NVIDIA Spectrum-X Ethernet with MRC Redefines AI Networking at Gigascale

NVIDIA, OpenAI, and Microsoft Unveil MRC Protocol to Supercharge AI Training

NVIDIA has announced that its Spectrum-X Ethernet fabric, now featuring the Multipath Reliable Connection (MRC) transport protocol, sets a new standard for gigascale AI networking. The technology is already deployed by industry leaders including OpenAI, Microsoft, and Oracle.

NVIDIA Spectrum-X Ethernet with MRC Redefines AI Networking at Gigascale — Source: blogs.nvidia.com

MRC enables a single RDMA connection to distribute traffic across multiple network paths simultaneously, dramatically improving throughput, load balancing, and availability for large-scale AI training fabrics. This replaces the traditional single-path approach with a dynamic, multi-lane system.

Background

As AI models grow exponentially, conventional Ethernet networks struggle to keep GPU clusters fully utilized. Network congestion and data loss cause costly idle time during long training runs. NVIDIA’s Spectrum-X platform was built specifically to address these bottlenecks.

Spectrum-X combines purpose-built hardware, deep telemetry, and intelligent fabric control. MRC represents the latest evolution—an open specification contributed to the Open Compute Project, born from close collaboration between NVIDIA, Microsoft, and OpenAI.

‘A New Benchmark for Reliability’ — OpenAI Executive

“Deploying MRC in the Blackwell generation was very successful and was made possible by a strong collaboration with NVIDIA,” said Sachin Katti, head of industrial compute at OpenAI. “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.”

Katti’s statement underscores the real-world impact: MRC allowed OpenAI to sustain high GPU utilization even under extreme network loads, a critical factor for training models like GPT.

What This Means

The introduction of MRC turns the network from a potential bottleneck into an accelerator. By load-balancing traffic across all available paths and dynamically rerouting around congestion, MRC ensures every GPU receives the bandwidth it needs throughout a training run.

When data loss occurs—inevitable at petabit scales—intelligent retransmission enables rapid, precise recovery with minimal impact on long-running jobs. Administrators gain fine-grained visibility and control over traffic paths, simplifying operations and accelerating troubleshooting.

Global Deployments Already Live

Microsoft’s Fairwater data center and Oracle Cloud Infrastructure’s Abilene facility—two of the largest AI factories built for frontier LLMs—already rely on MRC. These deployments prove the technology works at gigascale, not just in labs.

NVIDIA Spectrum-X Ethernet now provides the network foundation these hyperscalers need to run large-scale AI models with confidence. The platform’s purpose-built hardware, combined with MRC’s open specification, signals a shift toward truly AI-native networking.

Urgent Implications for AI Infrastructure

As demand for AI compute skyrockets, any downtime or underutilization of expensive GPUs wastes millions of dollars. MRC directly addresses this by keeping GPU clusters fed with data—no waiting for network recovery, no stalled training runs.

“Replacing a single-lane road with a smart street grid and a real-time traffic app” is how NVIDIA describes the leap. That analogy captures both the elegance and the urgency: every millisecond counts in the race to build smarter, faster AI.

The open-source release of MRC through the Open Compute Project ensures the entire industry can benefit. Expect rapid adoption as other cloud providers and enterprise AI labs seek to eliminate network chokepoints.