OpenAI · 2026-05-06 · major
MRC — OpenAI, NVIDIA, AMD, Broadcom, Intel, Microsoft Open-Source AI Supercomputer Networking Protocol
Six-vendor consortium releases Multipath Reliable Connection — an RDMA transport that spreads traffic across hundreds of paths, reroutes around link failures in microseconds, and lets 100K-GPU clusters run on two Ethernet tiers instead of three or four.
OpenAI gets the entire AI hardware stack — NVIDIA, AMD, Broadcom, Intel, and Microsoft — to agree on one open networking protocol for 100K-GPU clusters.
Key specs
| Gpus per cluster | 100,000+ |
|---|---|
| Failure recovery | microseconds |
| Link speed | 800 Gb/s |
| Ethernet tiers | 2 (vs 3-4 conventional) |
| Vendors | 6 (OpenAI, NVIDIA, AMD, Broadcom, Intel, Microsoft) |
| Production sites | OpenAI/Oracle Abilene, Microsoft Fairwater |
What is it?
Multipath Reliable Connection is a new RDMA transport protocol for AI training clusters, contributed to the Open Compute Project so any vendor can implement it. It is built into the latest 800 Gb/s network interfaces and extends RDMA over Converged Ethernet with source routing techniques from the Ultra Ethernet Consortium. The protocol is co-developed by OpenAI, NVIDIA, AMD, Broadcom, Intel, and Microsoft.
How does it work?
MRC spreads packets from a single connection across hundreds of network paths simultaneously, so no one path bottlenecks the all-reduce traffic that dominates training. When a link, switch, or path fails, MRC detects it and reroutes in microseconds — conventional fabrics take seconds. The path-spreading also collapses fabric topology: a 100,000-GPU cluster can run on two tiers of Ethernet switches instead of the three or four needed by today's 800 Gb/s networks.
Why does it matter?
Network failures and tail latency are now the dominant cause of wasted GPU time in frontier-scale training runs. MRC is already deployed in OpenAI's largest GB200 supercomputers, Oracle Cloud's Abilene Texas site, and Microsoft's Fairwater clusters — and was used to train GPT-5.5. By open-sourcing it through OCP, the consortium is signalling that AI fabric is a shared standard layer, not a vendor moat.
Who is it for?
AI infrastructure teams, hyperscaler network architects, OCP contributors
Try it
https://www.opencompute.org/