The rapid construction of artificial intelligence data centers is testing how quickly the U.S. power system can expand.
Bank of America analysts estimate that the country will need more than 230 gigawatts of new generating capacity over the next five years. Regulated utilities are expected to add approximately 93 gigawatts of accredited supply during that period. Data centers could contribute roughly 125 gigawatts of new electric load by 2030.
Developers are responding by considering on-site gas generation, battery storage and hybrid connections that combine self-generated power with electricity from the grid. Yet securing enough power is only one part of keeping an AI data center running reliably. Data center facility managers must also manage abrupt changes in demand created by high-density computing.
AI clusters can shift quickly between different stages of training and inference, producing power spikes that last only fractions of a second. These fluctuations can contribute to voltage sags, excess heat and wear on electrical equipment.
“Data-center developers are understandably focused on how many megawatts they can secure, but they also need to look closely at how that power behaves inside the facility,” said Richard Qiu, president of LiCAP Technologies. “AI computing can create sudden changes in demand that place stress on power systems over and over again.”
Equipment Shortages Raise the Cost of Power Instability
The equipment needed to deliver electricity to data centers is already difficult to obtain.
Lead times for some high-voltage transformers have stretched to several years, and shortages are affecting switchgear and circuit breakers as well. Some utilities are placing orders three to five years ahead to reserve equipment for planned projects.
Data centers could account for as much as 40% of the U.S. electrical-equipment market by 2030 under faster-growth estimates. That demand could make replacing damaged or overworked equipment even more expensive and time-consuming.
The shortage gives data center owners and facility managers another reason to limit the stress placed on the infrastructure they already have. Brief power spikes may seem small compared with a facility’s total electricity use, but repeated fluctuations can affect voltage stability, produce additional heat and require systems to be built larger than their typical load would otherwise demand.
Ultracapacitors are being considered for this fast-response role. They can charge and discharge within milliseconds and repeat that process many times with limited degradation. In a data center, they can supply power during a brief spike or absorb energy during a sudden change in load.
Batteries are generally used when power must be supplied for a longer period. Generators support extended outages but require time to start and reach the needed output. Ultracapacitors can cover the short interval before those resources respond or help manage fluctuations that do not require a full backup-power event.
“Each energy-storage technology has a range where it makes the most sense,” Qiu said. “Ultracapacitors are suited to frequent, short-duration power events. Handling those events separately can reduce the workload placed on batteries, generators and other electrical equipment.”
This type of configuration could also help data center system designers avoid sizing an entire power system around demand peaks that last for only a few seconds.
Backup Systems Are Taking on a Larger Role
A recent federal emergency order showed how data-center backup systems may be used when the wider grid is strained.
During an extreme heat wave, the Department of Energy directed PJM Interconnection to deploy available backup generation from large electricity users, including data centers. The order was intended to reduce pressure on the Mid-Atlantic grid during a period of record demand.
The decision also raised environmental concerns because many backup generators run on diesel or natural gas. PJM has deployed battery storage more slowly than regions such as California, leaving fewer alternatives when electricity demand spikes.
A layered backup system could give data center power teams more flexibility. Ultracapacitors can respond during the first moments of an interruption, batteries can provide power for a longer interval and generators can support the facility if the outage continues.
“Backup power works best when the response is matched to the duration of the event,” Qiu said. “A power dip lasting less than a second should not necessarily be handled the same way as a prolonged outage. Separating those functions can help facility power teams use their equipment more efficiently.”
Ultracapacitors could also support a controlled shutdown if power cannot be restored, giving active systems enough time to complete critical read-and-write operations and reducing the risk of corrupted data.
States Are Asking Data Centers to Carry More of the Burden
The data-center power crunch is also changing how state governments approach new projects.
New York recently paused state permitting for certain large data centers while regulators study their effects on electricity demand, water use, air quality and nearby communities. The order covers facilities capable of consuming at least 50 megawatts and calls for standards intended to prevent infrastructure costs from being passed along to other utility customers.
State officials are also considering whether developers should fund grid improvements, participate in demand-response programs or support new generation and battery storage dedicated to their facilities.
Similar questions are likely to emerge in other states with large concentrations of planned AI infrastructure. Developers may be asked to provide clearer evidence that their facilities can manage power demand, support local grid reliability and limit the need for costly upgrades.
Generation, transmission and storage will receive much of the investment. Power-quality systems deserve attention as well, particularly as AI equipment creates faster and less predictable load patterns.
The next phase of data-center construction will put greater pressure on data center developers and facility managers to account for what happens between the grid connection and the server rack. Managing those brief spikes and dips may help protect equipment that has become difficult to replace and keep facilities operating when the surrounding power system is under strain.
