The fight over whether artificial intelligence deserves the power grid gets decided next week, and not by anyone arguing about it online.

On July 7 and 8, the Federal Energy Regulatory Commission convenes a technical conference under docket AD10-12-017 on increasing market and planning efficiency through improved software. [1] The paper argued a week ago that the real contest here is not whether AI deserves power but which load forecast gets believed. This edition marks the moment that contest goes on the record: the supplemental notice set a June 30 deadline for second-day presentation slides, with timely submissions posted before the conference. [1] The materials that will shape how the AI build-out is planned landed yesterday.

Why a software conference is a technology story, and not just an energy one, is in the commission's own framing. Its notice says load forecasting has become more important because the bulk power system faces historic demand growth driven in part by data centers and other large new loads. [2] Translated: the number that decides what transmission gets built, what generation gets approved, and who pays for it is a forecast — and forecasts are assumptions wearing spreadsheets. Change the assumption about how many gigawatts of AI load are real, and you change the grid that gets built to serve it.

The conference agenda is not a referendum on AI. It is saturated with it. The original March notice says the second day may include software applications, including artificial intelligence or machine learning, to improve the efficiency and affordability of the bulk power system, alongside software for operations affected by the interconnection of large loads, resource adequacy, and energy adequacy modeling. [2] The June supplemental adds grid-enhancing technologies to the July 7 panels — the controls and hardware that squeeze more capacity out of lines already in the ground — and reserves July 8 for individual presentations, the slot where one vendor's model can be pulled apart on its own terms. [1] The June supplemental narrows the calendar to grid-enhancing technologies, load forecasting, and improved software, with panels on the seventh and individual presentations on the eighth. [1] AI is both the load being forecast and the tool proposed to forecast it.

This is where the online version of the story goes wrong. On X, a data center is a villain or a trophy — a warehouse of GPUs siphoning power from households, or a national-security asset being throttled by regulators. Mainstream coverage tends to flatten the same event into a grid-strain headline: demand is up, wires are short, brace for it. Both skip the mechanism. The commission is not staging a morality play about whether AI is worthy of electricity. It is asking whether the models used to predict AI's appetite are good enough to justify committing billions in transmission and generation to feed it.

The stakes of that quieter question are larger than the loud one. A forecast that overstates data-center load builds transmission that ratepayers fund for demand that never arrives. A forecast that understates it leaves the grid short when the load shows up, which is how you get the reliability crisis the alarmists keep predicting. Either error is expensive, and both are decided in the assumptions a vendor, a utility, or an academic can now be made to defend in a public docket rather than a press release.

That is the accountability the conference creates. If a software firm wants to claim its tool makes new AI load cheaper or safer to serve, the claim belongs in the conference materials due this week, where a regulator and a rival can inspect it. If a utility wants to argue that data centers make forecasting impossible, its assumptions are visible in the same file. The July 7 meeting turns a slogan war into a modeling argument with names attached.

The useful question is not whether AI should get the grid. The market has already answered that with capital. It is whose forecast the grid is allowed to believe, because that forecast — not the discourse — will set the transmission, the rates, and the siting for a decade. Next week, in a hearing room most of X will ignore, the assumptions get read into the record.

-- THEO KAPLAN, San Francisco