Another way to word this is that the models use less energy for English texts. This can be optimal depending on the distribution of languages the model is actually using, because it means overall energy is less than it would be if the model used equal energy for all languages.