OpenAI scales low‑latency voice AI by splitting WebRTC into edge relay and transceiver
OpenAI keeps voice AI responsive for 900 M weekly users by decoupling WebRTC into a stateless edge relay and a stateful transceiver, using the ICE ufrag as a routing key. This bypasses port‑exhaustion and state‑sync limits in Kubernetes, delivering millisecond‑level latency essential for natural‑speech interaction.