The AI Energy-Demand Forecast Scorecard: a reproducible audit of how the field forecasts data-centre electricity demand

A reproducible instrument that scores how the field forecasts AI / data-centre electricity demand — not a forecast of demand, and not a bubble call. Eleven published forecasts (IEA, Gartner, Deloitte, McKinsey ×2, LBNL, EPRI, S&P 451 BloombergNEF, Goldman Sachs, BCG) are scored on four signals: transparency, dispersion, revision behaviour, and heterogeneity. The dataset ships in full; reproduce.py (Python standard library only) computes every finding from the CSV — edit it and re-run, the numbers move.
  
Headline: of 11 forecasts, only 3 are reproducible from public method and data; 7 are confirmable at their primary  source and 10 verified anywhere — but the single highest figure (BCG ~1,050 TWh) cannot be verified at all. The forecasts use 6 incompatible formats, disagree by multiples at the US-2030 horizon without sharing a scope definition, and every documented self-revision points up. The field is, at present, easier to quote than to check — and that is the finding. 

 Changes in v0.2 (red-team pass). Finding 1's transparency count corrected to three nested, script-computed tiers —10/11 verified · 7/11 confirmable at the primary source · 3/11 reproducible from method+data (the earlier "9/10 confirmable at source" matched none of the data and is withdrawn). reproduce.py now computes the tiers rather than asserting them. Finding 4 clarified: three documented upward revisions, of which two are encoded numerically (BNEF +36%, Goldman +33% net) and EPRI's is documented in its native unit (a percentage, ~doubled); the EPRI magnitude was corrected from "~+60%". Each correction moved toward a sharper claim. Repackaged to portfolio standard —individual files plus a house-style PDF — replacing the v0.1 single zip. 

 Author: N. Milton (ORCID 0009-0003-4213-7769), independent. Produced with AI assistance (Claude Opus 4.8, Anthropic) for data retrieval, calculation, and drafting; the author set the question, method, and interpretation and verified every figure against the cited sources. As the assisting model is made by a company within the sector under analysis, that is disclosed for transparency. Open-science documentation only — not investment advice. CC BY 4.0.

https://nmairesearch.github.io/forecast-scorecard/