The Invisible Tax on Every AI Word: Why Inference Efficiency Is the Battle of 2026