How often does a proposed power plant actually get built, and how late? A base rate for 11,343 US generators from the stacked annual vintages of a statutory register

Abstract

Anyone pricing a first-of-a-kind industrial plant needs two things: how much it will produce, and whether it will exist at all. The second question is usually answered with anecdotes, because the base rate they should be compared against has not been published. A pioneer plant that gets built 70 per cent of the time is doing badly if ordinary plants get built 95 per cent of the time and well if they get built 60 per cent. This note publishes that denominator for the United States, from a source that was never designed to answer the question. The US Energy Information Administration's Form EIA-860 collects, every year, a list of operating generators and a list of proposed ones. Neither list is a study, but the annual vintages stacked are one: membership is a dated, mandatory filing and does not depend on the outcome. Fifteen vintages (2010-2024) yield a closed cohort of 11,343 proposed generators, of which 8,936 are resolved into built, abandoned or still censored. Between 84.5 and 86.6 per cent get built, the two ends being two definitions of built rather than two estimates. The median generator arrives two months late; 32.1 per cent arrive on time, the 90th percentile is 19 months and the maximum 207. The build rate falls 32 percentage points from the smallest size bracket to the largest, monotonically past 20 MW, among conventional plants alone. Cutting the rate by the phase the project declares makes it consultable at the moment someone has a specific project in front of them. Over 12,421 annual filings the probability of being built rises monotonically with phase, from 54.6 per cent for a project that has not started permitting to 97.8 per cent for one whose construction is complete. Punctuality does not: it inverts depending on which date the project is judged against. A generator declaring construction complete arrives on time 61.2 per cent of the time against the date it was itself declaring that year and 17.5 per cent against the first date it ever declared. Part of the mechanism is selection that the register cannot avoid: a project that finishes and enters operation between two annual filings never appears in that phase at all. An advanced phase therefore predicts existence, not punctuality. Validated out of sample the table reaches an AUC of 0.833, but declared lead time alone reaches 0.826 and phase alone 0.737, so the intuitive variable adds almost nothing to the boring one. Five threats to the number are measured rather than declared, including one that moves it: because both matching rules work within a single plant code, neither sees a developer file a unit under a new code, and at least 44 of the 1,424 abandonments are code reassignments rather than abandonments. That correction is a floor and not an estimate, since the screen cannot see reassignments that came with a change of rating. What the note does not claim is stated as plainly: it measures generators already carried on the form, so the figure is an upper bound on the survival of an announced project; it is not comparable with interconnection-queue completion rates, which measure an earlier stage; and it says nothing about whether a plant, once built, produced what it promised. Version 1 of 30 July 2026 was not deposited, so this record begins at version 2. The full cohort, the intermediate tables and the scripts that produce every figure are included.