Since 2018, U.S. data center carbon emissions have tripled. Data centers emitted 105 million tonnes of CO2 in the 12 months to August 20242accounting for 2.18% of national emissions (for comparison, domestic commercial airline emissions are approximately 131 million tons). About 4.59% of all energy used in the United States is used in data centers, a number that has doubled since 2018.
Artificial intelligence has been booming since the launch of ChatGPT in November 2022, and it’s difficult to pinpoint the extent to which AI is responsible for this surge. That’s because data centers handle huge amounts of different types of data—in addition to training or pinging AI models, they do everything from hosting websites to storing photos in the cloud. However, researchers say AI’s share is sure to grow rapidly as nearly every sector of the economy attempts to adopt the technology.
“It’s a pretty big surge,” said Eric Gimon, a senior fellow at the think tank Energy Innovation, who was not involved in the study. “There’s a lot of amazing analysis of how fast this exponential growth is going to happen. But in terms of computational efficiency or different types of chips, the business is still in its early stages.”
It’s worth noting that the source of all this power is particularly “dirty.” Because many data centers are located in coal-producing regions such as Virginia, the energy they use is 48% more “carbon intensive” than the national average. The paper, published on arXiv and not yet peer-reviewed, found that 95% of U.S. data centers are built in places with dirtier power sources than the national average.
Falco Bargagli-Stoffi, the paper’s author, said there are other reasons besides being located in coal country. “Dirty energy is available around the clock,” he said, and many data centers need it to maintain peak operations 24-7. “Renewable energy sources like wind or solar may not be available.” Political or tax incentives and local resistance can also influence where data centers are built.
A critical shift in artificial intelligence means emissions from the sector could soon skyrocket. AI models are rapidly moving from fairly simple text generators like ChatGPT to highly complex image, video, and music generators. Until now, many “multimodal” models have remained in the research phase, but that is changing.
OpenAI released its video generation model Sora to the public on December 9, and its website was so flooded with traffic from people eager to test it that the model still doesn’t work properly. Competing models like Google’s Veo and Meta’s Movie Gen haven’t yet been publicly released, but if these companies follow OpenAI’s lead, as they have in the past, they may soon. Music generation models from Suno and Udio are growing (despite lawsuits), and Nvidia released its own audio generator last month. Google is developing Project Astra, which will be a video AI companion that can talk to you about your surroundings in real time.
