How AI Data Centers Are Changing Power Demand on the Grid
A practical look at what the growth of AI infrastructure means for utilities, grid operators, and the investors who support them.
The growth of AI and hyperscale data centers is creating a sustained increase in electricity demand that utilities across the United States are working to understand and plan for. Even under conservative assumptions, cutting current industry forecasts in half, load growth over the next twenty years is still projected to run approximately three times faster than it did over the prior twenty, and at full forecast projections that figure is higher still. At either end of the range, this represents a scale and pace of demand growth that falls outside what most utilities have historically had to plan for, arriving on a timeline that is significantly shorter than traditional infrastructure investment cycles. This article assesses what is driving that demand, how it interacts with the grid at different levels, where the constraints sit today, and what tools utilities and grid operators have available to respond effectively.
3x Projected load growth over the next 20 years vs. the prior 20, even at half of current industry forecasts
How AI Data Centers Differ from Traditional Data Center Load
A traditional enterprise data center draws between 5 and 20 megawatts. A pre-AI hyperscale facility might reach 150 megawatts. A large AI training campus in 2026 draws between 500 megawatts and 1 gigawatt of continuous power. The next generation of planned facilities are designed at 2 to 5 gigawatts. That scale difference matters because it changes the category of grid infrastructure required to serve these facilities, not just the size of the connection.
The power draw characteristics are also meaningfully different. Traditional data centers draw steady, predictable power that utilities can plan around. AI training workloads produce periodic load oscillations because tens of thousands of GPUs execute identical computational steps in lockstep, alternating between a high-draw compute phase and a lower-draw synchronization phase every 3 to 10 seconds. Research published in August 2025 by NVIDIA, Microsoft, and OpenAI documented system-level power ramp rates exceeding 1,000 megawatts per second at gigawatt scale during these transitions.
AI training workloads produce periodic load oscillations at frequencies that can interact with the natural resonant modes of large transmission systems, a characteristic that conventional grid planning tools were not designed to account for.
Inference workloads, which now account for 80 to 90 percent of all AI compute consumption, present a different planning challenge. Inference load is driven by the random arrival of user queries and spikes in response to viral events or breaking news. It cannot be deferred without affecting the end user. This means the grid needs to maintain elevated reserve margins to absorb variability that cannot be forecast in the same way traditional industrial loads can.
Where the Constraints Sit Today
The grid constraints created by AI data center growth are not concentrated in one place. They appear at every layer of the system, for different reasons, on different timescales.
At the generation level, PJM's 2027 to 2028 capacity auction cleared 6,600 megawatts below its reliability target. Data centers account for 94 percent of all projected load growth in PJM through 2030. At the transmission level, PJM congestion costs escalated from 1.8 billion dollars in 2024 to 3.2 billion dollars in 2025. At the substation level, large power transformer procurement lead times have extended from approximately 140 weeks in 2023 to over 160 weeks by 2026, with a persistent 30 percent supply shortfall projected through 2030.
The geographic concentration of data center demand compounds these system-wide pressures. In Virginia, data centers already account for over 20 percent of total state electricity generation, a figure projected to reach >50% by 2030. In the PJM interconnection region, which serves 67 million people across 13 states, data centers are responsible for 94 percent of all projected load growth through 2030. PJM capacity prices rose 833 percent between the 2024 to 2025 and 2025 to 2026 delivery years, a cost ultimately reflected in the bills of residential and commercial customers across the region.
The Timeline Mismatch
7–10 years - Typical timeline to build the transmission infrastructure a new data center campus needs. While the campus itself takes 18 to 24 months.
The timeline mismatch is one of the most structurally difficult aspects of this challenge. An AI data center is constructed in 18 to 24 months. The substation serving it requires 3 to 5 years. The transmission line it needs takes 7 to 10 years. Every element of the supply chain moves more slowly than the demand it is supposed to serve.
