It doesn’t have to be the latest fancy gaming card - I run an old 2016 Tesla P100 ($200AUD~) which is fine for 16B q4 models.
They’re more configurable, there’s no subscription model, you contribute to them and meaningful ways, there is a privacy concern if you’re passing your own data and it’s perhaps strangely a more fulfilling experience.
I’ve got mine hooked up to Home Assistant so I can automate things with Siri and shortcuts, I’m using it finding patterns and code and several decades of digital notes.
If you’re not sure where to start, you can try some models out locally on your machine, using apps like https://lmstudio.ai or GPT4All with almost no ML knowledge. Then if you want to level up and customise things and run them on the server you can use localAI, Text generation web UI etc which provide openAI compatible APIs.
A good 13B model with decent prompting feels pretty on par with chatGPT (3.5) for general tasks, but can be configured to be quite good at very specific things.
https://github.com/oobabooga/text-generation-webui