The current AI boom is consuming vast amounts of energy, particularly in the United States. A report from the Harvard T.H. Chan School of Public Health examined 2,132 data centers in the U.S., which amounts to 78% of all such facilities in the country. These centers, filled with rows of servers, are used to train AI models and handle requests from applications like ChatGPT. They require significant energy both for operation and cooling.
Since 2018, the carbon emissions of U.S. data centers have tripled. In the year leading up to August 2024, these centers were responsible for 105 million tons of CO2 emissions, representing 2.18% of the nation’s total emissions. For comparison, domestic airlines account for about 131 million tons. Moreover, 4.59% of the total energy consumed in the U.S. is used by data centers, a figure that has doubled since 2018.
The growth of AI, especially since the launch of ChatGPT in November 2022, has contributed to this increase. Data centers process various types of data, from website hosting to cloud photo storage, but AI’s share is rapidly growing as industries adopt the technology.
Eric Gimon, Senior Fellow at Energy Innovation, notes that while there are many analyses on the rapid growth of AI, the industry is still in its early stages regarding efficiency and chip technology.
Most data centers in the U.S. are located in regions with high coal production, like Virginia, resulting in a “carbon intensity” of energy use 48% higher than the national average. The study, published on arXiv, found that 95% of U.S. data centers are in areas with dirtier energy sources than the national average.
Falco Bargagli-Stoffi, one of the study’s authors, explains that fossil fuels are available around the clock, which is necessary for data centers that require continuous operation. Renewable energy sources like wind or solar may not be as consistently available. Political and tax incentives, as well as local opposition, also influence where data centers are built.
The future of AI suggests that emissions in this sector will rise significantly. AI models are evolving from simple text generators like ChatGPT to complex image, video, and music generators. While many of these “multimodal” models are still in research, this is changing. OpenAI recently introduced its video creation model, Sora, to the public. Other companies like Google and Meta are expected to follow suit with their models, such as Veo and Movie Gen. Music generation models by Suno and Udio are gaining popularity, and Nvidia has launched its audio generator. Google is working on the Astra project, an AI video companion that interacts in real-time.
Gianluca Guidi, a Ph.D. student in artificial intelligence and the study’s lead author, points out that as AI shifts to images and videos, data size grows exponentially, leading to a surge in emissions.
The researchers aim to develop a reliable method to capture data center energy consumption snapshots. This task is complex due to the scattered data across various sources and authorities. However, they have created a portal to display emissions from data centers nationwide. The long-term goal of this data pipeline is to support future regulatory efforts to curb data center emissions, which are expected to increase significantly in the coming years.
Francesca Dominici, Director of the Harvard Data Science Initiative, anticipates growing pressure between the environmentally conscious community and Big Tech but does not foresee regulation in the next four years.