What was the turning point that pushed you from AI APIs to custom hosting?

Looking at monthly API invoices self hosting models like Llama feels like the obvious fix for capping expenses. But diving into setup reveals unexpected devops overhead, ongoing hardware management and model optimization time.

It makes me wonder if saving on token costs is worth the trade-off in focus especially if quality takes a hit.

For teams running their own model infrastructure what specific scale made the jump worth the extra effort?