Powering the Growth of AI and Data Centers

For the first two decades of the 21^st century, U.S. electricity demand remained relatively constant as energy efficiency gains enabled everything from lighting to air conditioning to do more with less. Yet now electricity consumption is poised to see incredible growth — around 2% annually for the next decade or more. What’s driving this growth? Population expansion, electrification and electric vehicles account for some of it, but the majority of projected increases is expected to come from the rise of artificial intelligence (AI) and the data centers that facilitate it.

Every question you ask of ChatGPT uses approximately 10 times more energy to answer than an old-fashioned Google search — and now Google uses AI in its results! However, the question AI can’t yet fully answer is, “where will all this additional energy come from?”

In the Carolinas, Duke Energy was already expecting high load growth in its Spring 2023 Supplemental Planning Analysis. By fall 2023, though, over two dozen additional large customer sites (defined as accounting for at least 20 megawatts [MW] of peak demand) were in development — predominantly data centers — representing at least 24,000,000 megawatt-hours of load that the utility will have to provide by 2033. Duke Energy specifically mentioned advanced cloud computing and blockchain operations as the primary types of data centers expecting growth in North Carolina.

Dominion Energy may have an even taller order to fill, as its service territory encompasses “Data Center Alley,” the Loudon County, Virginia, area that is the world’s largest data center hub. According to its 2024 Integrated Resource Plan (IRP), the utility has averaged around 15 data center connections (i.e., data center campuses) annually since 2013. Each of the last two years, these connections have totaled around 1 gigawatt (GW). Although most of the new data center activity is in Virginia, additional load of this magnitude impacts Dominion Energy’s North Carolina customers as well.

Electricity Demand of Data Centers

Data centers present a unique challenge for utilities because they operate 24/7 and can use huge amounts of electricity that must be of the highest reliability and quality. Data centers are trending larger than in the past, with many planned to draw at least 500 MW of electricity demand. Sites owned by technology companies like Microsoft and Google are also likely to require carbon-free energy to meet corporate sustainability goals, such that utilities will need to expand their renewable and nuclear fleets for these customers.

While these data centers have high load factors, meaning their energy consumption is relatively constant around the clock, there is some seasonal variability due to the need to keep equipment cool. In other words, data center loads could potentially be highest when the grid is already strained from air conditioning. However, the power draw of the servers within a data center can fluctuate based on the number of requests and activity required of them. This variability means that while a utility must build transmission infrastructure to power every server in a facility running at maximum capacity, the facility rarely if ever operates at that level.

Data centers require an uninterrupted supply of power, with most large sites requesting at least two interconnection points for redundancy. Data center equipment is so sensitive to power surges and voltage disturbances that any blip in power quality can cause a facility to switch over to backup batteries and generators. The larger the data center, the more likely that a switch to backup power will reverberate across the power system.

In July 2024, a lightning arrestor on a 230-kilovolt transmission line suffered a permanent fault that locked out a northern Virginia transmission line. In the subsequent milliseconds, the resulting power dip caused 60 data centers to automatically switch to backup power. As around 1.5 GW of demand suddenly disconnected, the distribution power spiked, and grid operators scrambled to alleviate the power surge and prevent more systems from tripping offline. Without quick thinking from PJM Interconnection and Dominion Energy, the incident could have triggered blackouts in the area. To mitigate such incidents, utilities often require data centers to support reliability investments that will ensure local power systems remain stable for all customers.

Utility Challenges

Given the sensitivity and sheer quantity of data center equipment slated to connect to the grid, utilities have their work cut out for them to upgrade and harden generation and transmission systems. Of all the challenges presented, the first is modeling these new loads.

Along with the roles that electricity availability and cost play, identifying the placement of a data center involves the developer following up with multiple electric utilities, landowners, state and local governments, and site contractors to determine the location that makes the most economic sense for their specific operation. If each of those individual queries is tallied by all the utilities contacted, then the amount of load growth on the horizon to support AI may be vastly overestimated. It’s difficult to determine the exact number of potential data centers since utilities are essentially competing to land the load in their respective territory and prospector information remains confidential between utilities. Furthermore, not every data center that gets a building permit will actually be built, so there is profound uncertainty around expected data center load growth.

Electric load follows server activity. Daily load curves for data servers can differ depending on whether generative AI training or real-time queries is the dominant load source. Among queries, use case is important, as business activities will coincide with business hours, and personal tasks tend to peak in mornings and evenings before and after work. These variations in intended uses are hard to predict for each server in a data center, especially when they could change over time. At the facility level, the unknowns get magnified.

Assuming future data center loads and locations are predicted accurately enough, the next hurdle for utilities is resource planning. Where can new generation be added to the grid? What type of generation should be used? How much will it cost? What permits are required? What is the lead time on ordering equipment?

