OpenAI’s MRC Initiative Reinforces Ethernet’s Expanding Role in AI Back-end Networks

OpenAI and industry partners introduced Multipath Reliable Connection, a networking protocol aimed at limiting congestion and failures in large AI clusters, supporting more predictable training across very large XPU systems.

Market Overview

The note frames MRC as a response to the way network slowdowns or failures can ripple through synchronous AI training jobs spanning thousands of GPUs.

It also ties the release to an ongoing shift toward Ethernet for AI cluster interconnects rather than relying solely on InfiniBand-style fabrics.

Key Findings

MRC extends RoCE, or RDMA over Converged Ethernet, and adds mechanisms intended to improve resilience and predictability for synchronous AI training workloads.

The article describes these mechanisms as multipath packet spraying and SRv6 source routing.

Technology and Standards

The note says the protocol supports Ethernet-based AI networking while positioning it as an addition to existing approaches rather than a replacement.

It also reports that OpenAI released the MRC specification through the Open Compute Project to enable alignment across silicon, cloud, systems, and AI infrastructure builders.

Operational Deployment

The report states that OpenAI says MRC is deployed across its largest NVIDIA GB200 supercomputers, including a site with Oracle Cloud Infrastructure in Abilene, Texas.

It adds that Microsoft’s Fairwater supercomputers also use MRC, and both examples were deployed using NVIDIA’s SpectrumX switches.

Industry Context

The note cites Dell’Oro Group data indicating that NVIDIA and Celestica captured 50% of AI back-end networks in 2025, with Arista ranking third.

It characterizes the broader implication as a shift in enterprise and cloud planning from compute-only comparisons to the ability to connect systems efficiently and operate networks reliably at scale.

This Analyst Signals brief reflects a neutral, fact-based summary of the original research note.