The electricity needs of data centers, especially those powering artificial intelligence (AI), could double worldwide by 2030, according to a new report released this month from the International Energy Agency (IEA).

In the U.S., a major AI powerhouse, the forecast looks particularly stark. The IEA suggests that the country’s AI data centers will consume more power than the production of energy-intensive materials— including aluminum, steel, cement and chemicals combined— in the next five years.

This explosion in energy demand strains power grids and threatens progress toward the U.S. government’s goal of 100% carbon-free electricity by 2035, said Kyri Baker, associate professor in the Department of Civil, Environmental and Architectural Engineering.

But Baker and her colleague, Bri-Mathias Hodge, professor in the Department of Electrical, Computer & Energy Engineering, are looking for solutions. They suggest that if future data centers are placed in the right location and equipped with energy storage technologies, they can run on 100% clean energy.

“We're still figuring out how AI data centers are going to impact the grid and emissions,” Baker said. “But one thing is certain: AI will continue to increase and be pervasive in our everyday lives. We need to start making plans to build and run these data center in a sustainable way.”

Conflicting demands

Much of the energy demand of AI data centers comes from training AI models, particularly the large ones that can write essays or mimic human conversations like ChatGPT, and cooling computer servers.

While asking ChatGPT a question can consume almost 10 times the amount of electricity as a simple Google search, that pales in comparison to the energy needed to train these models in the first place, according to the IEA.

Inside a data center. (Credit: Eimantas Buzas/Adobe Stock)

Today, a large-scale AI data center with a maximum power demand of 100 megawatts consumes roughly the same energy as 100,000 homes for a year, said Hodge.

“We want to lower emissions and cost, but also be competitive in AI research and development,” Baker said. “These are competing objectives, and it’s hard to say which one should dominate the other.”

Tech companies, who have to invest in costly specialized computer chips to train AI, often run these centers round-the-clock to maximize their return on investment. Because the wind doesn’t always blow, and the sun doesn’t always shine, they can’t rely solely on wind and solar power.

So some are turning to alternative sources. For example, Microsoft announced that it will use natural gas to power one of its data centers. Amazon recently purchased a data center next to a nuclear power plant.

But building natural gas and nuclear power plants is expensive and can take years. To meet the imminent demand, data center and grid operators are increasingly tempted to rely on fossil fuels, including coal, the most carbon-intensive option, notes Hodge.

In fact, rising demand has already delayed coal plant retirements across the country. In 2024, the U.S. retired only 4 gigawatts of coal-fired power generation capacity, a steep decline from the nearly 10 gigawatts retired annually over the past decade, according to the U.S. Energy Information Administration.

“We're shooting after a moving target here with decarbonization,” Hodge said. “A couple years ago, we were talking about how to fulfill the additional demand from electric vehicles with renewables, and now we have this AI boom that immediately dwarfs EV power needs.  We're not moving fast enough toward our climate goal, because the goalposts keep shifting.”

New technologies

Baker and Hodge, both fellows at the Renewable and Sustainable Energy Institute, are exploring solutions that can help fulfill massive power demands while safeguarding the grid and keeping carbon emissions targets in check.

They propose adding energy storage systems to the grid and data centers.

“Storage can help the grid become more resilient to power fluctuations,” Baker said.

“It also helps us lower emissions, because we can store excess renewables and use them later.”

Many states, including California and Texas, have been installing large lithium-ion batteries on the grid. But these batteries can only store enough power for up to four hours. Baker and Hodge are evaluating the costs and benefits of various storage devices that can last more than 10 hours, including hydrogen storage.

The idea is that when wind or solar produce more electricity than the grid can immediately use, utility companies can use the excess energy to split water into hydrogen and oxygen, and then store the hydrogen in large tanks. Later, when the power demand outpaces what’s generated, the hydrogen can react with oxygen to generate electricity, producing water as the byproduct and emitting no carbon.

These storage systems can also supplement nuclear or natural gas plants during outages or when they need to undergo maintenance.

For now, building energy storage facilities remains costly, so the team is calculating the most cost-effective mix of power sources and storage strategies.

“We're also looking at whether the storage facilities can sell electricity to the grid and potentially make some revenue, while at the same time providing services to the data centers,” Baker said. 

New AI hubs

In addition, the team is exploring the best places to build future data centers, based on availability of renewable energy, water for cooling and existing power transmission infrastructure.

This can help lower the upfront costs for AI companies, avoid exhausting the grid and lower environmental impact, Hodge said.

“If you want to prevent the worst impacts of climate change, the power system is the very first thing we need to decarbonize. Many other industries, like concrete and steel manufacturing, are much harder. Unfortunately, we're still taking baby steps,” said Hodge.

Still, both researchers are hopeful that the right investments and policies can steer AI growth in a more sustainable direction.