2026-07-15 Comment response to the published Request for information and request for comments

See attached file(s)The Real Safety AI Foundation submits the attached comments on the Department's analytic methods. The Foundation is a nonprofit concerned with the measurement, disclosure, and governance of harms arising from artificial intelligence systems. It has no financial interest in any covered product or covered equipment and represents no manufacturer. The Department asks what data, models, and assumptions should support its analysis under 42 U.S.C. 6295(o). The Foundation comments on one input: the energy consumption data published by the firms operating the largest and fastest-growing electrical loads now connected to the American grid. In 2025 Google reported a 33x reduction in energy consumption for the median Gemini Apps text prompt over one year. In 2026 it reported its largest load growth in history, a 37 percent annual increase in electricity demand. Both are true. The first is a ratio, the second a total, and nothing in the first permits computation of the second, because the divisor has not been published. Four recommendations. 1. Gold Standard Science and the unreproducible ratio. The Department has committed this framework to Executive Order 14303, which requires that science be reproducible and transparent. Reproducibility is the first term in that definition. A ratio whose denominator is withheld cannot be reproduced by anyone who does not already hold the withheld number. Asked for the prompt count, the firm's chief scientist answered that the company was not comfortable revealing it. There is no allegation of error in the ratio. The Foundation recommends that the Department state that an efficiency figure published without the quantity necessary to compute the corresponding total does not satisfy the reproducibility element, and that it identify what it will do when the only available figures for a product class are of that kind, since declining to rely on them and having nothing else are the same position. 2. Data collection should reach 42 U.S.C. 6296. The Department asks what data it needs and observes that its approach persists because of limitations on data. For these machines the limitation is different in kind. The data exist, to a precision no survey could reach, inside a small number of firms. They are metered continuously because they are billed continuously. They are not published. EPCA already supplies the instrument, at 6296 for manufacturers and 6316(a) for covered equipment. The Foundation offers no view on whether the Department should exercise that authority, and recommends that the framework state whether it exists. 3. The utility impact analysis assumes a market that has stopped existing. Under the sixth statutory factor the Department estimates effects on the Nation's needed generation capacity, while also stating that because of limitations on data its analysis assumes historic market distributions persist over the 30-year time horizon. That assumption is being applied during the largest load growth in the sector's history. The Foundation does not ask the Department to forecast that growth. It raises the point because a utility impact analysis built on persistent historic shares will misstate the baseline against which every standard in every product class is measured, and will misstate it in the same direction every time. 4. NASEM Recommendation 4-9 cannot be satisfied for this sector on the current record. The recommendation directs the Department to consider direct rebound and, if it believes rebound to be minimal, to document the reasons why. Documenting reasons requires data. The Foundation takes no position on rebound magnitudes, which have been contested for four decades. It observes only that the documentation duty attaches whichever way the Department comes out, and cannot be discharged for this sector by any amount of analytic diligence on the present record. The comment alleges no bad faith. Google published more about the energy cost of machine learning inference than any comparable firm, and its 2025 report is why this comment can be specific. The comment takes no position on the Energy Paradox, discrete choice modeling, or the discount rate. It does not argue that EPCA authorizes a limit on aggregate energy consumption, and asks for no standard and no coverage determination. Full comments, with citations and supporting authority, attached. Respectfully submitted, Travis Gilly, Founder and Executive Director Real Safety AI Foundation, Cary, Illinois t.gilly@ai-literacy-labs.org | ORCID 0009-0007-2954-6313 The commenter is under contract to the AI and Environment section of Oxford Intersections (Oxford University Press) and has a paper accepted to the 2026 ACM/AAAI Conference on AI, Ethics, and Society. He publishes legal scholarship on the measurement and disclosure of harms from artificial intelligence systems. The Foundation previously filed comments with HHS Office for Civil Rights in docket HHS-OCR-2026-0133.