It's very clear that more computation (and ram) on larger models will produce better results when talking about general model use. This will be true for a while.

Eventually the marginal benefits might plateau in combination with enough optimizations to make local use outweigh any cloud models.

More specific and narrow use cases are a different matter.