True.

The problem is serving is a skill readily mastered by the hyperscalers. That's their MO.

All they need is weights to serve. And the open models provide that.

OpenAI is relatively well placed in that they have inference chips they've designed and they own compute.