AI’s Energy Appetite: Environmental impacts and governance

As AI usage grows, so too does the energy demand on AI-related infrastructure. Experts in energy and energy policy discussed the environmental ramifications during the 2025 Vanderbilt AI Symposium.

In a panel titled AI Energy Appetite: Environmental Impacts and Governance, Ethan Thorpe, BA'24, a fellow at Vanderbilt Law's Private Climate Governance Lab, moderated a discussion between Vanderbilt Law Professor Michael Vandenberg, Dr. Leslie Abrahams, and Dr. Jonathan Gilligan.

Thorpe kicked off the session by stressing the importance of the subject on a national and international level.

"AI and the infrastructure that it powers have quickly become the inspiration for trillions of dollars in investment," he said.

Panelists offered an overview of the energy implications of AI. According to Abrahams, AI energy consumption is expected to jump from 4.5% of total U.S. electricity demand in 2023 to as much as 12% by 2028, an increase that represents the current demand of Texas or California.

"It was always the plan [to expand energy infrastructure]… (but) this expansion is happening a lot faster than anyone expected," Abrahams said. "It's also happening in a greater scale, so it's having a regional impact."

Vandenberg emphasized a few challenges with the expansion of AI-related energy use and demand.

"Electric utilities and others have an incentive to overpredict electricity demand in the future; if we overbuild, then we get a return on our investments," Vandenberg said. "The challenge is that we are building [the] power supply in a way that provides electricity that we can use to meet our very highest demands. So, one of the interesting things is not how much total energy we are using, it's when we are using it, what are those characteristics, and how can we shift those a little bit?"

The conversation then shifted to a discussion on what it means to invest in new and evolving energy expansion technology that may not pay off in the future. Gilligan explained how this uncertainty applies to both the quantity of energy needed and the way that the utilities are going to build out their infrastructure. He noted a prevailing statistic offered by OpenAI that a query conducted by a large language model takes 1/10 th of a Watt hour.

"Open AI says they serve about a billion queries a day. If that 10 th of a Watt hour per query number is right, that comes to about 100 million watt hours per day; divide that by 24 hours, and you need to come to about 5 million watts that it would take to serve the global demands for Open AI - that's 5% of a typical power plant," Gilligan said. "We would not be needing to build out power if these energy consumption estimates that people want to give us were even remotely true."

Gilligan also discussed the type of energy that utilities will rely on for expansion. If they turn to fossil fuels and gas, it could have disastrous consequences for the environment; if they turn to renewable alternatives, the positive impact could be momentous. The direction should be aided by monetary incentives, Gilligan said, since opening fossil fuel and coal plants is more expensive than renewable alternatives in terms of opening and shutting down.

"[The] huge growth of energy needed could potentially be beneficial in improving the nation's energy infrastructure, which is sadly out of date and vulnerable," Gillian said. "We could be building our capacity to transmit large amounts of energy from wind turbines and solar panels - and that would be a real benefit."

Lastly, the conversation shifted to private governance of AI - the idea that private organizations are filling gaps in climate adaptation, mitigation, and other environmental issues that you would typically see governments filling. This is seen in terms of energy and AI structures through company contracting requirements outlining a specific portion of renewable sources or a limit on emissions.

"Because of this contracting phenomenon, we've seen that 80% of the largest 10 firms in seven global sectors are already imposing environmental requirements on their supplies," Vandenberg said. "This will begin to kick in as it relates to AI, if those employers get attributed to the cost of the environmental impact of their employees' AI use."