Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseInsight · On RadarThe Verge→ original

Giant US Data Centers Use 82 TWh a Year — More Than Each of 148 Countries

Giant US data centers, built mostly for AI, consumed about 82 TWh (82 billion kilowatt-hours) in a year. That is 1.8% of all the country's electricity, and more than each of 148 countries, including Austria and Israel. Gas and coal supplied about 54% of that energy, so emissions reached 25.7 million tonnes of CO2, comparable to Denmark's. The practical takeaway: when choosing a region for a GPU cluster, look at the local power grid. Harvard found the 403 data centers using satellite imagery, because operators do not disclose their data. The study also does not account for on-site gas turbines or local air pollution.

AI-processed from The Verge; edited by Hamidun News
Giant US Data Centers Use 82 TWh a Year — More Than Each of 148 Countries
Source: The Verge. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceThe Verge ↗
Editor-in-Chief
Key metric #1In Wyoming (coal), a kilowatt-hour produces 7 times more CO2 than in Oregon (hydro)
Key metric #2Harvard found 403 data centers from satellite images: operators stay silent about their data
Key metric #3The estimate does not include on-site gas turbines or local smog

Picture two identical GPU clusters running the same model and drawing the same amount of electricity. One sits in Wyoming, where the grid leans heavily on coal; the other in Oregon, where hydropower is abundant. The first will emit roughly seven times more CO2.

That is the subject of a new study in Nature Sustainability: in the US, giant data centers ("hyperscale" facilities, buildings spanning millions of square feet that are built mostly for generative AI) consumed about 82 TWh in a year. A terawatt-hour is a billion kilowatt-hours, so that is 82 billion kWh, or 1.8% of all the country's electricity.

Only a minority of countries consume more; each of roughly 148 nations, including Austria, Israel and Venezuela, consumes less. If you are choosing a cloud region or planning your own cluster, this bears directly on your carbon reporting.

What Was Counted and What They Found

A team led by Gianluca Guidi at the Harvard T.H. Chan School of Public Health used satellite imagery to find 403 hyperscale data centers in the US. They estimated each facility's consumption from its electrical capacity. Then, using Environmental Protection Agency (EPA) data, they estimated how much carbon the local grid supplying each data center would have emitted between May 2024 and April 2025.

The result: about 82 TWh over the year. According to calculations by John Sterman, director of MIT System Dynamics and the MIT Sloan Sustainability Initiative, who was not involved in the study, that is more than about 148 countries consume in a year, which is most of the world's nations. The largest economies, including the US and China, are not on that list.

Emissions came to 25.7 million metric tonnes of CO2. The authors compare that with Denmark's annual carbon footprint, while Sterman notes that it exceeds the combined emissions of about 113 countries. The reason is simple: about 54% of the electricity for these sites likely came from gas and coal.

Location Decides: Wyoming vs.

Oregon

The most useful part of the study is the regional spread. Electricity generation in coal-reliant Wyoming produced roughly seven times more CO2 per kilowatt-hour than in hydro-rich Oregon. Guidi says two data centers with identical consumption can have completely different carbon footprints purely because of where they are located. He finds that encouraging: since the consequences are not predetermined, we have a choice about where to build and what to power the infrastructure with.

For practitioners, the takeaway is this: the region you pick for training and inference is a lever that, all else being equal, can deliver more than many code optimizations. If your company has climate commitments or reports to investors and customers, the question "which grid powers our GPUs?" belongs on the checklist next to latency and price.

Cornell professor Fengqi You, who was also not involved in the study, points to the risk of "long-term carbon lock-in." The power plants and grids built to meet growing data center demand operate for decades, so decisions made now will determine emissions far into the future. Sterman adds that emissions are irreversible: CO2 that has been emitted cannot be taken back. He proposes requiring operators to bring new clean generation online.

Limitations and Transparency

The figures remain an estimate. Operators do not disclose their actual energy consumption, so the researchers had to take the long way around, working from satellites and electrical capacity. The study does not account for local air pollution from new gas plants built to meet data center demand, nor for the gas turbines that some AI companies install directly on-site to avoid depending on the grid. In addition, hyperscale centers serve more than AI and make up only a fraction of the thousands of data centers in the country, a number that continues to grow.

An interesting detail: Sterman himself used Claude to match the study's figures against countries' consumption and emissions, and said so openly. In his words, AI can be extremely useful, but we need to find a way to limit the harm.

To close the data gap, Guidi and his colleagues are building a website to track the environmental impact of individual data centers. Separately, a bipartisan bill moving through Congress could require operators nationwide to disclose their electricity consumption. If it passes, developers and product teams will get more reliable data for choosing cloud sites and for emissions reporting.

The practical bottom line: AI's energy footprint is not a fixed property of the technology but the result of infrastructure decisions. Ask your cloud providers which grid powers the region you need, and do not accept averaged "green" assurances without a site-by-site breakdown.

Story

Key Questions Answered

What Was Counted and What They Found?

A team led by Gianluca Guidi at the Harvard T.H. Chan School of Public Health used satellite imagery to find 403 hyperscale data centers in the US. They estimated each facility's consumption from its electrical capacity. Then, using Environmental Protection Agency (EPA) data, they estimated how much carbon the local grid supplying each data center would have emitted between May 2024 and April 2025. The result: about 82 TWh over the year.

What are the key details on Location Decides: Wyoming vs. Oregon?

The most useful part of the study is the regional spread. Electricity generation in coal-reliant Wyoming produced roughly seven times more CO2 per kilowatt-hour than in hydro-rich Oregon. Guidi says two data centers with identical consumption can have completely different carbon footprints purely because of where they are located. He finds that encouraging: since the consequences are not predetermined, we have a choice about where to build and what to power the infrastructure with.

What are the key details on Limitations and Transparency?

The figures remain an estimate. Operators do not disclose their actual energy consumption, so the researchers had to take the long way around, working from satellites and electrical capacity. The study does not account for local air pollution from new gas plants built to meet data center demand, nor for the gas turbines that some AI companies install directly on-site to avoid depending on the grid.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…