How Data Center Load Affects the Grid at Different Scales
The impact of a data center on the surrounding grid changes significantly depending on the size of the facility and how it connects. There are five distinct scales at which different categories of grid stress are triggered, each affecting a different set of stakeholders.
| Scale | Grid Impact |
|---|---|
| 1 – 10 MW | Sustained load on distribution feeders and transformers. Neighboring customers may see degraded power quality. |
| 20 – 25 MW | Formal interconnection study required. Queue begins. Timeline pressure builds for developers. |
| 50 – 100 MW | Substation architecture changes required. Upgrade costs spread to all ratepayers in the territory. |
| 75 – 300 MW | NERC bulk power system reliability framework applies. Simultaneous disconnection becomes a documented hazard. |
| >20% of regional zone | Structural regional effects. Capacity markets reprice. All customers across the multi-state region absorb the cost. |
The first four thresholds relate to the behavior of an individual facility. The fifth relates to aggregate geographic concentration and requires grid-scale infrastructure and regulatory responses rather than facility-level solutions.
The Simultaneous Disconnection Challenge
One of the less well-understood aspects of AI data center grid interaction is how their protective systems respond during transmission disturbances. AI data center server equipment operates within tighter voltage tolerance windows than most large industrial consumers. UPS systems transfer to backup power on a voltage deviation of 5 to 10 percent sustained for more than 40 to 66 milliseconds. By comparison, the ITIC curve, the standard against which utility equipment is designed to ride through disturbances, considers an 80 percent voltage event lasting 50 milliseconds acceptable under normal utility operating conditions. The data center's protective system and the utility's reliability standard are calibrated to different thresholds.
Because many facilities across a region share similar protective settings, a single transmission disturbance can trigger simultaneous disconnection across a large number of data centers at once. On July 10, 2024, a single failed lightning arrestor on a 230 kilovolt transmission line in Virginia triggered a voltage depression lasting less than 70 milliseconds. That event caused 1,500 megawatts of AI data center load from 60 facilities across 25 substations to disconnect from the grid in under one second.
The reconnection phase creates a secondary concern that compounds the original disturbance. When UPS battery banks across many facilities begin recharging simultaneously after a fault clears, they generate a significant demand surge at precisely the moment the grid is still recovering from the event that caused the dropout. The July 2024 Virginia event, in which 1,500 megawatts disappeared from the grid in under one second, was not an isolated incident. As AI data center density increases across constrained corridors, the conditions that produced it are becoming more common, not less. The parallels to Spain's April 2025 nationwide blackout are instructive: a rapid, large-scale imbalance between generation and load that outpaced the grid's ability to maintain voltage and frequency stability, with consequences felt by every power user across the affected region. NERC formally noted that the Virginia pattern had not been anticipated by power system operators, and in May 2026 issued a Level 3 Essential Actions Alert requiring transmission owners to install dynamic fault recording devices for computational loads, with responses due August 3, 2026. The regulatory response reflects the seriousness of what an unchecked version of this problem could mean for grid reliability at a regional scale.
Regulatory Response and the Push to Maximize Existing Infrastructure
The conventional response to demand growth is to build new generation, transmission, substations, and distribution infrastructure. That investment is necessary and is happening across the industry. At the same time, the timeline mismatch between data center construction and grid infrastructure buildout means that new physical assets alone cannot close the gap within the timeframes the market is demanding.
A complementary approach is gaining significant regulatory and industry momentum: extracting more usable capacity from infrastructure already built. FERC and the broader industry refer to this as Grid Enhancing Technologies, or GETs, a category of hardware and software solutions designed to maximize the capacity of existing transmission lines rather than expanding them. In June 2026, FERC issued coordinated show cause orders under Docket No. RM26-4, requiring all six US regional grid operators to formally evaluate GETs as part of their large load integration plans.
