OpenAI’s MRC Protocol Boosts AI Supercomputer Networking

OpenAI dropped a new networking protocol called MRC (Multipath Reliable Connection) that’s designed to make massive AI training clusters way more efficient. Teaming up with AMD, Broadcom, Intel, Microsoft, and NVIDIA—and released through the Open Compute Project—MRC tackles the biggest headaches in supercomputer networking: congestion and constant failures. When you’re training models like GPT, a single delayed data transfer can idle thousands of GPUs, wasting time and power. MRC uses adaptive packet spraying to kill congestion, multi-plane designs for redundancy with fewer components, and static source routing to bypass failures without crashing the whole job. This is a huge deal for Stargate-scale clusters, where even the best hardware has a steady hum of link and switch failures. By sharing the spec openly, OpenAI wants to standardize networking across the AI ecosystem, making it easier to scale up reliably. Why it matters: This isn’t just a tweak—it’s a foundational rethink of how AI supercomputers communicate, and open-sourcing it could accelerate the entire industry’s ability to train next-gen models faster and cheaper.