That determines the cost effectiveness to make it worth it to train one of these models in the first place. Using someone else's weights, you can afford to predict quite inefficiently.