On the contrary I think in a few years most inference will be done locally. The smaller models have been slowly catching up to frontier for many tasks, and chips are being developed to run large models on phones and laptops (according to Qualcomm CEO they want to run 10T weights on a mobile chip).

To me the simple answer to why is control. If I run a model locally I get to control the model params, the system prompt, all of it. And of course there is no quota. But probably most importantly I am not tied to paying some company a monthly subscription forever.