AI energy demand is erratic. Power grids are unprepared.

Artificial intelligence devours energy. But this is not the only problem.

The rapid expansion of AI infrastructure is usually framed as an energy problem. It is predicted that data centers will consume an increasing share of the world's electricity. The International Energy Agency estimates that by the end of the decade they will account for three to four percent of global consumption.

Utilities are already adjusting long-term forecasts as they prepare for growth from hyperscale facilities and high-density computing clusters.

This look captures scale. But he misses the behavior. The problem is not only how much energy huge computing systems consume. The point is how increasingly dense and synchronized computing loads are beginning to change the very operation of the electrical grid. Unpredictable demand, which fluctuates dramatically both in time and space, poses new challenges for network operators.

AI's capricious appetites

Traditional network planning assumes relatively predictable consumer behavior. Industrial, commercial and residential loads generally follow well-established profiles that can be predicted with reasonable accuracy. Even significant increases in demand have historically been offset by reserves, transmission upgrades, and demand management programs.

Large computing infrastructures introduce a new class of electrical loads. Training models - the very computational task that makes AI smart - is usually highly synchronized: clusters of GPUs, TPUs and specialized accelerators work in parallel, tightly and on a schedule. But inference - the process of directly using these models - is distributed and controlled by users, which makes demand much less predictable both in time and place.

Both types of loads, although for different reasons, differ significantly from traditional industrial profiles.

Unlike many conventional manufacturing processes, these workloads can take off instantly - depending on learning cycles, distributed computing coordination, and task scheduling strategies. For the network, this is not just increased demand. This is a sharp demand. High-density computing workloads can cause large spikes in power consumption over ultra-short intervals, including fluctuations in the millisecond range.

Data center operators are already introducing mitigation technologies: batteries, power correction systems, supercapacitors. But collectively, rapid load changes from data centers place additional stress on backup generators, demand control systems, frequency control mechanisms (which keep the grid stable) and local transmission infrastructure.

It's important to understand that the variability associated with computing is different from the variability of renewable energy sources. Wind and solar fluctuations occur on the generation side and are linked to the weather. And computational variability arises on the demand side—due to task synchronization, scheduling, and computational intensity.

The interaction of increasingly dynamic conditions of both supply and demand introduces additional uncertainty into forecasting, reserve management, line load planning and network balancing. Research organizations like the National Renewable Energy Laboratory (NREL) are already highlighting the growing complexity of integrating highly dynamic resources into modern energy systems.

Location, location, location

The problem becomes more acute when computing activity is geographically concentrated. Large data centers are usually clustered in regions with favorable conditions: fiber optic communications, access to markets, tax incentives and historically low electricity prices. Northern Virginia, often referred to as “Data Center Alley,” is the most prominent example. The region has the world's largest concentration of data centers and carries a significant share of the world's Internet traffic.

Utilities operating in these locations have already cited the growth of data centers as the main driver of future load expansion. Virginia-based power provider Dominion Energy, for example, has repeatedly mentioned hyperscale demand growth in its integrated resource planning documents. Virginia has experienced one of the world's largest data center construction booms. Here in Manassas, Amazon Web Services and Iron Mountain data centers define the landscape.

A sudden surge in electricity consumption in a small area can overload substations, power lines and local balancing systems. And this despite the fact that in general there may be sufficient power reserves throughout the energy network. This creates local reliability problems that are not always visible behind overall demand figures.

The situation is aggravated by thermoregulation systems. Cooling in high-density data centers must respond instantly to changes in workloads. As the intensity of computing increases, so does the need for cooling, often nonlinearly. This coupling between computing and thermal systems causes load fluctuations to propagate across multiple levels of a facility's energy consumption.

High-density clusters create other problems at the local level. A large number of accelerators, switching power supplies and high-frequency equipment generate harmonics and non-linear loads. This further strains the distribution infrastructure. Modern data centers, of course, are equipped with interference suppression systems. But the scale and concentration of next-generation equipment may force operators and utilities to rethink approaches to local filtration, harmonic management, and infrastructure resiliency.

By the way, these same factors cause short-term electrical transients. They also hit local networks and power systems.

We need new rules

Another problem is that many of the existing regulatory and operational frameworks were designed to meet relatively stable industrial demand. Previously, they tried to limit large and sharply changing loads. Why? Sudden changes complicate balancing, increase equipment wear, and reduce system predictability. High-density computing clusters do not fit into this framework.

Hence the pressure - we need to change both operational approaches and regulations. Demand management mechanisms may allow some of the computing workload to be shifted or curtailed during peak periods. Data center operators are already experimenting with flexible scheduling, batteries and self-generation. And network operators are reviewing development plans and connection rules for such large and flexible consumers.

The Electric Reliability Council of Texas (ERCOT), for example, has publicly acknowledged that large flexible loads, including data centers, are increasingly impacting long-term planning and grid stability. Lines continue to grow across the US, a direct reflection of the pressure on power generation and transmission.

But here's the rub. Network expansion times are measured in years, not quarters. A structural imbalance arises. The computing infrastructure scales quickly. Electric - no.

The main conclusion: large computing centers are not just another category of industrial consumers. They change the very structure of demand for electricity in time and space. To reduce everything to total consumption is to lose sight of these secondary operating effects. Simply adding capacity does not solve the problems of load spikes, synchronization, local congestion, transients, shrinking reserves, and increasingly stringent load monitoring requirements.

The question is not only how much energy these systems consume. The question is how they change the operating conditions of the network itself. This is not about slowing down the development of AI. It's about recognition: hyperscale computing is a new type of electrical load. And as AI infrastructure scales, roadmaps must consider not only overall energy consumption, but also demand volatility, timing effects, and geographic concentration. Grid resilience will increasingly depend on understanding exactly how these assets consume energy, rather than just how much.