The number of U.S. data centers is growing, largely to power artificial intelligence programs. That has led to concern about the environmental consequences of data centers — and their impact on the energy grid itself. What will happen if scores of new data centers come online?
A new study by MIT researchers indicates that the impact of data centers could vary significantly, depending on how their energy use is structured.
Specifically, if data centers move a significant portion of their energy consumption to non-peak hours, it might actually help lower average energy costs. The environmental impact, in terms of type of energy consumed, would differ by location, with some places likely seeing a greater buildout of renewables and others experiencing a relative increase in fossil fuel use.
“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 in the MIT Sloan School of Management and co-author of a new paper detailing the study. “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. And that’s what we’ve been interested in understanding.”
Specifically, the paper finds that a flexible arrangement for data-center energy consumption, compared to an inflexible one, would produce cost savings of up to 5 percent in Texas, 4 percent in the Mid-Atlantic region, and 2 percent in the western U.S. states. To achieve that, data centers would have to move more than 20 percent of their consumption — sometimes more like 50 percent — to non-peak hours.
The paper is titled “Flexible Data Centers Reduce Power System Costs But Can Increase Emissions,” and appears today in the journal iScience. The authors are Juan Ramon L. Senga, a postdoc in MIT’s Center for Energy and Environmental Policy Research; Shen Wang, a postdoc in MIT’s Center for Energy and Environmental Policy Research; and Knittel, who is the George P. Schultz Professor at MIT Sloan and the associate dean for climate and sustainability at MIT.
The 20 percent solution
The expansion of data centers has raised questions about additional stress for the U.S. grid, the global effects of increased fossil-fuel consumption, and the local environmental effects of data centers. The current study examines the first two of these issues.
To conduct the research, the scholars extensively simulated scenarios in which data centers expand, using the so-called “Gen X” model of the U.S. power grid, for a year’s worth of energy use.
The study focused on the grid systems in three areas: Texas, the Mid-Atlantic region, and the “Western Interconnect,” comprising the 11 large western states in the lower 48 states of the U.S. The researchers studied these regions because they collectively host most of the country’s data centers — about 82 percent of U.S. data centers by 2030, according to one analysis.
A bit counterintuitively, the researchers found that adding data centers could lower energy costs in some scenarios. Typically, about 60 percent of grid expenses are fixed costs, like power lines, while about 40 percent consists of energy costs. Adding data centers to the grid could, in effect, apportion the fixed costs over a higher volume of energy use.
“It’s really just math,” Knittel says.
But there is a catch. Lower costs might only happen if data centers increase their average consumption faster than their peak-hours consumption, when energy is most expensive. As it happens, most data centers do have flexibility built into their energy-use patterns, since they usually run at about 80 percent capacity.
In the study’s modeling, that flexibility often consists of shifting use from early-morning and early-evening peaks, to more midday energy consumption, when the energy load is lower and solar is at full capacity. The simulations show this makes a difference.
“There are two dimensions that data centers have to make decisions about,” Knittel says. “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?”
Pretty soon, real money
Additionally, data centers have different amounts of flexibility based on the types of AI-related computation they host. Data centers being used for AI training data tend to consume energy at a steady rate, but as a result could provide more flexibility for shifting power loads compared to inference data centers, which are used more for online search queries. In the latter case, consumption is driven more by end-user Internet habits.
Overall, Knittel emphasizes, the magnitude of cost savings suggested by the study, ranging from 2 percent to 7 percent, is significant.
“Three percent is a big number,” Knittel says. “When you’re talking about the grid, 3 percent or 6 percent doesn’t sound like a lot. But when you’re multiplying it by 100 billion dollars, it becomes real money.”
When it comes to environmental impact, the modeling finds that the projected level of data center growth by 2030 would be very significant in terms of carbon dioxide emissions. Compared to a world with no data center growth, the study finds those emissions would rise by 58 percent in Texas, 20 percent in the Mid-Atlantic region, and by 24 percent in the western U.S. That underscores the need to be strategic about data center consumption.
But the modeling also finds that the implications of data center buildout for clean-energy use vary by region. In Texas, where 54 percent of grid power is wind energy, having more data centers with flexible patterns of energy use could reduce emissions, by increasing demand for wind energy. The study finds that in this scenario, there could be 40 percent fewer CO2 emissions.
However, in the Mid-Atlantic region, where there is a reasonable amount of solar energy but relatively less wind power, more data centers with flexible consumption patterns could increase both renewable energy and fossil-fuel energy consumption. Here the modeling suggests an increase in CO2 emissions system-wide of 3 percent.
“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.”
“That’s why we have policy”
For any of this to happen, however, the data centers would have to implement the flexible energy-use schedules modeled in the study. And it’s not clear that companies using data centers would be motivated to do that. To Knittel, this suggests officials might have to craft regulations in this area.
“That’s why we have policy,” Knittel says.
More specifically, he adds, there is one big policy lever officials could use to achieve this goal: offering quicker initial hookups to the grid in return 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,” Knittel says. “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 to have to wait. We do this with power plants too.”
Certainly, Knittel adds, as firms competing with each other, “Tech companies say they won’t provide flexibility alone. But if everyone in the industry has to, it’s okay.”
The current study is the first to examine the “end-to-end” implications of the centers for costs and emissions. The results, the scholars feel, bear further evaluation — and it is a topic they are continuing to model.
“Those are two dimensions I think we should all be considering here,” Knittel says. “The end result is really up to us, and up to policy.”
The research received support from the Future Energy Systems Center of the MIT Energy Initiative.