With a focus on clean power generation, utilities may look to renewable and nuclear power to meet these large loads, but these technologies have their own challenges. Utilities and grid operators require any proposed generation projects to undergo a series of studies before they can be built. Traditional nuclear plants usually require more than a decade from planning and permitting to construction and operation, with many regulatory requirements along the way. Even new nuclear technologies, such as small modular reactors, have long development times and financial unknowns. Renewable resources such as solar and wind farms may be constructed faster, but yearslong interconnection queues can slow down their operational time frames. Lawrence Berkeley National Laboratory reported that as of 2023, the typical generation project seeking connection to the U.S. grid spent nearly five years in the interconnection queue before it was able to begin commercial operation.

In addition, energy storage must be added to ensure 24/7 availability for intermittent generators like solar facilities. Even natural gas peaker plants, which are traditionally the fastest type of electricity generation to add to the grid, are currently being affected by supply chain backlogs on turbines and generators. Constructing any project that involves a gas turbine may necessitate talking to manufacturers as much as eight years before they’ll be needed to operate.

To add to the complexity even further, simply adding generation sources is to no avail if utilities can’t move that power to their customers. Transmission and distribution upgrades are no small task. High-voltage power lines alone can take three to six years or longer to build. Then the utility must perform a planned outage to safely connect a high-voltage line to the power system, which requires delicate coordination. Once the cross-country transmission is constructed, transformers step down the voltage into usable power. Unfortunately, manufacturer backlogs on transformers are increasing the time and cost of making upgrades, while utilities are highly regulated and ultra-sensitive to any actions that may impact the bottom line of customers. Balancing these competing interests is both art and science.

Data Center Developments

Amid the challenges outlined above, there is hope that data centers will help to meet their own energy needs. Some are even co-locating electricity generation on-site to ensure a dedicated power supply or using power purchase agreements to team up with nearby generation.

In fall 2024, Constellation Energy announced a 20-year partnership with Microsoft to rebrand and restart the Three Mile Island (TMI) nuclear plant in Pennsylvania as the Crane Clean Energy Center. The restart will be of TMI Unit 1, which safely and effectively operated until 2019, when it was retired for economic reasons. It will not include the TMI Unit 2 reactor that infamously incurred a meltdown in 1979 and is still being decommissioned nearly 50 years later. In the partnership, energy not used by Microsoft’s nearby hyperscale data center can be sent to the grid to help meet the demands of other customers.

Traditionally, data centers were not prime candidates for demand response or load curtailment programs, but new research suggests that some may be well suited for load flexibility. Many have backup uninterruptible power supply (UPS) systems that can supplant grid power for multiple hours. Utilities can capitalize on this by offering demand response programs lucrative enough to coax data centers into participation. Other utility programs may capitalize on the operational flexibility of the data equipment within a facility. Inference data centers specialize in training large language models and other forms of AI. This training activity can be scheduled for off-peak periods when the facility can take advantage of cheaper electricity and avoid straining the grid. Furthermore, with the rise of cloud computing and web-based resources, some data processing activities can be moved to other data centers in less-constrained grid locations to free up resources when needed. This coordination of server processing across geographical locations is called “geo-shifting.”

In addition to utility-based initiatives, data center facility owners and computer makers are always working on the latest innovations. Power usage effectiveness (PUE) is a metric used to measure the energy efficiency of a data center. PUE is calculated by dividing the total energy consumption of a data center by the energy used just by the server equipment. Data center operators have made significant strides in efficiency in the past couple of decades, with the newest designs exceeding a PUE of 1.05 (the closer to 1.0, the better). To put this in perspective, a PUE of 1.2 was considered as close to perfect as possible in 2008. A data center load of 50 MW in 2008 with a PUE of 1.2 correlates to an energy savings of over 6 MW with a modern PUE of 1.05, not accounting for efficiency gains from the server equipment itself.

Data center operators are implementing a number of measures to improve PUE. New microchips may handle tasks more efficiently, equating to more output with less energy input. Adding variable frequency drives to cooling fans and pumps can increase cooling efficiency, which optimizes both the energy consumption of computing equipment (which works better at colder temperatures) and the energy use of the entire data center. Even computer coders are looking for efficiency gains. Finding ways for large language models and AI to perform high-value tasks with more efficient and elegant code can mitigate the energy demand needed for those applications.

Conclusion

With highly competitive electricity rates and a diverse energy mix, North Carolina remains an attractive state for data centers and projects in other emerging and re-emerging industries, such as electric vehicles, renewables, steel production and semiconductors. Adding industry opens jobs and attracts top talent, one of the reasons North Carolina is perennially one of the fastest-growing states.

Powering the growth of AI and data centers presents both a formidable challenge and a rare opportunity. As one of Advanced Energy’s stakeholders recently commented, “This is the most exciting and yet frustrating time in my career.” Proactive strategies and coordination among utilities, technology companies, regulators, and other stakeholders are key to ensuring that our power systems remain robust and resilient within a rapidly evolving landscape.