MRC Goes Open — Ethernet AI Fabric Architecture Reaches Gigascale Production
OpenAI and Microsoft reveal MRC, the open RDMA transport that eliminates flow collisions in Ethernet AI fabrics at one hundred thousand GPU scale — confirmed in production at the clusters training today's frontier models. Plus Anthropic ships agentic engineering tools, SONiC lands SRv6 and DPU dark-mode, and a benchmark shows vision agents burn forty-five times more tokens than API calls.
Welcome to Amaze Networks for Thursday, May seventh. OpenAI and Microsoft just went on the record about the transport protocol running inside their largest AI training clusters — and it's running on Ethernet.
Not InfiniBand. A new protocol called Multipath Reliable Connection — M-R-C for short — now contributed to the Open Compute Project as an open spec. And there's a full architecture paper from both organizations explaining exactly how it works in production.
Let's start there because I think this is one of the most significant AI fabric stories we've covered. What's the problem MRC actually solves?
New episodes, every weekday.
Amaze Networks drops at 4 AM CT, Monday through Friday. Spotify and Apple Podcasts submissions in progress.