US labs charge about seven times what a token costs them to serve. This is the story of that gap — how it opened, who pays for it, and whether the trillion dollars behind it ever comes home.
Training is a fixed cost paid up front; serving is a real cost that grows with every answer.
US prices sit far above serving cost — an 80% cut on an unchanged model proves it.
China prices near cost because weights are open and hosts compete, not because inference is cheaper.
Chinese labs push the fixed cost onto clouds, hedge funds and the state.
Convergence hasn't started; the flagship price rose. This is the claim carrying the capital risk.
The token is the wrong ruler; cost per finished job is what matters.