The AI boom is not only changing software. It is also changing the energy profile of the data centers that train models and respond when people use tools like ChatGPT.
A new paper from teams at the Harvard T.H. Chan School of Public Health and UCLA Fielding School of Public Health gives a clearer view of the scale. The researchers examined 2,132 data centers operating in the United States, representing 78% of all facilities in the country.
What the new data shows
Data centers are the buildings that hold rows of servers. Those servers are used for many tasks, including training AI models, handling user requests to AI systems, hosting websites and storing cloud files.
They need large amounts of electricity for two basic reasons: running the computing hardware and keeping that hardware cool. As more digital work moves into these facilities, their energy footprint grows with it.
Since 2018, carbon emissions from data centers in the US have tripled. For the 12 months ending August 2024, data centers were responsible for 105 million metric tons of CO2. That accounted for 2.18% of national emissions.
The source article notes one comparison: domestic commercial airlines are responsible for about 131 million metric tons. Data centers are not at that level, but the gap is no longer abstract.
Energy use has climbed as well. About 4.59% of all the energy used in the US goes toward data centers, a figure that has doubled since 2018.
Why AI’s exact share is hard to isolate
The paper does not put a single number on how much of the increase comes from AI specifically. That is because data centers do not serve only one purpose.
The same broad infrastructure can support AI model training, user requests to models, websites, cloud storage and other forms of data processing. That makes it difficult to separate AI emissions cleanly from the wider data center footprint.
Still, the researchers say AI’s share is growing rapidly. The reason is straightforward: nearly every segment of the economy is trying to adopt the technology, and wider adoption means more work flowing through data centers.
Eric Gimon, a senior fellow at the think tank Energy Innovation who was not involved in the research, described the increase this way: “It’s a pretty big surge.” He also cautioned that the industry is still early in figuring out efficiencies and different kinds of chips.
The electricity mix matters
The location of data centers is central to their emissions. Many are located in coal-producing regions, including Virginia. As a result, the carbon intensity of the energy they use is 48% higher than the national average.
The paper, published on arXiv and not yet peer-reviewed, found that 95% of data centers in the US are built in places with electricity sources that are dirtier than the national average.
Falco Bargagli-Stoffi, an author of the paper and Assistant Professor at UCLA Fielding School of Public Health, said there is another practical reason these facilities may rely on dirtier power. “Dirtier energy is available throughout the entire day,” he said.
That matters because many data centers need to maintain peak operation 24-7. Bargagli-Stoffi added that renewable energy, like wind or solar, might not be as available. Political or tax incentives and local pushback can also influence where data centers are built.
Multimodal AI raises the stakes
The next phase of AI could intensify the problem. The field is moving beyond fairly simple text generators like ChatGPT toward more complex systems that generate images, video and music.
Until recently, many multimodal models were still in the research phase. That is changing. OpenAI released its video generation model Sora to the public on December 9, and its website was flooded with traffic from people trying to test it, leaving it still not functioning properly.
Other companies have similar systems that have not yet been released publicly. The source article names Veo from Google and Movie Gen from Meta, and notes that they might arrive soon if those companies follow OpenAI’s lead as they have in the past.
Music generation is also expanding. Models from Suno and Udio are growing despite lawsuits, and Nvidia released its own audio generator last month. Google is also working on Astra, described as a video-AI companion that can converse with users about their surroundings in real time.
The emissions concern is tied to the size and complexity of the data involved. Gianluca Guidi, a PhD student in artificial intelligence at University of Pisa and IMT Lucca and visiting researcher at Harvard, is the paper’s lead author. He explained the shift plainly: “As we scale up to images and video, the data sizes increase exponentially.”
What researchers want to measure next
One goal of the project was to create a more reliable way to capture how much energy data centers use. That has been difficult because the relevant data is spread across several sources and agencies.
The researchers have built a portal showing data center emissions across the country. The longer-term goal is to create a data pipeline that can inform future regulatory efforts to curb emissions from data centers, which are predicted to grow enormously in the coming years.
Francesca Dominici, director of the Harvard Data Science Initiative, Harvard Professor and another coauthor, expects pressure to increase between environmental and sustainability-conscious communities and Big Tech. Her forecast, however, is cautious: “But my prediction is that there is not going to be regulation. Not in the next four years.”
For now, the takeaway is clear. AI’s energy demand is becoming part of the larger data center emissions story, and the shift from text to richer media could make that story much larger.