Fresh intel from state utility regulatory filings

Public utility commissions are where electricity actually gets decided, and they publish an enormous amount in the process. The problem is the form: esoteric, long, hard-to-read PDFs, with the redaction landing precisely on the number you want. Halcyon’s approach is to use large language models to pull structure out of these documents at scale, which is what makes a cross-state canvass possible. The first thing that falls out is scale. Pennsylvania’s commission convened an en banc hearing on large load interconnection, meaning all the commissioners sitting at once. Bullard had never seen one; an advisor told him it almost never happens outside of something market-defining like restructuring a generation fleet. It pulled every intervenor and expert witness into one room, and the testimony is where the eye-popping numbers sit. Duquesne Light, with a relatively small service territory, said four projects in advanced stages of connection could add 40% to its demand, and one large hyperscale project could add 30% on its own. Rappahannock Electric Cooperative in Virginia has a 1.2 gigawatt summer peak and individual interconnection requests four times that size.

Numbers like that force a regulatory construct that does not exist yet, and the dockets are where it is being invented. Bullard reduces it to two questions: how does the build get paid for when you are building on behalf of a handful of customers rather than a hundred thousand, and what happens if the customer never shows up. The second is the one he thinks has been left out of the conversation. Anyone in power generation takes for granted that not every announced project gets built, but that discipline has never been applied to the demand side, and asked whether every planned hyperscale data center on paper will be built, Bullard’s answer is flatly no. Kann extends it a step: even the ones that get built are being underwritten by tariffs that assume payment over a decade or two, and that assumption can change fast.

The dockets also show the second-order effect, which is incumbent large loads elbowing into proceedings that are ostensibly about data centers. East Kentucky Power Cooperative is working a custom large-load tariff through its commission, and Nucor, the mini-mill steel producer and the region’s biggest load, filed as an intervenor. Bullard reads that as a useful reminder that other large loads exist, alongside things like the Ford and SK On battery plant nearby, and as a caution about a world where everybody gets a bespoke tariff. Kann pushes it further, into an inference the filings do not make: because large-load interconnection capacity is genuinely limited, anything that is not a data center is now competing with data centers for sites, and he says it is getting harder to site manufacturing, industrial electrification, even medium- and heavy-duty fleet charging. Bullard adds the asymmetry underneath that. Data centers are a class of customer almost uniquely underexposed to the price of electricity relative to what access to it is worth. Raise power prices 30% and the cost of steel from an electric arc furnace moves materially; do the same to a training run, where the premium is on speed, and it barely registers. That is a clean argument in principle, and when Kann asks the obvious follow-up, whether large-load tariffs are in fact being set at a premium to other industrial rates and whether data centers are accepting them, the filings go quiet. Bullard says it is partly redaction and mostly that these are early and under deliberation, with the terms that matter unknown: whether the premium is an adder to an existing rate or a new flat rate, whether it escalates, how long it runs.

The rate side is where the filings are most legible and least reassuring. Bullard deliberately looked at small service territories, which get less scrutiny, serve a smaller customer base that is often almost entirely residential with few large ratepayers, and can go a decade between rate resets, so the reset arrives as a nonlinear jump rather than a drift. A Vermont village electric and light department asked for roughly a 21.5% increase effective July 1 on a timeline with none of the years-long California-style process, and the state Department of Public Service pushed back that it needed a conference first. A Mississippi operator is collapsing three residential tiers into a single statewide flat rate for water and sewer, which lands hardest on whoever was in the lowest tier. Bullard’s point is that rate increases are inherently regressive and that these are appearing everywhere at once. The gas cost work is the counterweight and the episode’s best demonstration of why this mining exercise is worth doing. The NextEra chief executive’s roughly $2,400 per kilowatt figure for a new gas plant, delivered from a conference stage, had become gospel through third-hand repetition. Bullard set out to demystify it and found the opposite of a debunking: canvassing more than 100 plants and about 55 gigawatts across all 50 states, he gets a figure in the low $2,200s per kilowatt for plants landing in 2031 and 2032, as best the garbled transcript can be read, legitimately double the price of a couple of years earlier. The figure that fails is the official one. The Energy Information Administration’s 2024 cost-of-generation benchmark of about $1,100 to $1,200 per kilowatt is, in Bullard’s words, a conclusion not found in evidence.