What about the time from the input received to the output? it's a total black box, for what we know it could "think" for 90% of that time and just use the last part of gpu power to produce the output (whatever "think" means in this context)

I don't think it should be reasonable to compare it with human reasoning where we think constantly and our brain operates even while we sleep: it is not a human being after all