Off-peak power shifting by AI data centers could reduce grid costs and flatten demand curves.

The number of U.S. data centers is growing rapidly to power artificial intelligence programs, raising concerns about environmental consequences and strain on the energy grid. A new study by MIT researchers shows that the impact could vary significantly depending on how data centers structure their energy consumption.

If data centers shift a substantial portion of their energy use to off-peak hours, it could actually lower average energy costs. However, the environmental benefits would vary by location—some regions would likely see greater renewable buildout, while others could experience increased fossil fuel consumption.

"The key with data centers is: How can we add them to the network without adding a lot to our peak usage?" says Christopher Knittel, an economist at MIT Sloan School of Management and co-author of the research. "One way for data centers to do that—to add to average usage but not the peak usage—is if they provide some grid flexibility during those high-cost periods."

The study, "Flexible Data Centers Reduce Power System Costs But Can Increase Emissions," published today in *iScience*, finds that flexible energy consumption could produce cost savings of up to 5% in Texas, 4% in the Mid-Atlantic region, and 2% in western U.S. states. To achieve these savings, data centers would need to shift more than 20% of their consumption—sometimes closer to 50%—to nonpeak hours.

The authors, Juan Ramon L. Senga and Shen Wang, postdoctoral researchers at MIT's Center for Energy and Environmental Policy Research, along with Knittel (the George P. Shultz Professor at MIT Sloan and associate dean for climate and sustainability), used extensive simulations modeling a year of energy use across the U.S. power grid.

The study focused on three grid regions expected to host about 82% of U.S. data centers by 2030: Texas, the Mid-Atlantic region, and the "Western Interconnect" (the 11 large western states). These regions were selected because they are anticipated to account for the vast majority of future data center expansion.

Counterintuitively, the researchers found that adding data centers could lower energy costs in some scenarios. About 60% of typical grid expenses consist of fixed costs like power lines, while 40% are energy costs. By spreading these fixed costs over higher energy volumes, data centers could reduce per-unit costs. However, this only occurs if data centers increase their average consumption faster than their peak-hours consumption, when energy is most expensive.

Most data centers have built-in flexibility, typically operating at about 80% capacity. In the modeling, this flexibility often involves shifting use from early morning and early evening peaks to midday, when energy loads are lower and solar generation peaks. The simulations demonstrate this flexibility makes a measurable difference.

"There are two dimensions that data centers have to make decisions about," Knittel explains. "One is how much of their load in any one time period is flexible. And two, how many hours, plus or minus, can they move that computation?"

Different data center types offer varying flexibility. Those used for AI training consume energy steadily but offer more capacity to shift power loads, whereas inference data centers serving online search queries follow more irregular consumption patterns driven by end-user internet activity.

Knittel emphasizes that the cost savings—ranging from 2% to 7%—are substantial in real terms. "3% is a big number. When you're talking about the grid, 3% or 6% doesn't sound like a lot. But when you're multiplying it by 100 billion dollars, it becomes real money."

On environmental impact, the modeling projects that data center growth by 2030 would significantly increase CO2 emissions compared to no expansion: 58% in Texas, 20% in the Mid-Atlantic, and 24% in the western U.S. However, regional differences matter substantially.

In Texas, where 54% of grid power comes from wind energy, flexible data centers could reduce emissions by 40% by increasing wind energy demand. Conversely, in the Mid-Atlantic region with more solar but less wind capacity, flexible consumption patterns could increase both renewable and fossil fuel use. The modeling suggests a 3% increase in systemwide CO2 emissions, as data centers shift load to hours when solar and wind production decline, allowing coal plants to remain operational.

"When data centers provide some flexibility in that latter scenario, the data centers actually move hours to when sun and wind energy production is slowing, and that allows a coal plant to stay on," Knittel observes. "So it doesn't necessarily attract more renewable investment. It attracts more coal investment."

For these benefits to materialize, data centers must adopt flexible energy-use schedules—but it remains unclear whether companies would voluntarily do so. Knittel suggests regulatory intervention may be necessary. "That's why we have policy," he says.

One key policy tool could be offering faster grid connections in exchange for time-of-use flexibility. "One big concern about these data centers now is how long it takes for them to connect to the grid. One way to provide flexibility now is what's called 'connect and manage,' which is connecting you faster to the grid if you agree to provide flexibility. Tech firms would take that deal. They would rather connect a year earlier and throttle down computation a few hours a day than have to wait." This approach is already used with power plants.

Industry-wide requirements could overcome corporate hesitation. As Knittel notes, while individual tech companies resist flexibility mandates, universal standards across the industry would level the competitive field.

This study represents the first comprehensive "end-to-end" examination of data center implications for both costs and emissions. The researchers plan to continue modeling this critical issue. "Those are two dimensions I think we should all be considering here," Knittel concludes. "The end result is really up to us, and up to policy."