One of the most practical GETs available today is Dynamic Line Rating, or DLR. Most transmission lines are rated using static engineering assumptions set without real-time knowledge of actual conditions such as ambient temperature, wind speed, and current loading. DLR replaces those fixed assumptions with live sensor data, allowing utilities to calculate the actual thermal capacity of a line in real time. Lines that appear fully utilized under static ratings often have meaningful additional safe capacity when measured against actual conditions. Analysis of a May 2026 PJM congestion event showed that real-time monitoring could have provided 12.3 percent additional capacity on the constrained line, potentially reducing approximately 100 million dollars in congestion costs borne by customers over a 72-hour period.
20–30% Estimated unused headroom on many existing transmission lines under static ratings, which real-time sensing can help utilities safely access.
The Grid Monitoring Gap
The monitoring infrastructure most utilities rely on today was designed for a different load environment. SCADA systems collect telemetry at the substation level on minute-scale timescales. They are effective for the loads and conditions they were built to serve. For AI data center loads, which produce sub-second power transitions, tight voltage tolerance trip points, and simultaneous disconnection risk across dozens of facilities, substation-level monitoring does not provide the resolution or the coverage needed to manage the grid effectively.
Three specific monitoring gaps create the most significant operational vulnerabilities. First, when a fault occurs on a distribution line part of servicing a data center campus, utility crews must physically patrol the line to locate it. US average CAIDI rose to a record 7.4 hours in 2024, with crews spending approximately two-thirds of that time searching for the fault rather than repairing it. For an AI campus running a frontier model training job, that outage duration has direct financial consequences.
Second, grid operators currently have no real-time visibility into the electrical behavior of large loads. They cannot see when a training job is running, when a checkpoint event is about to produce a significant power swing, or when multiple facilities with identical UPS settings are approaching a simultaneous trip condition. Without that information, pre-emptive action is not possible.
Third, conventional monitoring has no ability to detect the precursor conditions that indicate an impending fault, such as rising conductor temperature, progressive insulation degradation, or vegetation growth approaching clearance limits. Faults arrive without warning.
How EGM Addresses the Monitoring Gap
EGM's Meta-Alert platform is built to provide the transmission and distribution level visibility that conventional SCADA does not. Its patented Accurate Fault Location and Detection system, independently validated by the US Department of Energy's National Laboratory of the Rockies across 26 blind test scenarios, locates faults to within a single pole span of approximately 300 feet. That capability compresses fault response from 3 to 5 hours of line patrol to approximately one hour from alarm to repair.
The platform monitors over 60 electrical, physical, and environmental parameters in real time at each sensor cluster location, including voltage, current, fault current signatures, conductor temperature, harmonic distortion, wind speed, and humidity. The Dynamic Line Rating module uses live conductor and environmental data to continuously calculate actual available capacity on monitored lines, with deployments demonstrating up to 50 percent improvement in usable capacity compared to static ratings.
For AI data center corridors where interconnection queue timelines are measured in years, dynamic line rating can reveal additional firm capacity on existing feeders that is available now, without new construction.
For the simultaneous disconnection problem specifically, EGM sensors placed at the utility service entrance, the boundary where UPS systems make their disconnect decision, capture voltage events before the trip occurs. That provides the earliest possible warning of an impending mass disconnection cascade, directly addressing the visibility gap NERC identified in its review of the Virginia 2024 event.
The system installs on energized lines without service interruption, requires no field calibration, and is fully operational within hours of installation. A typical 500 megawatt campus deployment uses 8 to 14 sensor clusters deployed across the full power delivery chain in 60 to 90 days. The platform directly satisfies the dynamic fault recording requirements of the NERC Level 3 Alert, making it both an operational tool and a compliance solution for utilities serving large computational loads. Validated return on investment across deployed systems is under one year.
The growth of AI data center demand is creating a new set of planning and operational requirements for the grid. Understanding the load characteristics, the scale thresholds, and the monitoring gaps is the starting point for developing responses that are both practical and effective for utilities and the customers they serve. To learn more about EGM's Meta-Alert platform or request a Grid Vulnerability Assessment, visit egm.net
