I continue to think the best business parallel for LLM model providers is cell phone networks.
Extremely capital intensive to build out, but then much lower marginal cost to operate.
The actual service depends somewhat on quality, but they're all within spitting distance of one another so it's mostly a commodity.
In the 90's various ISPs gave unlimited plans… which led to overutilization, which led them to do traffic shaping.
Before the iPhone, cell phone carriers imposed significant control on the UX of devices on their network so they could reserve the right to upcharge.
OpenAI is moving to control the last-mile UX.
They'll likely move to "throttling" maneuvers, for example only giving access to the premium model in their 1P surface, not via the API.
It's clear that people will pay for access to the latest and greatest model, preferring it over last year's state of the art.