Unlocking large scale AI training networks with MRC (Multipath Reliable Connection)

OpenAI has released a networking protocol called Multipath Reliable Connection, or MRC, through the Open Compute Project, the industry group that already shapes how hyperscalers build servers and racks. The pitch is straightforward: as training runs stretch across tens of thousands of GPUs, the network connecting them becomes just as important as the chips themselves, and standard connection protocols start to buckle under that scale.

The core problem MRC tackles is resilience. In a cluster with thousands of network paths, something is always failing — a cable, a switch port, a link flapping in and out. Traditional reliable connection protocols route traffic in a way that ties performance to the health of individual paths, so one bad link can drag down an entire training job, or worse, stall a checkpoint that took hours to reach. MRC instead spreads traffic across multiple paths simultaneously, so a single degraded route doesn't become a single point of failure for the whole system.

This isn't just an academic exercise for OpenAI. The company runs training clusters at a scale where even small inefficiencies compound fast — a network hiccup that costs a few minutes across 50,000 GPUs adds up to a lot of wasted compute-hours, and by extension, a lot of wasted electricity and money. By publishing MRC through OCP rather than keeping it in-house, OpenAI is betting that better networking infrastructure benefits everyone racing to build bigger models, not just itself.

There's also a broader signal here about where the bottlenecks in AI actually sit. For years the conversation around scaling has centered on GPU counts and chip architecture. MRC is a reminder that the plumbing between those chips — the switches, the cables, the protocols moving gradients and activations back and forth — is quietly becoming its own competitive frontier. OCP's involvement means MRC could show up in gear from multiple vendors, not just whatever OpenAI happens to be running internally.