skip to main content
x
Photo montage by Peter Means

The data center dilemma

Behind every AI interaction is a vast network of physical infrastructure. But as demand for data centers surges, so do questions about their impact on surrounding communities.

Few topics these days are as polarizing as data centers.

About two-thirds of Americans opposed new data centers being built in the U.S., according to a recent poll. If the data centers were being built in their own communities, their opposition grew even stronger. In Virginia, which currently has more data centers than any other state, residents have voiced concerns about electricity and water demand, noise pollution, environmental impacts, infrastructure costs, and more.

The data center challenge isn’t going away. Virginia Tech researchers are tackling it head on by engineering solutions to the problems that have made data centers such a lightning rod for controversy.



Lu Zhang

Associate Professor, Building Construction
Professor Lu standing in the middle of large academic tables with pixel art overlaid.
Photo by Peter Means

The challenge: Many Americans oppose local data center construction

“One of the greatest challenges facing data center development is growing community opposition and regulatory uncertainty. But as data center demand accelerates with the global AI race, the question is not whether we should build more data centers, but how we can build them better.

“I think the biggest misunderstanding is that data centers are either entirely good or entirely bad. The reality is much more complex. A project may generate substantial tax revenue and strengthen digital and energy infrastructure, while nearby communities may also experience noise, land-use changes, pressure on electricity and water systems, and rising utility costs, with concerns about how these affect quality of life. Both can be true at the same time. Data centers are not all the same, and many of their impacts can be mitigated through better planning and design.”

The research solution: Engage community members early in the planning and design process 

“To make informed decisions, communities need project-specific information that is provided early, presented in understandable terms, and allows them to compare alternatives. Most importantly, information must be connected to genuine decision-making.

“My research focuses on developing innovative solutions for a new generation of data centers: facilities that are planned, designed, constructed, and operated with greater intelligence, environmental responsibility, and awareness of the people and places around them.

“We are developing a human-in-the-loop, AI-enabled digital platform that can automatically generate, optimize, and compare multiple design alternatives. The platform supports early community engagement so community members can see how their input may change a site plan, infrastructure strategy, mitigation measure, community-benefit agreement, or approval condition. 

“The goal of our digital platform is not to replace human judgment. It is to automate the intensive process of exploring alternatives, calculating impacts, and measuring tradeoffs. Engineers, developers, policymakers, and communities can then focus their judgment on the most promising options and make better-informed decisions together.

“We aim to shift data center development from a reactive process to a proactive one that anticipates impacts, optimizes solutions, and improves projects while there is still time to make meaningful changes.”


Dimitrios Nikolopoulos

John W. Hancock Professor of Engineering, Computer Science
Dimitrios Nikolopoulos leaning up against a table with a wall of windows behind him. There is an overlay of pink numbers in the image.
Photo by Peter Means

The challenge: Data centers are inefficient

“When it comes to energy usage in data centers, adding two numbers is cheap. Fetching them from memory is expensive. The real cost sits in moving data between memory, processors, and networks.

“For example, a language model reads billions of numbers from memory for every word it writes. Then it reads them all again for the next word. Multiplied across billions of requests, that becomes a significant energy challenge. 

“There is another inefficiency: AI systems don't necessarily adjust how much computing power they use based on the difficulty of the task. A relatively simple request sent to a large AI model can still incur that model's full energy cost.”

The research solution: Create AI systems that budget energy use and operate closer to home

“My research group builds systems that treat the quality of an answer as a budget. We spend energy only where it buys quality. That means developing AI systems that can adjust how much computational effort they use based on what the task requires, meeting that need with the fewest joules possible. 

“We’re also addressing the energy costs of moving data. We're exploring ways to move AI inference out of large data centers and closer to where the data originates. Instead of automatically sending information hundreds of miles to a data center, we can ask whether some of that computation could happen closer to the source, in smaller computing nodes.

“If this works at scale, the data center of the future becomes a continuum. Warehouse-sized data centers could handle the hardest computational work, while many smaller nodes handle other tasks closer to people, instruments, and sensors.

“Rather than using the same amount of computing power for every task, energy use could better reflect the complexity and value of the work being done. Asking where computation belongs — and how much computation a task actually requires — could help reduce the resources needed to support growing demand for AI.”


Eric Sjostedt

Presidential Postdoctoral Fellow, Civil and Environmental Engineering
Eric Sjostedt leans up against his work desk. There is an overlay of numbers and lines in the stylized photo illustration.
Photo by Peter Means

The challenge: Data centers draw large amounts of water from local systems 

“Nationally, data centers may represent a relatively small share of water demand. But a data center using 1 million gallons of water per day is drawing roughly the same amount as 2,500 households. Put several large facilities on the same local water system, and the scale changes quickly — and water is fundamentally a local resource.

“The engineering choices made within a data center can have impacts far beyond the building itself. Direct water use at data centers is predominantly associated with cooling, but when considering their total water footprint, the majority of water use is associated with generating the electricity that cooling requires. And there isn't one cooling technology or one solution that makes sense everywhere. Location matters.

“Evaporative cooling technologies are extremely effective at dissipating heat, but they require significant amounts of water. Dry cooling alternatives can use almost no water at the facility, but they generally require more energy because air is less efficient than water at transferring and dissipating heat. The right choice depends in part on where the data center is located, what resources are available there, what the local regulations are, and what other demands already exist on the local infrastructure.

“Right now, we're trying to address an information gap. When a data center proposal comes to a community, how can a water or wastewater utility better evaluate it and make sure it can continue providing safe, reliable, and affordable service to its existing customers?”

