Choose, not Hoard: Information-to-Model Matching for Artificial Intelligence in O-RAN