Even more important in a local context is the difference between token generation and prompt processing speed. We tend to focus on the former, but for multi-turn/agentic workflows the latter can dominate.

Yeah definitely. I've recently commented on that: https://news.ycombinator.com/item?id=48557890