Running models locally is a moving target. By the time we have GPT3.5 quality at home we'll have better models in the cloud.

I'm sure the workloads will shift locally more and more, if for no better reasons than latency and privacy.