Researchers and industry are working to address the challenge of rising energy needs linked to AI and data centers amid concerns around environmental impacts and strain on the energy grid.
UW-Madison’s Wisconsin Energy Institute recently hosted a discussion on this topic, which highlighted the soaring energy requirements linked to increased reliance on artificial intelligence. Machine learning is being used to analyze the avalanche of data being created every day, and Prof. Matt Sinclair said a “massive amount” of computing is needed to handle this task.
Sinclair, an assistant professor of computer sciences at UW-Madison, highlighted the “exponential growth” in the size of AI models seen in recent years.
“Because of that, we just can’t really run these workloads on any one computer. So we’re increasingly splitting them to run across more and more computers in parallel to answer these really complex questions,” he said, adding “this has had huge implications on energy consumption.”
Until around 2018, about 2% of all energy consumption in the United States was attributed to data centers, Sinclair said. But since then, that has increased rapidly. By 2023, that percentage had more than doubled to 4.4%. By 2028, it’s expected that 7% to 12% of all U.S. energy consumption will go to data centers, he explained.
While these data centers aren’t exclusively dedicated to AI, Sinclair said it’s largely fueling the increase.
While some in the field are focused on designing more efficient hardware to reduce energy needs, such as specialized “accelerators” designed to handle AI processing, others are using existing technology designed for graphics processing for machine learning purposes.
Tyler Huebner, a member of Google’s energy market development team for the central U.S. and a former commissioner for the state Public Service Commission, touched on Google’s “tensor processing units,” or TPU. These were developed by the company for neural network machine learning.
Still, Sinclair said making more efficient machines alone won’t solve the problem, as users will just buy more of them.
Other options include designing smaller AI software models, writing “really efficient code” to better handle the data workload, and even applying advanced cooling techniques such as mineral oil immersion.
Huebner also discussed some of Google’s energy-related projects, ranging from scaled-up geothermal energy for a data center in Nevada to locating data centers alongside clean energy resources and small modular nuclear reactors.
“We’ve signed an agreement with a company called Kairos Power to pilot and hopefully build up to 500 megawatts of small modular reactors that could be sited around the country,” Huebner said.
Sinclair added: “The bottom line is, we’re going to have to do something.”
The discussion comes in the wake of multiple major data center projects being announced in Wisconsin, including Microsoft’s planned $3 billion AI facility in Mount Pleasant and a 2,000-acre project in Port Washington. Microsoft also recently acquired land for a smaller data center cluster in Kenosha.
Amy Barrilleaux, communications director for Clean Wisconsin, noted Microsoft’s installation will be the state’s largest electricity user when completed, with a power need equivalent to 300,000 homes. She compared that to adding an entire city’s worth of energy needs to the state over the next few years.
All of these projects are located within the service territory for We Energies, a subsidiary of WEC Energy Group, Barrilleaux noted. She added We Energies recently announced plans for a $2 billion natural gas plant project, tying it to the Microsoft project’s projected energy needs. The natural gas project has yet to be approved by the PSC, she said.
She argued the utility company “doesn’t have a history of being a leader” on clean energy, with less than 3% of its energy mix currently coming from wind and solar.
“This gas plant buildout has a lifespan of 30 years,” Barrilleaux said. “And we know we don’t have 30 years to have a bunch of new fossil fuel burning in Wisconsin, if we want to meet our climate goals.”
In an emailed comment on Barrilleaux’s remarks, a spokesperson for WEC Energy Group pointed to the company’s plans to “transform our power generation fleet” with a balanced mix of wind, solar, energy storage and natural gas.
“Our proposed natural gas generating facilities are an important step to ensure reliability for our customers during the transition to a cleaner energy future in Wisconsin,” the spokesperson said. “Now more than ever — it is critical for us to have quick-start gas plants available and running when the wind doesn’t blow and the sun doesn’t shine.”
These plants will help the company meet “robust demand” along the I-94 corridor, the spokesperson said, noting they’re compliant with the grid operator’s reliability rules and help meet the need for more flexible energy resources. WEC Energy Group has also reduced emissions from electricity generation 54% from 2005 levels, and is investing $9 billion in new renewable energy by 2029.
“Thanks to these investments in new solar, wind and battery storage, we will more than quadruple our carbon-free energy in the next five years,” the spokesperson said.
Costa Samaras, director of the Scott Institute for Energy Innovation at Carnegie Mellon University, noted about 20% of electricity demand growth over the next 15 years is coming from AI and data centers. That’s more than buildings and industry, 15% each, but less than half that of cars, making up about half of the projected demand growth.
“We have to figure out how to manage the data center electricity challenge, because that’s the small part of the electrification challenge,” Samaras said. “And if we can’t figure out that without increasing emissions, we’ve got no chance on cars, buildings and factories.”
Watch the video.