AI energy forecasts may be missing large-scale inference demand by Benevolent_Rain

Up­date af­ter pub­li­ca­tion: Mo Putera pointed me to SemiA­nal­y­sis data on real Opus 4.7 agen­tic work­flows, which pushed my cen­tral es­ti­mates for both to­ken us­age and blended to­ken price sub­stan­tially down­ward. The origi­nal ver­sion used ~140M to­kens per 8-hour work­day-equiv­a­lent and ~$5 per 1M to­kens; I now use as­sump­tions closer to ~10M to­kens and ~$1 per 1M to­kens, but still with wide prob­a­bil­ity dis­tri­bu­tions. This makes the en­ergy bot­tle­neck look less im­me...