The pricing argument that open-weights advocates have relied on assumes substitutability — that an enterprise seeing GPT-5.5 at $5/$30 per million tokens can credibly threaten to self-host Llama or Mistral instead [4]. That threat requires GPU capacity, operational teams, and an evaluation and fine-tuning pipeline that the typical closed-API customer does not maintain. When the cost of building that infrastructure , the substitution calculus collapses before it reaches a procurement meeting. The pressure that economic theory predicts never arrives at the…