I believe you when you say you're not changing the model file loaded onto the H100s or whatever, but there's something going on, beyond just being slower, when the GPUs are heavily loaded.

I do wonder about reasoning effort.