You can’t talk about the utility industry these days without highlighting artificial intelligence (A.I.) and the data center construction boom it’s driving. With some utilities’ projections showing demand growth pushing into multigigawatt territory, concern has focused on the sheer size of new generation capacity that could be required to support supersized A.I. centers.

However, it’s also the nature of the loads creating headaches for local utilities. A.I. data operations are highly variable, meaning that electricity demand can spike and fall in fractions of a second. This usage profile poses stability and power quality hazards for utilities and their customers throughout the centers’ connected grids.

It’s not just capacity

Data centers aren’t a new load category for electric utilities, and we’ve lived through previous booms in this sector before. Development also peaked during the early days of the internet and the rise of online marketing, back in the late 1990s through early 2000s. More recently, the growth in cloud computing has created a market for larger centralized facilities.

While such developments can drive electricity demand higher, their loads remain pretty stable once the kilowatts start flowing. A.I.-related data centers, though, operate differently, with demand profiles that can shift dramatically from second to second. This is especially true for training operations when tasks might be spread across many graphics processing unit servers at a time. These power-hungry devices can ramp up in unison in seconds or even microseconds, driving demand spikes. Operations can then power down just as rapidly, which can create a surge on the connected grid. As a result, these facilities can affect grid reliability across a utility’s network.

The risks detailed

More specifically, A.I. data centers pose challenges to grid operations and other utility customers because of the intense variability of their power use. For utilities, swings of hundreds of megawatts within seconds can disrupt the balance between electricity generation and demand. This can create frequency deviations that lead to transmission instability. Additionally, the high-frequency power electronics used in A.I. hardware can introduce harmonic distortions into the grid. These distortions can lead to voltage instability, which can damage other customers’ household appliances and business equipment, increase the risk of electrical fires and degrade the lifespan of transformers and other utility infrastructure.

Another challenge is the reactivity of A.I. data centers’ uninterruptible power supply (UPS) systems. It makes sense that operators would want to protect their highly sensitive hardware, but there is a danger that hyper-vigilance could potentially lead to cascading blackouts. This nearly happened in July 2024 in Northern Virginia when a minor equipment malfunction caused several milliseconds-long disturbances. 

While these irregularities would otherwise not affect grid performance, they were significant enough for UPS systems at 60 connected data centers to switch their loads off the grid and onto backup generation. The resulting demand drop of 1,500 megawatts forced the grid operator PJM Interconnection and area utility Dominion Energy to quickly reduce the amount of electricity flowing into their lines or face system-wide failures. That worst case was avoided, but the situation underscored just how much damage small voltage shifts could lead to as A.I. loads begin to predominate on local utility systems.

Addressing challenges

Like many other aspects of data center design, this issue of how to address load variability is still under development. On the technology side, synchronous condensers are one option project owners are turning to for help. In November, ABB, Cary, N.C., announced a partnership with microgrid developer VoltaGrid, Houston, for multiple U.S. projects the company has under construction. The condensers are paired with flywheels to provide inertia to ride through short-circuit faults and maintain stable voltages. 

Companies also are looking at ways UPS systems can be improved to support, rather than complicate, grid operations. In June, FlexGen Power Systems, Durham, N.C., which develops battery-­management software, and Rosendin, San Jose, Calif., announced a partnership to integrate technology from both companies into a battery-based UPS design able to smooth out voltage variations to protect on-site and off-site electrical equipment.

Many utilities and state regulators are looking for ways that A.I. data centers’ large load profiles can be used advantageously to support overall grid operations. As a result, grid-interactive UPS approaches could become more common. In these cases, on-site utility-scale batteries could serve multiple functions. Their ability to rapidly absorb or inject power could enable them to ease load transitions and sudden frequency variations. These batteries also could serve as a demand-response resource, supporting all or part of a data center’s load during peak-use periods. Such capabilities could prove essential as these energy-intensive facilities continue to expand.

stock.adobe.com / DIMA RYKOV