Why would reallocating resources make a single inference run worse in quality?