A recent Towards Data Science piece measured the power draw of an RTX 3090 and found that cost per token tracked throughput, not parameter count. The author replicated the experiment on an M3 Ultra Mac Studio with 96 GB of unified memory using a custom tool called TokenWatt, which reads whole-SoC power rails through Apple’s IOReport interface and subtracts a rolling idle baseline. The tool was calibrated against a Shelly Plug US Gen4 that meters actual energy at the wall, with error bands between ±2.6% and ±4.5%. The author’s electricity rate is a flat $0.31/kWh. Five models were run in a sustained generation loop at three durations (120, 360, and 720 seconds), three passes each, with a 15-minute idle baseline between runs. The finding that the surprise holds on Apple Silicon and gets bigger: the author’s 120-billion-parameter model is about five times cheaper per token than a model a quarter its size. A wall-verified figure of “$0.109 per million tokens” is cited with a stated uncertainty.