The growth of artificial intelligence could spur the construction of new datacenters such as this one in Northern Virginia. Source: Getty Images.

About 30 miles west of Washington, DC, Virginia's Loudoun County is the largest datacenter market in the world, and the electric utility that serves it, Dominion Energy Inc., has long pointed to the industry as a "growth machine" driving increased demand for power and increased investment to meet that growing demand.

Accommodating that growth could prove to be a challenge as artificial intelligence is expected to require significantly more electricity as the technology scales, forcing datacenter operators, utilities and government officials to prepare for those needs even as the precise impact remains unknown.

Electricity demand from datacenters in Virginia increased by about 500% from 2013 to 2022, according to a June presentation by Dominion to mid-Atlantic grid operator PJM Interconnection LLC, with 81 datacenters totaling 3.5 GW in capacity connecting to the utility's system since 2019.

PJM now requests Dominion and other load-serving entities in the Dominion transmission zone to provide a 15-year datacenter forecast instead of a five-year outlook. The migration to cloud computing, 5G technology and AI are seen as some of the biggest drivers of growth in the PJM region.

While about 80% of the datacenter industry in Virginia is located in about 30 square miles of Loudoun County, there are "established and growing markets" outside of Northern Virginia, Alan Bradshaw, Dominion's vice president of strategic partnerships, said in an interview. Other datacenter growth areas include around Richmond, Va., and in the southwestern portion of the state.

In its 2023 integrated resource plan, Dominion, whose utility operates legally as Virginia Electric and Power Co. and is known as Dominion Energy Virginia, projected datacenter peak demand to reach 13.3 GW by 2038, which Bradshaw explained refers only to those datacenter customers within the utility's service territory.

"We're starting to see datacenters move into some of the electric cooperative service territories ... and when they're in those service territories, the co-op is our customer. So, we have to bring the transmission infrastructure to their service territory," Bradshaw said. "So, it's a much bigger pie than just the 13.3 [GW] in our service territory."

The datacenter industry "has grown on average 0.5 GW a year in the last three years" in Dominion's service territory, the company said, "achieving a peak metered load of almost 2.8 GW in 2022."

In its integrated resource plan filing, Dominion cited PJM's 2023 load forecast showing summer peak demand in the Dominion transmission zone increasing at a compound annual growth rate of about 4.4% and net energy increasing 6% between 2023 and 2038. In the prior year's forecast, PJM showed summer peak demand increasing about 2.0% and net energy increasing 2.9% between 2022 and 2037, Dominion noted.

Over the next 10 years, PJM's forecast shows annual peak growth of nearly 5% and energy load increasing 7%, Dominion said, an increase "driven primarily by data centers and, to a lesser extent, electrification in both the company's service territory and in other service areas within DOM Zone."

AI implications

Two key branches of AI are machine learning, also referred to as training, and inference.

"What we're hearing is that ... the server racks that are required for that training mode will pull ... somewhere between five and seven times more [energy] than the server racks that we're seeing today," Bradshaw said, adding that part of that is tied to the transition from CPUs to graphics processing units.

"Depending on what you're building, it can take hours, days, weeks [or] months to go through that process," Bradshaw said.

Inference is the process of making decisions or a calculation based on available data.

"What we're hearing about inference is that's going to require about two to three times more energy than the typical server with today's technology," Bradshaw said. "So clearly, much more demand, electrical demand, for AI than what we've seen to date."

While Dominion is "trying to get our arms around" AI, the company's load forecast assumes some level of growth and "expects that continual evolution of technology to continue to occur," Bradshaw said.

"Our forecast does not fully integrate the impact of five to seven times more energy in a server rack, at least not now," Bradshaw said. "It will as we learn more."

Banking on efficiency

Improvements to the operating efficiency of datacenters could help temper demand growth.

"The datacenters, typically, most of them replace servers about every three to five years as technology gets better," Bradshaw said. "Typically, the servers get more efficient. Now sometimes that means it reduces load. Sometimes it means they can put more in the same space, right? So it kind of offsets a little bit."

