On May 5, 2026, OpenAI announced a major infrastructure breakthrough: the MRC (Multipath Reliable Connection) network protocol, co-developed with AMD, Broadcom, Intel, Microsoft, and NVIDIA. The specification has been open-sourced through the Open Compute Project (OCP) and is already running on the largest NVIDIA GB200 supercomputers, training next-generation frontier AI models.
The Old Bottleneck: One Packet Drop Holds an Entire Cluster Hostage
Training massive AI models requires tens of thousands of GPUs to exchange data in perfect sync. Under traditional networking, a single delayed packet or a switch failure can stall the entire training step, idling expensive GPUs and forcing costly restarts or long routing convergence. As cluster sizes balloon, the problem grows exponentially.
OpenAI decided to redesign the network layer from scratch, aiming for fault tolerance at the 100,000+ GPU scale.
Three Core Innovations: Multi-Plane, Adaptive Spraying, Static Routing
MRC achieves ultra-low latency and near-zero downtime through three radical architectural changes:
- Multi-Plane Topology: The 800 Gb/s network interface is split into smaller links (e.g., 8 x 100 Gb/s), each connecting to different switches to form parallel planes. The system needs only 2 switch layers to link over 100,000 GPUs, compared to 3–4 layers in legacy designs. This cuts cost, power, and component count significantly.
- Adaptive Packet Spraying: Instead of pinning a flow to a single path, MRC scatters packets across hundreds of paths. Load balancing is dynamic — if a switch is congested, it triggers "Packet Trimming," forwarding only headers to trigger fast retransmission and reduce false positives.
- Static Source Routing (SRv6): BGP dynamic routing is eliminated. The sender embeds the complete path in each packet, and switches follow a simple static table. When a link fails, MRC reroutes around it at microsecond speed, so training jobs experience virtually no interruption.
Live on the World's Largest GB200 Supercomputers, Training Next-Gen Models
MRC isn't a theoretical exercise. OpenAI confirmed the protocol is fully deployed on its biggest NVIDIA GB200 supercomputers — including the site with Oracle Cloud in Abilene, Texas, and Microsoft's Fairwater system — and is actively training multiple next-generation frontier models. OpenAI states: "In production, even with multiple link flaps per minute or planned first-tier switch restarts, training workloads see almost zero impact. No special maintenance coordination is needed."