Nvidia opened the Multipath Reliable Connection (MRC) transport protocol to the broader industry on May 6, 2026, releasing it as an open specification through the Open Compute Project after first deploying it on Spectrum-X Ethernet hardware inside frontier AI training clusters at OpenAI, Microsoft and Oracle. MRC is an RDMA transport that lets a single connection spread traffic across multiple network paths, with hardware-level failure bypass that detects a broken path and reroutes around it in microseconds. The move turns a Spectrum-X-native protocol into a shared standard for the gigascale AI fabric.
The deployments named are not test beds. Microsoft's Fairwater and Oracle Cloud Infrastructure's Abilene data center — two of the largest AI factories built for training frontier large language models — already rely on MRC to hit their bandwidth and uptime targets. OpenAI runs MRC in production on its Blackwell-generation systems. The protocol was co-developed with AMD, Broadcom, Intel, Microsoft and OpenAI, a list that reads like a deliberate signal that Nvidia wants Ethernet, not just InfiniBand, to anchor the next wave of training networks.
Sachin Katti, head of industrial compute at OpenAI, framed the result in operational terms. "Deploying MRC in the Blackwell generation was very successful and was made possible by a strong collaboration with NVIDIA," Katti said. "MRC's end-to-end approach enabled us to avoid much of the typical network-related slowdowns and interruptions and maintain the efficiency of frontier training runs at scale."
Key facts
- 01Nvidia released Multipath Reliable Connection (MRC) as an open specification through the Open Compute Project on May 6, 2026.
- 02MRC is already running in production at OpenAI, Microsoft Fairwater and Oracle Cloud Infrastructure's Abilene data center.
- 03Spectrum-X failure bypass detects a path failure and reroutes traffic in hardware in microseconds.
- 04Nvidia developed MRC in collaboration with AMD, Broadcom, Intel, Microsoft and OpenAI.
- 05Customers can run Spectrum-X Ethernet Adaptive RDMA, MRC or custom protocols on ConnectX SuperNICs and Spectrum-X switches.
The technical pitch is straightforward. A standard RDMA connection pins traffic to one path, which works fine until that path congests or breaks, at which point thousands of synchronized GPUs stall waiting on a single straggler. MRC distributes a single connection's traffic across every available path, balances load on the fly, and on packet loss retransmits precisely rather than restarting flows. Nvidia's analogy is replacing a single-lane road through town with a grid plus a live traffic app.
“MRC's end-to-end approach enabled us to avoid much of the typical network-related slowdowns and interruptions and maintain the efficiency of frontier training runs at scale.”— Jaeden Schafer
Failure bypass is the part that matters most for training economics. In a multi-thousand-GPU job, a brief network disruption can idle the entire cluster and force a checkpoint reload. Spectrum-X handles detection and reroute in hardware in microseconds, fast enough that long-running jobs don't notice. For an operator paying for tens of thousands of Blackwell GPUs by the hour, that is the difference between a clean training run and a re-spent week.
OpenAI is also pairing MRC with multiplanar network designs on Spectrum-X. A multiplane fabric runs several independent network planes in parallel, each offering an alternate path between GPUs, and Nvidia's Spectrum-X Multiplane capability adds hardware-accelerated load balancing across those planes. The combination is how the company is keeping latency predictable while scaling to hundreds of thousands of GPUs in a single cluster.
Spectrum-X is also being kept deliberately pluralist on transport. Customers can run Spectrum-X Ethernet Adaptive RDMA, MRC or other custom protocols natively across Nvidia ConnectX SuperNICs and Spectrum-X switches, all with multiplane support. That flexibility is the wedge against InfiniBand-only deployments and against rival Ethernet AI fabrics from Broadcom and others — pick your transport, keep the hardware.
The OCP release is the strategic piece. Nvidia could have kept MRC as a Spectrum-X exclusive and used it as a lock-in feature. Instead, by handing the spec to OCP with AMD, Broadcom and Intel as co-developers, Nvidia is betting that broad adoption of an MRC-shaped transport benefits Spectrum-X more than exclusivity would — because Spectrum-X is the platform where MRC was tuned first and runs best. This follows AI Chat Daily's earlier coverage of the Spectrum-X open-spec move on the protocol's announcement.
The caveat worth naming is that an open specification is not a guarantee of interoperable implementations. AMD, Broadcom and Intel collaborated on the spec, but each will ship its own NICs and switches on its own timeline, and protocol-level compatibility does not equal performance parity. Nvidia's pitch — that MRC was "proven first and optimized on Spectrum-X hardware" — is also a hint that early non-Nvidia implementations may underperform the reference. Buyers running Fairwater- or Abilene-scale fabrics will keep weighing Spectrum-X against vendor-neutral Ethernet roadmaps from the Ultra Ethernet Consortium camp.
For Nvidia, the playbook here is the one that has worked all decade: ship the silicon, prove the protocol in production at the most demanding customers, then standardize it once the lead is locked in. With OpenAI, Microsoft Fairwater and Oracle's OCI Abilene already running MRC at frontier scale, Spectrum-X arrives at the open-standards table as the incumbent — and the rest of the AI networking market now has to decide whether to build to Nvidia's spec or watch its largest customers keep buying Nvidia's switches anyway.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
