The issue is that local models are dumb and tend to make mistakes than look good at a first glance. So any "saving" is quickly ruined by having to do an extensive review. You might as well just write things yourself.

I use it as code scaffolding, which means in a way I’m often rewriting it. For me, writing from scratch isn’t the same amount of effort as using a code scaffolding tool.

I disagree, even though I'd love for it to be different. With models like Opus, I can give it a good architecture and expect good results. For many of the less expensive models, that is not the case, they make mistakes, you need to over specify, they get stuck in a loop, etc. As you get to the models you can realistically run locally, it gets so frustrating I'd rather be writing the code myself.

At what point will local inference catch up to today’s cloud inference? Will it ever? If it doesn’t, does that imply a certain dead-end for the LLM inference industry?

I don't think at any point in foreseeable future we will have terabytes of RAM for dedicated LLM chips at home.