I maintain that the “dominant” game theoretic move here as a frontier model maker might be to cease all model spending, invest only in harnesses (C Code, Codex, etc) and mint money on enterprise subs. The deep question is to what extent orchestration layers and models must be integrated.

I wrote about this recently, in part, here:

https://paulkedrosky.com/commoditization-orchestration-and-the-new-ai-stack/