OpenAI Publishes Split-Relay WebRTC Architecture for Voice AI at 900M-User Scale — Go Implementation, SO_REUSEPORT, Thread Pinning

OpenAI Engineering·medium signal

OpenAI engineers published a technical deep dive on their rebuilt WebRTC infrastructure for real-time voice AI. The 'split relay plus transceiver' architecture separates packet routing from protocol termination: a lightweight UDP relay forwards packets without decrypting media, while a stateful transceiver handles ICE/DTLS behind it. This lets OpenAI run WebRTC in Kubernetes without exposing thousands of UDP ports. Built in Go with SO_REUSEPORT, thread pinning, and low-allocation parsing. Directly applicable to anyone building voice agents on the Realtime API.