provide a robust, attributional tool for assessing the environmental footprint of hyperscale data centers using the most recent EPA eGRID plant-level data. The authors emphasize that their approach offers a statistically robust assessment due to the distinctive visibility and well-documented infrastructure of HDCs. They argue that their findings highlight the disproportionate concentration of data centers in regions with carbon-intensive grids, a critical factor often overlooked in broader sectoral analyses. The study underscores the urgency of data-driven decision-making to inform policies aimed at mitigating the environmental impact of data centers, particularly as AI adoption drives further expansion. The authors explicitly note that their estimates are bounded scenario-based figures rather than strict lower bounds, acknowledging limitations such as incomplete facility coverage and the lack of facility-level PUE data. They position their work as a necessary step in recognizing data centers as a relevant economic sector requiring targeted policy attention, while distinguishing their facility-level attribution from server-level or AI-specific emission projections found in other literature.