The AI Energy-Demand Forecast Scorecard: A Reproducible Audit of How the Field Forecasts Data-Centre Electricity Demand

A reproducible audit of how the field forecasts AI and data-centre electricity demand, scoring twelve published forecasts from named institutions on five signals: transparency, method class, dispersion, revision behaviour, and comparability. It measures forecaster behaviour; it is not itself a forecast, and not a bubble call. Of the twelve current forecasts, only four are reproducible from public method and data, eight are confirmable at their primary source, and eleven are verified against some credible source, while the single highest figure (BCG, about 1,050 TWh) cannot be verified at all because it is paywalled. The forecasts are expressed in five incompatible formats, and even inside the one slice where six figures share a country, a unit and a horizon, they disagree about whether cryptocurrency is in scope. They span 2.8 times at the US 2030 horizon, or 2.2 times excluding the unverified row. A companion reproduce.py computes the scorecard table and all five signals from the shipped dataset, so the headline figures are derived rather than asserted. Version 0.3 is a material refresh. The v0.2 finding that every documented revision pointed upward is WITHDRAWN and restated: LBNL's 2025 Update, published on 18 June 2026, revises its historical estimates downward and restates its 2028 projection inside its own prior range, giving a tally of three upward revisions, one downward and one flat. The roster has been rebuilt on primary sources, moving LBNL to the 2025 Update and EPRI to its February 2026 edition, adding the IEA's US base case, and retaining both superseded editions as a flagged revision trail. US 2030 dispersion accordingly falls from about 5 times to 2.8 times, and the earlier spread is shown to have been bounded at both ends by the two softest rows in the table. Two further findings are added: the field's base rate, drawing on two prior energy-demand panics that did not materialise, and a scope mismatch that survives inside a single comparable slice. An earlier claim that the McKinsey 2025 figure would be scoreable against EIA actuals is corrected: no official US measurement of data-centre electricity exists or is currently collected, so this remains a process scorecard rather than an accuracy one. Independent analysis, not investment advice. Produced with assistance from Claude Opus 4.8, an Anthropic model; Anthropic is a participant in the AI industry whose electricity demand this work assesses, so the assisting model is not a neutral party to the subject matter and the primary sources cited should be preferred to this summary.