Nvidia is opening Multipath Reliable Connection, the RDMA transport protocol behind its Spectrum-X Ethernet fabric, to the rest of the industry through the Open Compute Project. The company announced the release on May 6, 2026, after MRC ran in production on Blackwell-generation clusters at OpenAI, Microsoft and Oracle. Failure bypass detection and automatic rerouting now happen in microseconds, in hardware, across fabrics scaling to hundreds of thousands of GPUs.
The protocol was co-developed with AMD, Broadcom, Intel, Microsoft and OpenAI — a notable lineup given how rarely those names ship a joint spec. MRC lets a single RDMA connection spread traffic across multiple network paths simultaneously, which Nvidia frames as the difference between a single-lane road and a managed grid with live traffic rerouting. The practical payoff: higher GPU utilization, better load balancing, and fewer stalls when a link degrades mid-training.
Microsoft's Fairwater data center and Oracle Cloud Infrastructure's Abilene site, two of the largest AI factories built specifically for frontier LLM training, are already running MRC over Spectrum-X. Both are part of the wave of gigascale builds that have pushed networking, not just compute, to the center of AI infrastructure planning. Nvidia's pitch is that scale-out Ethernet has become the bottleneck — and that Spectrum-X is the answer.
Key facts
- 01Nvidia released Multipath Reliable Connection (MRC), an RDMA transport protocol, as an open specification through the Open Compute Project on May 6, 2026.
- 02MRC was co-developed with AMD, Broadcom, Intel, Microsoft and OpenAI, and first deployed on Nvidia's Blackwell generation.
- 03The protocol's failure bypass detects path failures and reroutes traffic in microseconds, scaling to hundreds of thousands of GPUs.
- 04Microsoft's Fairwater and Oracle Cloud Infrastructure's Abilene data center are running MRC over Spectrum-X Ethernet in production.
- 05OpenAI's Sachin Katti said MRC let frontier training runs avoid typical network-related slowdowns at scale.
Sachin Katti, head of industrial compute at OpenAI, said the rollout worked. "Deploying MRC in the Blackwell generation was very successful and was made possible by a strong collaboration with NVIDIA," he 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." Quotes like that are the entire point of opening the spec — they double as a customer reference and a recruiting pitch to the next operator considering a non-Nvidia fabric.
“MRC's failure bypass technology detects a network path failure and reroutes traffic automatically in hardware in just microseconds, keeping hundreds of thousands of GPUs synchronized during frontier training runs.”— Jaeden Schafer
The technical case rests on what happens when a network path dies during a training job. With thousands of GPUs needing to stay in lockstep, even a brief interruption can stall an entire run and burn idle compute. MRC's hardware-level failure bypass detects the broken path in microseconds and reroutes around it without involving the host stack, which Nvidia argues is the only way to keep latencies predictable at this scale.
Layered on top is multiplane networking — multiple independent fabrics, each offering an alternate communication path between GPUs. OpenAI is deploying multiplane designs with Spectrum-X alongside MRC, with hardware-accelerated load balancing distributing traffic across planes. Nvidia's Spectrum-X Multiplane capability is what makes that scale to the hundreds-of-thousands-of-GPUs target without sacrificing latency.
Customers running Spectrum-X aren't locked into MRC. Spectrum-X Ethernet Adaptive RDMA, MRC, and custom protocols all run natively across Nvidia's ConnectX SuperNICs and Spectrum-X switches. That flexibility matters because hyperscalers tend to want their own transport experiments alongside vendor defaults, and Nvidia is signaling it would rather host those experiments than fight them.
Gilad Shainer, who authored Nvidia's announcement, framed the release in the language of open standards. "As AI factories continue to scale, the network must do more than move data quickly. It must be intelligent, resilient and based on open standards." Submitting MRC to OCP is the concrete move behind that line — though the protocol was, as Nvidia notes, "proven first in production with performance optimized on NVIDIA Spectrum-X Ethernet hardware."
The skeptic's read is straightforward. Open-sourcing a transport protocol that was co-designed for and first optimized on your own silicon is a competitive maneuver as much as an act of generosity. AMD, Broadcom and Intel are listed as collaborators, but the gravitational pull stays with Spectrum-X switches and ConnectX SuperNICs, where MRC was tuned. Rivals selling Ethernet switching into AI clusters now have to either adopt MRC and chase Nvidia's implementation curve, or argue their alternative is worth the integration cost.
That dynamic mirrors what Nvidia has done across the stack — release reference designs and software layers that anyone can adopt, while keeping the highest-performance implementation on Nvidia hardware. Following last week's coverage of Nvidia's NemoClaw agent framework wrapping the OpenClaw spec, MRC is the second open-by-Nvidia release in a fortnight aimed at locking in the company's role as the default substrate for gigascale AI. The Pentagon's recent classified-AI awards to Nvidia, Microsoft and OpenAI sit in the same orbit.
For operators planning the next round of AI factory buildouts, the calculus shifts. Ethernet has now eaten enough of the InfiniBand use case that the question is no longer whether to run AI on Ethernet, but whose Ethernet stack reroutes fastest when a fiber goes dark. Nvidia's answer — measured in microseconds, validated at OpenAI, Microsoft and Oracle, and now openly specified — is the one competitors will spend the next year trying to match.
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.
