Power supply or grid stress? Both need to be solved to meet data center demand
While small modular reactors have been discussed as a potential solution, there are hurdles to making them a reality.
Note to readers: This series of articles focuses on researchers whose work improves efficiency, addresses concerns, or offers alternative solutions to some of the pressing issues created by data centers.
To fully grasp the enormity of the challenge of powering data centers, it helps to understand the scale of what’s actually required to do so.
A small-to-mid-sized city might use 200–300 megawatts. Meanwhile, xAI’s Colossus 2 data center, outside Memphis, has already exceeded that amount, with hopes to eventually build out to a full 2 gigawatts. But even that would only put a dent in the potential 130 gigawatt data center demand by 2030.
“In my world of research, people think the hardest part is finding solutions. But the hardest part is actually finding interesting problems,” said Kirk Cameron, managing director of the Virginia Tech Institute for Advanced Computing, who has worked on data centers for much of his career. “We’re in the initial stage of people really figuring out what the problem is, and the problem is that these things are insatiable in their power needs.”
Data centers existed long before the AI boom. They were first home to online mail servers like Gmail and Hotmail, and later expanded to house the kinds of cloud services that took off in the late aughts and are still in heavy use today. The build-out in Northern Virginia suburbs like Reston and Ashburn, as well as new projects near Roanoke and Richmond, made proximal sense both for the underseas internet cables that connect the U.S. to the rest of the world (some of which connect in Virginia Beach) and the needs of the federal government and its extensions in and around Washington, D.C.
Those data centers all continue to serve those functions, but pale in comparison to the additional capacity — and power — needed for AI-driven data centers today.
That issue is what drove the initial conversation among Virginia Tech researchers. Between its broad fields of study and the academic environment that a land grant institution offers, the Data Center Summit on May 5 is the kind of collaborative event Virginia Tech is uniquely suited to host in 2026.
“With Virginia Tech’s existing expertise, an event like this is timely,” said Ali Mehrizi-Sani, director of the Power and Energy Center and professor in the Bradley Department of Electrical and Computer Engineering. “In this space, we have many capabilities, both in research and developments we’ve done across the university.”
One big topic of discussion right now surrounds the idea of small modular reactors (SMRs), or on-site power plants, that data center companies might be able to build to supply their own power right where it’s needed, without impacting residential and commercial energy delivery and putting extra strain on the grid.
SMRs have two major issues, though, according to Mehrizi-Sani. The first is the mismatch between the generation and the load side. These reactors tend to put out a very steady amount of power, which doesn’t really adapt well to the wild fluctuations in power demand from a data center. You’d need supplemental power storage, possibly in the form of batteries, which are currently expensive, low capacity, and have short lifespans.
The second issue might be more prohibitive: “‘S’ is for ‘small,’ but these are not actually small,” said Mehrizi-Sani, and they aren’t really commercialized yet. Even on the small end of an SMR, we’re talking about a power plant that generates roughly 40 megawatts, enough to power the town of Blacksburg.
These power plants could use a number of different fuel sources, but each has drawbacks. Natural gas plant turbines take at least 18 months to build and burn fuel that adds more carbon to the air. Solar and wind farms need more physical space than many other types of energy at the same power scale. The cleanest, most efficient source that has been discussed is nuclear power, but that presents its own, additional obstacle: it needs people educated in how to run them.
While other researchers at Virginia Tech are working on projects that may accelerate that training in the commonwealth, there’s only so fast that the industry can accelerate.
“Typically for something to move from research idea to an actual capability, you’re looking at a decade,” said Mehrizi-Sani.
Until then, Mehrizi-Sani and everyone else gathering in Alexandria will continue to brainstorm and collaborate on solutions to the most pressing challenges data centers present.
All articles in this series
Power supply or grid stress? Both need to be solved to meet data center demand
Reducing power consumption, protecting data through algorithms
There's a power/water trade-off in data center resource allocation
Thinking small: How small language models could lessen the AI energy burden
