I think you missed the rate limits. they are way to low to use this model in production to iterate over a big dataset each day to extract features or what not. It is simply not usable- With a self hosted approach you can set the rate limit to whatever you want, which makes it much more usable.

And who do you think will set up the openai api in production to make use of the model --> Mlengineers. no business person will have the time to make prompt engineering and connect this model to services in the data lake of the company. This is kind of biased from my perspective.

A lot of people confuse the Chat GPT interface with the openai api. If you just want to use the first one, go ahead and just use it. But with the second one, no business people is involved anymore and it becomes either way an engineering problem for both, openai model and self hosted llm (finetuned one).