OpenAI built a networking protocol with AMD, Broadcom, Intel, Microsoft, and NVIDIA to fix AI supercomputer bottlenecks

OpenAI and partners including AMD, Broadcom, Intel, Microsoft, and NVIDIA launched MRC, a new multipath networking protocol for AI supercomputers.

MRC spreads GPU traffic across multiple routes, recovers from failures in microseconds, and reduces the number of switch tiers needed in very large clusters.

The spec and research paper were released via the Open Compute Project, and MRC is already running on OpenAI’s largest NVIDIA GB200 systems.

The goal is to make frontier model training faster, more reliable, and cheaper by cutting congestion and downtime in giant GPU networks.