OpenAI rearchitects WebRTC infrastructure to deliver low-latency voice AI for over 900 million weekly active users

OpenAI's real-time AI team redesigned its WebRTC stack to handle three infrastructure constraints: one-port-per-session media termination does not fit OpenAI infrastructure well, stateful ICE and DTLS sessions need stable ownership, and global routing must keep first-hop latency low.

The new architecture uses a transceiver model—a WebRTC edge service that terminates client connections and converts media into simpler internal protocols for inference, transcription, and speech generation. The transceiver is the only service owning WebRTC session state, including ICE connectivity checks, DTLS handshakes, and SRTP encryption keys.

This design allows backend services to scale like ordinary services instead of acting as WebRTC peers themselves, and enables the service to run on Kubernetes where workloads can scale up and down as demand changes, avoiding the operational burden of managing large UDP port ranges per pod.