I read the paper for this, and I think it's far more of a propaganda exercise than an attempt to provide real or meaningful data. The biggest give away is that they choose to give median figures rather than the means they say they collect, and no indication of the distribution.
This figure is probably massively misleading since it is probably based on numbers from "Gemini Flash" which they use to handle a subset of queries and runs at much lower energy levels than the broader models There are other things I take issue with, but the choice of median is the biggest red flag