Chris Walker, director of sustainability at Amazon Web Services Inc. (AWS), a division of Amazon.com Inc., described AI as "the transformational technology of our time."

"It's not new to us though. Amazon has invested heavily in the development and deployment of AI and machine learning for more than 25 years, both in the customer-facing services but also within our internal operations. It is embedded in our DNA," Walker said in an interview, adding that AWS is helping customers "achieve sustainability objectives through the use of AI and machine learning."

AWS noted that organizations are increasingly looking to migrate from on-premises datacenters to cloud computing because of "increased agility, cost savings and increasingly because AWS is much more sustainable than on-premises infrastructure."

In addition, AWS noted that 451 Research has found AWS infrastructure to be 3.6 times more energy efficient than the median of surveyed enterprise datacenters in the US.

"AWS has early access to the latest server technology, and we're looking to constantly adapt our designs and our operations inside of our datacenters around more energy-efficient server platforms, which are faster than enterprise [datacenters]," Walker said.

AWS also uses "advanced modeling methods within our engineering space" to optimize datacenter design, according to Walker, which allows for greater reliability and better energy efficiency.

A "high-performance machine learning chip" reduces time and costs associated with training generative AI models, Walker said, "cutting training time for some models from months down to hours."

"This, in turn, means building new models requires less money, less power and has a potential cost savings of up to 62% and energy consumption reductions of up to 29% versus comparable instances," Walker added.

Government action

Virginia Gov. Glenn Youngkin signed an executive directive Sept. 20, the provisions of which include directing the state Energy Department to study the impact of AI on power generation, particularly the "expected increase in energy demands ... necessitated by increased adoption of AI."

"Obviously, there is going to be a tremendous impact," Virginia Department of Energy Director Glenn Davis said in an interview.

Youngkin's administration is working with utilities and datacenter companies to properly model the growth and is assessing the potential for small modular nuclear reactors or even hydrogen to serve as behind-the-meter resources to power datacenters, Davis said.

Dan Thompson, a principal research analyst at S&P Global Market Intelligence specializing in datacenters, listed Alphabet Inc. subsidiary Google LLC, Amazon and Microsoft Corp. among the major tech names committed to investing in infrastructure to support growth.

"Access to power is kind of the key challenge," Thompson said, acknowledging that determining how much current datacenter growth can be attributed to AI remains hard.

"While we cannot attribute 100% of the current construction boom in certain markets of the datacenter industry to AI, it is reasonable to assume that it is being helped by the growth of AI," Thompson wrote in a Sept. 11 report.

Walker, at AWS, could not comment on how much electricity demand is tied to AI.

Growing too fast?

The rapid acceleration of datacenter growth during the height of the COVID-19 pandemic prompted Dominion to reevaluate its long-term transmission planning, temporarily suspending new datacenter connections in late summer 2022. Dominion resumed datacenter connections after a two-month pause, during which it studied 10 to 15 years of planned transmission projects to determine which could be built earlier. Some are now complete, and others are expected to come online in 2025 or 2026.

A PwC report commissioned by advocacy group the Data Center Coalition cited the suspension as an example of a utility ensuring transmission infrastructure keeps pace with growing demand.

"In response, some datacenter owners and operators have partnered with the utility to identify unique solutions for new data center projects," the report said, adding this has included a datacenter owner partnering with the utility to build a 300-MW substation on-site.

Dominion emphasized that all of the datacenters that have asked for transmission service have received it. Bradshaw said Dominion is open to exploring several solutions to power datacenters as part of resource planning, which is modeled along the lines of Youngkin's all-of-the-above energy plan.

"You're probably not going to have a solar field right next to a datacenter ... because they have a very high load factor and they run 24/7," Bradshaw said. "It would take quite a bit of solar to do that."

S&P Global Commodity Insights reporter Jared Anderson produces content for distribution on Platts Connect. S&P Global Commodity Insights is a division of S&P Global Inc.

451 Research is part of S&P Global Market Intelligence. S&P Global Market Intelligence is a division of S&P Global Inc.

S&P Global Commodity Insights produces content for distribution on S&P Capital IQ Pro.