NVIDIA: Spectrum-X Multipath Reliable Connection becomes OCP open standard for gigascale AI networks

On May 6, 2026, NVIDIA announced that Multipath Reliable Connection (MRC), the RDMA transport protocol developed for the Spectrum-X Ethernet platform, has been published as an open specification through the Open Compute Project (OCP). The protocol was previously proven in production on Spectrum-X hardware across three major AI supercomputers.

What is Multipath Reliable Connection?

MRC is a transport protocol that enables a single RDMA connection (Remote Direct Memory Access — accessing a remote machine’s memory without CPU involvement) to distribute traffic across multiple network paths simultaneously. NVIDIA uses an analogy in the announcement: instead of a single-lane road, MRC lays out “a network of streets with a real-time traffic app” that reroutes drivers around congestion and closed roads.

Technically, the protocol delivers hardware-accelerated fault detection and rerouting in microseconds, intelligent retransmission for fast packet-loss recovery, and fine-grained traffic visibility and control for administrators.

Who is already using MRC in production?

Three major AI deployments already rely on MRC:

• OpenAI — uses multiplanar network designs with MRC in combination with NVIDIA’s Blackwell GPU generation. Sachin Katti from OpenAI stated that “MRC’s end-to-end approach enabled avoiding typical network slowdowns and interruptions.”
• Microsoft — the Fairwater data center relies on MRC for performance and energy efficiency.
• Oracle Cloud Infrastructure — the Abilene data center uses MRC for large-scale frontier LLM deployment.

All three cases fall into the category of gigascale AI networks — clusters where GPU counts run into the tens of thousands and traditional RDMA designs begin choking due to hot spots and slow fault recovery.

Why does the OCP standard change the game?

The Open Compute Project is an industry forum founded in 2011 for open data center equipment specifications. When NVIDIA donates a protocol to OCP, other networking equipment vendors can implement MRC in their own ASICs and switches — without licenses, without paying royalties to NVIDIA.

MRC development was already collaborative: NVIDIA mentions AMD, Broadcom, Intel, Microsoft and OpenAI as contributors to the specification. This dramatically reduces vendor lock-in risk for hyperscalers that have already invested in Spectrum-X switches.

What does this mean for AI infrastructure?

Gigascale AI networks have until now been the territory of exclusive NVIDIA-Mellanox designs. Opening the MRC specification signals that NVIDIA is willing to share the standard so the entire industry can scale faster — the assumption being that GPU sales, not switching IP, will remain the primary revenue source.

For organizations building their own AI clusters, MRC as an open standard means a potentially wider choice of switch vendors while maintaining the same level of performance that OpenAI, Microsoft and Oracle already have in production.