it's interesting that you posed the question whether or not you should spend $20,000 to run your own models locally. I've been asking myself that exact same question the last few weeks, specifically whether or not I should spend $500, $1,500, $3,000, or more?
What I came up with was that the smart move right now is to stick with my limited 6 GB GPU, use my Google workspace subscription, use my Google cloud free tier availability, and lease time on vast ai