AI Model Energy and Carbon Disclosure: A Source-Audited Study of 103 Releases (2024-2026)

AI model cards and release documents are supposed to help people assess models. For energy and carbon, they often leave out the number a reader needs, or print figures that cannot be compared directly. I built the Release Energy Ledger, a source-traceable record of 161 distinct developer releases, and audited a selected set of 103 dated 2024-2026 releases that passed a page check. On the designated release document, 85 (82.5 percent) print no energy, carbon or compute figure, 1 prints compute only, and 17 (16.5 percent) print an energy or carbon-related figure: 16 training-specific and 1 life-cycle footprint. None prints release-specific inference energy. Meta accounts for 10 of the 17; without Meta, the count is 7 of 93. The figures differ in method, unit, equipment boundary and emissions accounting, and only 8 of the 17 state a basis, a boundary and a method together. Changing the unit of analysis gives 11 of 80 documents, 10 of 59 model families and 7 of 23 developers. The strict-quote subset (5 of 74) is a different evidence frame, not a lower bound. A page re-read confirmed the printed figures for all 17; blinded AI coding agreed with the ledger on 91.9 to 100 percent of fields, and a separately instructed AI adjudicator agreed on 42 of 42 selected cells. These results describe the documents in this selected set, not actual emissions, the universe of releases or legal compliance. No human has validated any label. The ledger and counting script are released with the paper as a versioned data record (https://doi.org/10.5281/zenodo.23808341).