> Realistically if you use a self-hosted LLM for your job, you might be using it, what, a solid 6 hours per day?

You can grow the utilization rate well beyond that if you don't always care about getting a quick, real-time response. (And if you do, then maybe the cloud model was the better deal after all!)