OpenAI Releases AI Networking Protocol as Open Standard

A New Backbone for AI Training

OpenAI on Wednesday released the Multipath Reliable Connection (MRC) protocol, a new networking specification designed to make massive AI training clusters faster, more efficient, and more resilient. The protocol has been published as an open standard through the Open Compute Project, and was developed in collaboration with an unusual coalition of industry rivals including Microsoft, AMD, Broadcom, Nvidia, and Intel.

The move addresses two of the biggest bottlenecks in modern AI infrastructure: network congestion across thousands of GPUs and the inevitable hardware failures that can derail multi-week training runs.

How MRC Works

At the heart of MRC is a technique called packet spraying, which scatters data across hundreds of network paths simultaneously rather than funnelling it through a few high-traffic links. According to Mark Handley, OpenAI's networking lead, this produces flatter networks that consume less power and compute resources. When a path fails, MRC detects the issue and reroutes traffic within microseconds, allowing training jobs to continue without interruption.

The protocol also pairs with SRv6, or IPv6 Segment Routing, which prescribes exact data paths through the network instead of requiring individual switches to make routing decisions on the fly. The result is lower energy consumption at the switch level and more predictable performance.

Already Deployed at Scale

MRC is not a research preview. It is already running in OpenAI and Microsoft's largest training facilities, including the Oracle site in Abilene, Texas, and Microsoft's Fairwater supercomputers. These are the same systems that have been used to train GPT-5.5 and other frontier models.

Greg Steinbrecher, OpenAI's workload lead, framed the release as an industry-wide push rather than a competitive play. He stressed that fragmentation across in-house protocols has been bad for the networking industry and that aligning everyone behind a single open standard accelerates progress for all players.

Why It Matters

By establishing MRC as an open standard, the coalition is signalling a clear shift toward Ethernet-based AI fabrics and away from proprietary networking stacks. With compute demand surging and data centre build-outs accelerating, a shared protocol could reshape how the next generation of AI infrastructure is designed and operated.