Most people are running a whole lot less than $20's worth of tokens per day on cloud platforms. (Is that assuming a frontier model? 1M output tokens per day?) Local hardware could easily take up that workload, at least the part of it that's non-time-critical.