An analysis, with arithmetic run by Opus 5, estimates AI "thinking wattage" against the human brain across three measures. At human pace (~5 tokens/s), inference on Jalapeño- or Vera Rubin-class hardware draws 3–20 W versus the brain's 20 W — parity reached around 2024–2025. Raw operations per joule place silicon within 2–10× of synaptic efficiency. But per delivered work, agentic sessions run 30–700 W, 2–35× worse. Fully loaded human energy footprints are kilowatts, and training amortizes better than childhood.
No score is assigned. Sources and their independence are shown in the citation chain below.