This research paper proposes a new architecture for federated learning of large language models (LLMs) over mobile networks, addressing challenges in mobile RAN transport. The proposed approach uses in-network aggregation to reduce the number of transfers and enable selective provisioning of optical connectivity.