I think focusing on end-to-end time confuses things more than it helps. A system can have 10X throughput with the latency being unchanged. You don't need to reduce latency or cycle time to have a 10X increase in throughput.

The better argument is that Software Engineers spend a lot of time doing things that aren't writing code and arent being accelerated by any AI code assistant