OpenAI exposes voice AI low-latency technology: WebRTC + Kubernetes + Go supports 900 million users

Core conclusion

OpenAI recently released an in-depth technical article "How OpenAI delivers low-latency voice AI at scale", which fully discloses the real-time communication architecture behind the ChatGPT voice mode. Core points: OpenAI uses the Pion WebRTC library written in Go language, combined with Kubernetes to deploy relay nodes in 22 regions around the world, achieving millisecond-level voice interaction latency.

The article received 220+ likes on Hacker News, sparking extensive discussions about WebRTC architecture choices, Pion library maturity, and voice AI product experience.

Key Points

• Event time: 2026-05-04 -Affected objects: AI voice application developers, ChatGPT voice users, WebRTC technology community
• Core change: OpenAI abandons the traditional one-port-per-session model and uses the transceiver architecture to implement a single-port solution

Background and trigger events

OpenAI’s Advanced Voice Mode is one of ChatGPT’s most popular features. However, getting an AI voice conversation to respond within hundreds of milliseconds involves an extremely complex real-time communications technology stack.

The traditional WebRTC model is designed for point-to-point communication (each session requires a separate port), but OpenAI needs to handle concurrent voice traffic of hundreds of millions of users, and following the traditional solution will lead to port exhaustion and operation and maintenance nightmares.

Technical takeaway: OpenAI chose Pion, a pure Go implementation of the WebRTC library, as the core networking component instead of Google's libwebrtc C++ library. This choice allows OpenAI to efficiently handle tens of thousands of concurrent connections using Go’s goroutine model.

Key Impact (by Dimension)

Adaptation suggestions

If you are an AI voice application developer

1. Evaluating Pion as a WebRTC solution: The selection of OpenAI demonstrates the feasibility of a pure Go WebRTC library in large-scale production environments.
2. Focus on the pipecat open source framework: The voice AI pipeline solution recommended by the HN community is maturing rapidly.
3. Adopt Transceiver architecture: Migrate from one-port-per-session to transceiver model, using fewer ports to carry more concurrent sessions.
4. Advantages of the Go language: Go's goroutine model is naturally suitable for processing tens of thousands of concurrent connections in real-time systems.

If you are a ChatGPT voice user

• Behind the low-latency experience of voice mode is an extremely complex infrastructure. Understanding this will help you use the product appropriately.
• If you feel that the voice reply is too fast and interrupts your thinking, you can give feedback to OpenAI through official channels

Related extended information

• HN ：222
• Pion WebRTC - Go WebRTC
• pipecat-ai/pipecat - AI

Tool entry

The following technology stacks appearing in the text are widely used in the field of AI speech: OpenAI, ChatGPT, Kubernetes, DeepSeek, Hermes Agent. These tools are being used by an increasing number of developers of automated workflows and AI applications.

Internal link guidance

• Want to build your own AI automated workflow? Watch: How to build an AI content automated distribution system with n8n + ChatGPT: a complete 30-minute tutorial
• Real case: Someone used AI tools to create a product with a monthly income of $5,000: Indie Developer: n8n + OpenClaw Automation Workflow Earning $5,000/mo
• More ways to use AI Agent tools: AI Agent Tools 2026 Complete Tutorial: 5 Tools to Build an Automated Pipeline in 30 Minutes