OpenAI rebuilds WebRTC for voice AI

The real story is that voice AI is now an infrastructure problem as much as a model problem. OpenAI’s answer is boring in the best way: preserve standard WebRTC at the edge, simplify the backend, and make routing deterministic.

• –The thin relay plus transceiver split keeps ICE, DTLS, and SRTP state in one place while letting inference services scale like normal backend services
• –Routing on ICE credentials avoids a hot-path lookup and helps preserve first-packet latency under Kubernetes
• –The fixed UDP surface is a serious operational win: easier to secure, load balance, and autoscale than huge per-session port ranges
• –This design is especially relevant for 1:1 voice agents where turn-taking latency matters more than multiparty media features
• –For developers building on Realtime API-style systems, the lesson is that protocol semantics at the edge beat custom client hacks