Google has announced a next-generation Federated Learning (FL) system that provides externally verifiable privacy guarantees while improving computational efficiency. The system, which leverages Trusted Execution Environments (TEEs), marks a significant advancement in how machine learning models can be trained on decentralized, private data without requiring users to trust the server operator. Gboard, Google's keyboard app, has already adopted the technology and is seeing substantially faster compute times compared to the previous FL system. The new architecture introduces four core operational elements: encrypted local data uploads from devices, a Key Management System (KMS) that verifies which server-side computations can access the data, isolated workload execution within TEEs, and fault-tolerant recovery mechanisms. Crucially, the system publishes access policies to Rekor, a public transparency log, allowing external auditors to monitor which server workloads devices are participating in. The KMS and data-processing binaries are reproducibly built from open-source code, enabling independent verification of the system's privacy guarantees. This advance addresses a long-standing challenge in federated learning: how to prove that device data was never logged, inspected, or misused by the server operator. By combining TEEs with differential privacy algorithms and secure aggregation, Google has created a fully verifiable end-to-end FL system. The approach has immediate applications across Google's consumer products while establishing a model for how enterprise organizations might deploy similar privacy-preserving machine learning infrastructure.