AI training loads are reshaping the grid at frequencies SCADA cannot resolve. The utility must serve the load that shows up, absorb the consequences, and explain costs it cannot yet measure.
AI data centers do not draw power the way factories, office buildings, or even traditional data centers do. Large-scale GPU training synchronizes tens of thousands of processors, alternating between power-intensive computation and lower-power communication phases. The result is rhythmic power oscillations pulsing in the sub-synchronous frequency regime.
These oscillations can excite torsional resonance in turbine-generator shafts at power plants connected to the same transmission network. The result is physical destruction of generation equipment at a power plant miles from the data center that caused it.
This is not a hypothetical. Northern Virginia hosts roughly 70% of America's data center capacity, all on Dominion Energy's transmission system. A Bloomberg analysis of approximately 770,000 residential power sensors found that Loudoun County experiences harmonic distortion at four times the national average.[1]
More than three-quarters of the worst power-quality readings nationwide are within 50 miles of major data center clusters, affecting an estimated 3.7 million Americans.[1]
A hyperscaler can relocate a project to a different state, a different grid operator, a different country. The utility has no such option. It must serve the load that shows up, absorb the grid impacts, justify the infrastructure costs to regulators, and answer to ratepayers when bills go up or power quality goes down.
The largest technology companies in the world are making siting decisions based on available power, tax incentives, and fiber connectivity. They are not selecting for grid health.
Data centers choose where to build. Utilities do not choose their customers.
After two decades of nearly flat U.S. electricity demand, AI data centers have driven a surge in load growth at a pace the grid has not seen in generations.[2]
In PJM, the regional transmission organization serving 65 million people across 13 mid-Atlantic and Midwestern states, the demand forecast for the 2026–2027 capacity auction increased by 5,250 MW, almost entirely driven by data centers.[3]
Dominion Energy is receiving requests from individual data center campus developers for "several gigawatts" of power and has proposed its first base-rate increase since 1992.[4]
The sub-synchronous oscillations produced by synchronized GPU workloads operate at timescales that SCADA systems cannot resolve. Utilities are planning for gigawatts of data center capacity but they are not measuring what this load is doing to the grid.
No one has published a comprehensive estimate of the total economic impact of AI data center load on the U.S. grid, in large part because the grid dynamics driving that impact are not being measured. The costs are real, distributed, and accelerating.
Bloomberg estimates harmonic distortion from data centers threatens billions of dollars in damage to home appliances and electronics. 3.7 million Americans live in areas where distortion exceeds industry thresholds.[1]
PJM clearing prices rose from $28.92 per MW-day to $329.17 per MW-day for the 2026–2027 delivery year. The independent market monitor attributed 63% of the increase to data centers.[5]
Utilities across the country are filing for rate increases driven substantially by data center load growth. The costs flow to every ratepayer in the region.
Sustained harmonic distortion accelerates insulation breakdown and overheats transformers. Equipment designed to last decades is operating under conditions it was never engineered for.
Grid reliability experts describe harmonics as a leading indicator of system stress. The downstream public safety costs of unmonitored grid degradation are unquantified but real.
You cannot price what you cannot see. You cannot manage what you do not measure.
Every hyperscaler negotiation, rate case, interconnection study, and public explanation comes down to the same question: can the utility prove what the load is doing to the grid?
Data centers create fast electrical behavior at timescales SCADA was not built to capture: harmonic distortion, sub-synchronous oscillations, rapid power swings, and localized power-quality impacts.
PredictiveGrid captures that missing signal continuously. Residential-edge power quality, substation sub-second behavior, harmonic content at the point of common coupling, and time-aligned context across the footprint a hyperscaler is loading.
Now the utility has the evidence base to plan, negotiate, regulate, and explain what is happening.
PingThings PredictiveGrid gives utilities the measurement layer this problem requires. Harmonic content at the point of common coupling. Sub-second substation behavior. Residential-edge power quality. The evidence base for every rate case, interconnection study, and hyperscaler negotiation.