Which do you think is better: a large local language model or a cloud-based API?

A developer is considering switching from cloud-based APIs like Claude and GPT to a self-built on-premises inference system, citing monthly bills of $100 to $200 and an initial hardware cost of $2,000 that would be recouped in about a year. The developer is unsure if on-premises performance can match cloud quality for their tasks.

I’ve been using the Claude and GPT APIs for various projects, but the monthly bills are starting to get out of hand—they range from about $100 to $200 per month, depending on usage. So I’m considering switching to a self-built, on-premises inference system. The initial hardware cost is about $2,000. Based on my current API usage, I’d recoup that cost in about a year. But I’m not entirely sure if an on-premises model can truly match the performance of the cloud for the tasks I’m working on. I forgot to mention that the web version worked fine for me before—it was just that I ran out of quota. Do you think I need to set up an on-premises inference system?