For years, the discussion around AI infrastructure has focused on chips, compute capacity and the race to build larger data centers. This summer has made power constraint behind that expansion harder to ignore
Extreme heat has pushed cooling demand higher across the United States just as data centers are adding large, concentrated loads to the grid. The Department of Energy now describes the country as moving from decades of relatively stagnant electricity demand into a period of significant load growth, with hyperscale AI data centers among the major contributors.
The combination should change how we think about the infrastructure behind AI. The challenge is not simply supplying enough electricity over the course of a day. Utilities and data-center energy teams also have to deliver the right amount of power at the right moment, maintain voltage and frequency, and recover quickly when supply or demand changes.
AI is arriving during a much tougher grid cycle
New NOAA data shows just how intense that pressure is becoming. July 2026 was the warmest month on record for the contiguous U.S., averaging 76.9°F, with every state in the Lower 48 running at least 1°F above its 20th-century average. Thirteen states recorded one of their five warmest Julys, pushing cooling demand higher at a time when utilities are also preparing for a new generation of large, power-intensive loads.
At the same time, projections for data-center electricity use continue to climb. Lawrence Berkeley National Laboratory estimates that data centers could account for 11.8% of total U.S. electricity consumption by 2030 under its reference case, with scenarios ranging from 9.5% to 15.3%.
Those numbers are significant on their own, but annual consumption doesn’t tell the entire story. Data centers are concentrated geographically, require highly reliable power and can create major new demand in places where transmission and generation infrastructure were planned years before AI workloads reached their current scale.
The Department of Energy has noted that data center loads can put pressure on regional grids because of their size, location constraints and need for firm power. Meanwhile, federal regulators are already reconsidering how very large loads connect to the transmission system as utilities try to accommodate them without pushing unnecessary costs or reliability risks onto other customers.
More generation will help, but the grid also needs speed
It’s tempting to frame the answer only in terms of building more generation. More electricity supply is clearly necessary, and renewables are likely to provide a substantial share of it. But adding megawatt-hours alone will not solve every power quality and reliability challenge created by large, rapidly changing loads.
The International Energy Agency expects renewables to meet nearly half of the additional global electricity demand from data centers through 2035. It also expects natural gas, nuclear and other sources to play important roles, which reflects how difficult it would be to keep data centers running continuously with any single generation technology.
Renewables bring their own characteristics. Solar production changes with cloud cover and disappears at night, while wind output varies with weather conditions. Those resources can supply enormous amounts of energy, but higher penetration increases the value of technologies that can respond when generation and demand move out of balance.
That distinction between energy and power deserves more attention in the AI infrastructure conversation. A grid may have sufficient energy available over hours or days and still experience short duration mismatches that create voltage deviations, frequency instability or stress on equipment. The infrastructure therefore has to be designed for both capacity and response speed.
The IEA has identified greater power-system flexibility as an important requirement as grids absorb more variable renewable generation alongside concentrated loads such as data centers. Transmission expansion, batteries, flexible demand and other resources will all have roles in providing that flexibility.
Storage technologies should be matched to the job
Energy storage is sometimes discussed as though it is a single category with one set of capabilities. In practice technologies should be matched to the duration, power level and cycling profile of the job. Lithium-ion batteries are very good at storing substantial amounts of energy and discharging it over longer periods, making them valuable for applications ranging from renewable energy shifting to backup power and grid services.
Short-duration, high-power events, on the other hand, create a somewhat different engineering problem. A sudden load change may last seconds rather than hours, but the equipment responding to it has to react almost immediately and may have to repeat that cycle thousands or even millions of times over its operating life.
This is where ultracapacitors can complement batteries. They can charge and discharge very quickly and tolerate a high number of cycles, allowing them to absorb or provide power during brief, high-power events while a battery or another resource handles the longer-duration energy requirement.
Hybrid systems therefore deserve more consideration as AI loads grow. Rather than asking one storage technology to do every job, designers can assign each technology the duty cycle it handles best: batteries for sustained energy needs and ultracapacitors for transient loads, rapid power fluctuations and other high-power events.
There is also a lifecycle argument for separating those duties. Repeated high-current events generate heat and place additional stress on battery cells, so diverting some short-duration power events to another storage technology can reduce the number of demanding cycles imposed on the battery. It doesn’t remove the need for batteries, but it can allow them to function closer to the conditions they were designed to handle.
Data centers will have to become better grid participants
The relationship between data centers and the grid is also likely to become more interactive. Historically, electricity infrastructure was largely built around generators responding to consumer demand. AI facilities are large enough that their behavior matters at the system level. AI hyperscalers’ facilities may have on-site generation, batteries, uninterruptible power supplies and sophisticated power-management systems, creating opportunities to manage when and how it draws electricity from the grid.
Examples of this could include limiting certain loads during constrained periods, using on-site storage to smooth short spikes or coordinating backup resources with utilities. Lawrence Berkeley National Laboratory has already identified load flexibility, better interconnection processes and improved planning as part of a wider set of options for connecting large electricity users more quickly.
The infrastructure race extends well beyond the data center
There is a broader industrial issue here as well. The United States is investing heavily in AI computing capacity, but servers and buildings are only part of the physical supply chain required to support that expansion. Transmission equipment, power electronics, batteries, ultracapacitors, transformers and other grid technologies will also have to scale. Many of those products require specialized manufacturing capacity that cannot be expanded overnight.
Domestic manufacturing therefore belongs in the infrastructure discussion. The same urgency being applied to semiconductor production and data-center construction should extend to the equipment required to move, condition and store electricity around those facilities. Compute capacity cannot scale faster than the power infrastructure that supports it.
There’s little value in completing a data center months earlier if the local grid can’t connect it, or if key electrical components remain stuck in long procurement queues. The Department of Energy’s latest transmission assessment is already pointing to grid capacity as a constraint as data centers, advanced manufacturing and other large loads seek new connections.
This summer is a useful warning
The grid is being asked to manage rising electricity demand, extreme weather, new industrial loads, changing generation resources and aging infrastructure simultaneously. AI makes those pressures more visible because its power requirements are growing unusually quickly. The IEA estimates that data centers will account for roughly half of U.S. electricity-demand growth through 2030, which means decisions being made about power infrastructure now will affect how easily new computing capacity can be added later in the decade.
This summer offers a useful preview. Adding generation is essential, but the AI power challenge is not only about producing more electricity. It is also about moving, storing and delivering power at the speed each application requires. Transmission, storage, grid flexibility and fast-response technologies will all be needed if AI infrastructure is to scale without making reliability the next bottleneck.