The research solution: Create an information exchange that can help communities plan for data center water usage

“There isn't a common body of knowledge utility companies can turn to for guidance, particularly for a rural utility that may never have dealt with a data center proposal before.

“We're working to create that common body of knowledge, as well as a more structured, repeatable way for developers, operators, and utilities to exchange information. Utilities need to understand what technologies a facility will use, how it will operate, how quickly it will come online, and what infrastructure will be required to support it. Is the existing infrastructure capacity sufficient? Is the facility close to the water, power, telecommunications, and other infrastructure it will need? What upgrades will be necessary, and who will pay for them?

“If we can improve the exchange of information among developers, utilities, communities, and other infrastructure providers, data centers could be planned alongside the systems and communities that will support them — rather than being proposed in a location and leaving utilities scrambling to catch up. They need to be in the room when these decisions are being made.

“If we do this right, it becomes about more than building a data center. It could also help strengthen the infrastructure systems that communities depend on for the future.”


Jonathan Bluey

Assistant Professor of Practice, Building Construction
Jonathan Bluey sits at his work desk. There is an overlay of a dot pattern in the stylized photo illustration.
Photo by Peter Means

The challenge: There aren’t enough skilled workers to build sustainable data centers 

“Data centers aren't new, but they've grown tremendously in size and complexity. As demand for AI and cloud computing increases, so does the need for the infrastructure and skilled workforce required to support them.

“The U.S. construction industry will need an estimated 349,000 additional workers in 2026 and 456,000 in 2027 just to keep labor supply and demand in balance. Meanwhile, data center construction continues to accelerate.

“Meeting that demand isn't just about building more facilities. It's about ensuring that the people designing and constructing them understand the technologies, resources, safety requirements, and sustainability considerations involved.”

The research solution: Design a course to train workers to build data centers safely and more sustainably

“At the Myers-Lawson School of Construction, we are creating a 15-week AI Data Center Construction course to better prepare future leaders to work in the data infrastructure industry.

“Understanding how a data center construction project is built is a critically important part of the solution for problems like sustainability. The question is always, "What's the solution?" followed by finding practical answers to the issues that exist today.

“Energy is one of the most important considerations. Data centers require a lot of energy. With sustainability in mind, I keep returning to the same principles: people and infrastructure are governed by physics and economic factors. Engineering energy, for me, is focused on maximizing purpose and minimizing entropy.

“I am collaborating with industry professionals and subject matter experts to create this systematic course that covers an AI data center project from start to finish. Students will be taught everything from contracts, estimating, scheduling, and sitework to the structural, mechanical, electrical, and plumbing aspects of a data center project. We’ll showcase the highest standards of sustainability, safety, technology, and quality.

“The course is for the future leaders, designers, and builders of these facilities — or anyone who wants to better understand how this infrastructure is built today.”


Ranga Pitchumani

George R. Goodson Professor, Mechanical Engineering
Ranga Pitchumani leans on an office cube divider. There is an overlay of zeros in ones in this photo illustration.
Photo by Craig Newcomb

The challenge: Data centers aren’t consuming power steadily

“We've lived with data centers for quite some time. They support our financial services, the health care systems we rely on, and our everyday email communications.

“But AI data centers can have very different power requirements. The easiest analogy I give people is to think about an EKG. Traditional data centers have less variability in power consumption, like a normal EKG. But with AI data centers, especially those training large models or running certain AI workloads, power demand can hum along like a normal EKG, then suddenly spike dramatically, placing additional demands on the power infrastructure.

“I think one of the biggest misunderstandings is the perception that a data center has relatively uniform or constant power consumption. That's really not the case. Different workloads, resource utilization, and cooling requirements are all things that need to be predicted as accurately as possible.”

The research solution: Forecast power usage

“We're developing tools that can predict power demand over different periods of time, including what's called probabilistic forecasting. Rather than just putting a precise number on what the power demand will be, we can estimate the probability of different outcomes. That helps in planning resources, making operational decisions, reducing costs, and improving efficiency.

“If I can predict what the computing load is going to be, can I proactively adjust cooling rather than waiting for things to heat up? When should I start cooling? How much cooling is needed? These are important questions because better forecasting benefits not just the data center, but also the utilities that deliver power to it.

“We're not just looking at the total power consumption of a data center. We're asking how much power each computing node and server requires. If a particular server is going to be used a lot, can we target the cooling requirements to that server as opposed to the entire building? That could make a big difference in efficiency.

“We're working with Virginia Tech's Advanced Research Computing facility, which is like a mini version of an actual data center. They provide us with historical and real-time data about computing jobs and the systems running them, including GPUs, CPUs, and memory. We're using that information to build models that forecast power requirements for individual computing nodes, as well as cooling needs for the data center. We're also testing those predictions against actual operating data.

“Eventually, we'll work with Virginia Tech Electric Service to understand how power delivery can be tailored to meet those requirements reliably and cost-effectively.

“Solar energy can also play an important role, but not necessarily in the way we think of solar panels powering a house. It's part of a larger ecosystem that includes wind, energy storage, and other power sources. We're building on our previous work in renewable energy forecasting to better predict how much renewable energy will be available alongside how much electricity a data center will need. Together, those forecasts can help us make better decisions about energy generation, power delivery, and data center operations.

“By forecasting what's coming, we can proactively take steps rather than react to problems after they arise. How can we use these tools to support thoughtful data center growth that benefits society? That's the conversation that needs to happen among the public, utilities, data center developers, and policymakers.”

If you want to support cutting-edge faculty research like the projects featured in this magazine, go here to support the College of Engineering. For more information, call (800) 533-1144.