The throughput of our processes are limited by the slowest part of our process pipeline.

How is the lantency defined in this case? Is it infinity? But I think we can delay the transfer stage (orange block) to decrease the lantency.

I think we can think of latency as the time it takes when a request is sent until it is fulfilled, so even if the transfer stage is delayed, latency would not be decreased because the delay should be considered as a part of the latency.