Why AI needs a new kind of supercomputer network

The OpenAI Podcast · 6 May 2026 · Episode 18

About this episode

Training frontier models is not just a matter of adding GPUs; one small fault can break the whole coordinated dance. Mark Handley and Greg Steinbrecher explain the new network design behind OpenAI's latest training runs and Multipath Reliable Connection, a protocol developed with AMD, Broadcom, Intel, Microsoft and Nvidia and opened up to the industry.

About the show

The OpenAI Podcast — OpenAI. Long-form conversations with the people building at OpenAI — research, ChatGPT, Codex, infrastructure and where the industry is heading.