but I'd assume for real world energy usage the KV cache would be very impactful if not eventually dominant.

They talk about this in the appendix where they go over the (estimated) effects of large amounts of input tokens (up to 100k). This isn't really relevant for Gemini Nano because it only has a max 32k context window, and the deployment in Chrome probably caps it at far less than that.

I'm inclined to believe the main analysis is reasonably accurate. The numbers are similar to what I get on my local machine with local models. Granted, I tested with smaller models (7b parameter Mistral in this case) on weaker hardware (AMD 6700XT), but on a quick test I get about 50 tok/s locally at 180 W power use, which is about 0.5 Wh for 500 tokens. AMD GPUs suck for AI, so I think it's plausible that dedicated compute hardware would get basically the same energy efficiency on a frontier model.

Gemini Nano on a phone NPU is obviously going to be far more efficient -- by all accounts it gets the same or better tok/s I am getting at like 1/50th the TDP.