← NM AI Research · research portfolio
How the field forecasts data-centre electricity demand — scored on dispersion, the forecasters' own revisions, and transparency. A scorecard of forecaster behaviour, not a forecast.
Interactive front-end over the frozen v0.2 dataset · data as of 2026-06 · canonical record: DOI 10.5281/zenodo.20572928 · NM AI Research, ORCID 0009-0003-4213-7769 · CC BY 4.0
Forecasts view. The forecasts are deliberately left in their published units, scopes and horizons. Dispersion ("they can't all be right", measured) is computed only within a comparable slice — same scope, metric, unit and horizon. Narrow Scope + Unit to compare like with like; leave them on "All" and the banner reports how many incompatible formats the field is using, which is itself a finding (heterogeneity).
Transparency funnel. Three nested tiers computed from the table: verified (the figure can be confirmed against any credible source) ⊃ confirmable at the primary source (the author published it themselves) ⊃ reproducible (the model can be rebuilt from public method + data). The gap between verified and reproducible is the headline: most figures can be quoted, few can be rebuilt.
Revisions shows where a forecaster changed its own number; all logged revisions point up. A third revision (EPRI) is documented in its native unit — % of US electricity, 4.6–9.1% (2024) to 9–17% (2026 ed.), roughly doubled — but not encoded numerically here because this row's TWh is itself derived from that %.
This scores how the field forecasts; it is not a forecast and takes no view on which number is right. "Opaque" means the model is proprietary / not reproducible, not that the figure is wrong — several opaque figures are confirmed at the author's own release. BCG's 1050 TWh is the lone unverified entry (a separate paywalled study, not the in-hand BCG PDF). Units and scopes are not interchangeable: TWh, GW and % are different things, US is not Global, 2030 is not 2035.
Disclosure. No third party reviewed, funded, or directed this work. The author holds no position that depends on any of these forecasts being right.
Reproducibility. This page embeds the frozen forecast_scorecard_data.csv verbatim (regenerated by build.py); the full reproduce script and write-up live in the Zenodo record. Edit the CSV, run build.py, the findings move.