I would like to know more detail about this claim.
The scenario you are describing is one were 64-128 OS threads are fully blocked waiting for IO. If that's the case, is it likely that you will have additional unused IO resources that could be being utilized?
Also, what overhead do you see as the main limit on spawning a lot of threads? Is that the CPU time of context switching? If so, in this scenario CPU is not the bottleneck, and switching between processes will be nothing, especially with a 32-64 core CPU.
This is a genuine question as I have never worked on an application that got close to maxing out either approach.