Step-by-step guide to testing energy efficiency of local AI models

I compared the costs of "running models locally" versus calling cloud APIs directly. After actually running through it, here's the bottom line: local AI isn't some mystical art, but whether it's "worth it" needs to be backed by data. A few days ago I came across Stanford's Intelligence per Watt paper; they put "intelligence volume" and "power consumption" into a single formula, which is a pretty straightforward approach. I gave it a try myself and organized this workflow for beginners.