How do you decide where to put a data center that will consume as much power as a small city?
Data centers have existed to support early computers and the internet for decades. Every data center in the U.S. draws power from a regional electrical grid and water from a local watershed—costs that are usually calculated after they’re built, if at all.
But in the last five years, the boom of artificial intelligence has driven up the energy demand. Hyperscale data centers, designed to handle AI and cloud computing, consume three to five times more power per square foot than traditional data centers, with large facilities drawing as much power as nearly 100,000 homes and millions of gallons of water daily. But as technology companies scour the country for their next site, cities have often been resistant to data center expansion for two reasons: strain on the local power grid and negative health effects of pollution on nearby residents.
Now, a Harvard biostatistician and an electrical engineer have combined their expertise to build a framework for helping policymakers decide where data centers should go.
Francesca Dominici, the Gamble professor of biostatistics, population, and data science at the Chan School of Public Health, has spent her career studying how air pollution affects human health. She’s collaborating with Le Xie, the McKay professor of electrical engineering at the Paulson School of Engineering and Applied Sciences, who studies the impact of large-scale systems on the electrical grid. (Xie is also a member of the board of managers of the power grid operator PJM Interconnection, though he notes his views on the project don’t reflect the views of the company.)
Together, the two are using collected data and modeling to predict the energy and health impacts of a possible data center before its final location is chosen.
Though the project is still in its early stages, the researchers say they aim to build the framework in three parts: a data pipeline that tracks the resource consumption and emissions of existing data centers; models that estimate the health effects on surrounding communities due to data center-driven pollution; and an interactive toolkit that developers, utilities, and local leaders can use to plan future sites.
“[Data center expansion] is very political. It’s becoming very contentious,” Dominici said. “So, the project itself has the aspiration of tackling this issue from a scientific, rigorous, and unbiased perspective.”
Understanding the intersection of energy demand and health effects is critical to assessing the overall impact of a data center, Dominici and Xie explained. But often, the stakeholders involved in approving and building a facility span industries. Grid operators may not fully understand environmental impacts, Xie explained, while environmental and public entities may not fully understand the electrical side—and that gap can make deliberations more difficult.
Dominici and Xie hope that, by combining their respective areas of expertise, the project will help bridge knowledge gaps across industries to make the process of developing data centers easier for the parties involved.
The work begins with data collection from Dominici’s team, which she says is focused on building “the most comprehensive and validated data set” of existing and planned hyperscale data centers.
Researchers are gathering publicly available data from permitting documents and satellite images, but Dominici noted that it is often much harder to track down the energy use and emissions from data centers that are built “behind the grid,” using private generators, rather than drawing from the regional grid. To collect as much data as possible, the group is also purchasing proprietary data on existing facilities.
After building their dataset, Dominici and Xie said they aim to use models to describe—and predict—the energy demand, emissions, air pollution, and health effects of data centers on surrounding communities. The goal, Dominici said, is “to empower both ends of the spectrum, both the community, the regulators, the utilities, and the developers, on a way forward.”
Xie said “one of the major bottlenecks” for AI data center development is the inflexibility of the grid, which makes it difficult for data centers to draw enough reliable power. His team is working to model the impact of data centers on the grid, looking at both the high amount of constant energy they draw as a daily baseline and the sudden power spikes that occur when computing demand surges. They are also looking at ways to make data centers more flexible so they can adjust their power draw in ways the grid can absorb.
Dominici and Xie then hope to translate their findings into usable estimations of a given data center’s impact that developers, residents, and city officials can reference as they make decisions over where to place facilities.
“Le and I are very united in the goal of putting something in the hands of the policymaker, the center developer, and the community, that allows them to make progress on these highly controversial issues right now,” Dominici said.
They hope to make the framework usable on a local level, aiming to estimate the health effects of emissions and pollution roughly by zip code.
“The ultimate ambition will be to be as granular in space and time resolution as we can because we want it to be useful,” Dominici said, adding that it is often city officials charged with making decisions over specific data centers.
The project was backed by a University effort to bolster collaborations between academia and industry. Dominici and Xie’s project was one of seven selected for Harvard’s Ascend award in June, which distributed grants ranging from $250,000 to $750,000 following review by the Office of the Vice Provost for Research and a panel of external industry experts.
