What we got wrong, and what changed.

Site vsite 1.1

Visual redesign: card-based layout, inline illustrations, newsroom avatars, purple system across the front page, trust pages, the Power Index and both ledgers. No data or copy in the Power Index changed. About page wording updated to describe the operator as a research-harness company rather than an agency project.

AI Data Center Power Cost & Grid Stress Index v1.0

First public release on research.whatsthehost.com. Data through 2026-08-17 (EIA-930), May 2026 (EPM Table 5.6.A), eGRID2023 Rev.2. Page data externalised to index.json, claims.json and inferences.json; CSV export added to the scoring script so the dataset is reproducible from code.

AI Data Center Power Cost & Grid Stress Index

The Columbus, OH row's local-constraint text in the table and the CSV still read "at least 85% of contracted capacity … with a three-year exit fee" after claim C-026 had been corrected to AEP's first-party wording on 2026-08-17. Corrected to: at least 85% of the energy they are subscribed to use, over 12 years including a four-year ramp; the exit-fee amount is not stated in any first-party source and is not published. Process change: a ledger correction must now be grepped against every rendered output before a story is re-published.

AI Data Center Power Cost & Grid Stress Index — claim ledger

7 blockers found, 7 fixed. Quantitative pass found the percentile population was contaminated by duplicate metro rows and the year-over-year window was misaligned, biasing all growth figures down by roughly one point. Editorial pass found one paraphrase stored as verbatim, one status mis-stamped, one claim mis-dated by 13 months with the wrong institution, one federal order mischaracterised, and several counter-positions attributed to named living people who had not made them. All fixed; every fix is recorded in the discrepancy field of the affected claim.

AI Data Center Power Cost & Grid Stress Index — claim ledger

1 blockers found, 1 fixed. Supplied a working eScholarship PDF URL for C-002, which had been failing with a 404 since the first audit. Re-verified the same day: all three cited figures present verbatim. Status moved FETCH_FAILED → VERIFIED_LIVE. The re-check also corrected our own 2014 baseline note (60 TWh, not 58) and surfaced C-073, the report's author list.

AI Data Center Power Cost & Grid Stress Index — claim ledger

3 blockers found, 3 fixed. Joe pushed back on giving up on dead URLs and asked for site-scoped and title-based searching. Three unverified claims resolved as a result, and two of them corrected substantive errors. C-036 re-sourced to EPRI's own release, replacing our computed 'roughly double' with EPRI's own '60% higher than prior 2024 estimates'. C-037 re-sourced to Tyler Norris's ISO-NE presentation, which gives 98 GW at 0.5% curtailment where we had published 'roughly 100 GW'. C-026 re-sourced to AEP's first-party release, which showed the 85% applies to subscribed ENERGY not contracted CAPACITY, that there is a 12-year term we had omitted, and that our 'three years of minimum charges' exit fee was unsupported and has been removed. Lesson recorded: a 404 on a landing page is a routing problem, not an evidence problem.

AI Data Center Power Cost & Grid Stress Index — inference ledger

I-02 downgraded HIGH to MEDIUM; Goodhart removed as a decorative model; state-versus-RTO, nominal, preliminary and confounder caveats added; a mechanistic falsifier added alongside the correlational one.; I-03 downgraded HIGH to MEDIUM; the stated historical base rate was half wrong and is corrected to chronic over-forecasting; the Amara application contradicted its own evidence and was removed.; I-01 downgraded HIGH to MEDIUM; a two-year forward prediction cannot meet this ledger's definition of HIGH; falsifier replaced with three measurable published series.; I-14's headline 3.99 cent gap could not be reproduced from the study's own data. The metric is now defined explicitly (simple mean of four named states, one named EIA table) and the figure corrected to 4.96 cents.; Six 'who disagrees' fields attributed constructed positions to named living people without citation. All rewritten to record counter-arguments as arguments unless a document supports the attribution.; Mark Christie is described as a former FERC Chairman, which he is.; I-11's Jevons application was a slogan rather than a demonstrated chain, and an uncited DeepSeek 'natural experiment' claim was removed.; Falsifiers for I-01, I-04, I-05, I-09, I-12 and I-13 were unmeasurable or tested the wrong proposition; all replaced with dated, named, publicly observable tests.; C-062 struck from I-08's evidence as an unverified allegation.; A known_gaps section was added listing seven substantive objections neither ledger addresses.

AI Data Center Power Cost & Grid Stress Index — inference ledger

C-002 re-sourced to a working URL and re-verified; status FETCH_FAILED → VERIFIED_LIVE.; C-073 added: Koomey's co-authorship of the report whose range he is quoted doubting. Folded into I-03 as evidence, with the reading that the ranges are honest and the downstream point-citation of them is not.

AI Data Center Power Cost & Grid Stress Index — inference ledger

I-07's headroom evidence updated to the precise 76 GW / 98 GW figures from Norris's own presentation, replacing the rounded landing-page number.; I-03's EPRI evidence replaced with EPRI's own '60% higher than prior 2024 estimates' framing and unflagged, now that C-036 is verified.; I-09's Ohio evidence unflagged with the corrected tariff terms.