OpenAI’s voice AI only feels natural when conversation moves at the speed of speech — and network lag kills that instantly. To handle over 900 million weekly active users while keeping interactions crisp, the team rebuilt their WebRTC infrastructure from the ground up. The old one-port-per-session model didn’t scale with their infrastructure, and stateful ICE/DTLS sessions needed stable ownership. Their solution? A split relay plus transceiver architecture that keeps standard WebRTC behavior for clients but changes how packets route inside OpenAI’s servers. This matters because WebRTC is the backbone for real-time audio in ChatGPT voice, the Realtime API, and interactive agents. By building on this open standard, OpenAI avoids reinventing media transport protocols and focuses on connecting audio to their models. The result is faster connection setup, lower jitter, and more natural turn-taking. Why it matters: As voice AI moves from novelty to necessity, infrastructure that handles real-time audio at planetary scale separates the products people actually talk to from the ones they abandon after one awkward pause.