Since you obviously can’t use unified RAM, VRAM is going to make the most notable difference in locally run LLMs. But even if you do get a more powerful GPU, unless you drop thousands of dollars on one with more VRAM (on top of other components you may need to upgrade as well), you are, unfortunately, not going to have anything comparable to frontier cloud models.
The $20 a month is going to be far more affordable. Again, this is all by design so that you are locked into cloud models once you get hooked onto LLMs.