When you say “in” them, are you referring to their training data, or their model weights, or the infrastructure required to run them?