OpenAI Introduces MRC (Multipath Reliable Connection): A New Open Networking Protocol for Large-Scale AI Supercomputer Training Clusters

Why It Matters

MRC turns networking from a hidden bottleneck into a predictable, resilient layer, directly lowering GPU idle time and training expenses for large‑scale AI models. Its industry‑open design could reshape data‑center architectures across the AI ecosystem.

Key Takeaways

• MRC spreads packets across hundreds of paths, slashing congestion
• Microsecond‑scale failure rerouting keeps training jobs alive
• Two‑tier topology supports 131k+ GPUs with fewer switches
• Open‑source via OCP, enabling broader industry adoption

Pulse Analysis

Training today’s frontier models demands not only raw compute but also a network that can move petabytes of data without pause. Traditional RoCE fabrics route each transfer over a single path, creating hot spots that force GPUs to wait for delayed packets. OpenAI’s MRC tackles this by spraying packets across dozens of parallel routes and embedding SRv6 source routes directly in the packet header, offloading routing decisions to NICs. The result is higher bandwidth utilization, dramatically reduced tail latency, and a network that behaves predictably even when links or switches falter.

The protocol’s engineering choices are as pragmatic as they are innovative. By delegating routing intelligence to the NIC, MRC eliminates the need for dynamic routing tables in switches, allowing instantaneous microsecond‑level failover. Its multi‑plane design fragments an 800 Gb/s interface into eight 100 Gb/s lanes, enabling a two‑tier switch fabric to connect over 130,000 GPUs—a configuration that would otherwise require three or four tiers. This reduction translates into roughly one‑third fewer optics and 40% fewer switches, cutting capital expenditure and power consumption while also shrinking the blast radius of any single failure.

For enterprises and cloud providers, MRC offers a clear economic incentive. Every second a GPU sits idle costs thousands of dollars in wasted compute; by keeping traffic flowing even during component failures, MRC safeguards both model training timelines and budgets. Moreover, its open publication through the Open Compute Project invites other AI leaders to adopt or extend the protocol, potentially standardizing a more resilient networking layer across the industry. As AI models grow larger and training cycles lengthen, such networking resilience will become a decisive factor in competitive advantage.