Cost per token, not raw speed, decides the right inference GPU. The formula, why cheaper cards often win, how utilization changes everything, and a worked example.