GPU idle rates above 30%, operational headcount scaling linearly with cluster size, and no visibility into recurring costs. The AI infrastructure build-out has a margin problem, and the fix is ...