NextFin News - OpenAI has unveiled a new networking protocol designed to eliminate the data bottlenecks that frequently stall the training of massive artificial intelligence models. Developed over two years in collaboration with a coalition of industry heavyweights including AMD, Broadcom, Intel, Microsoft, and NVIDIA, the Multipath Reliable Connection (MRC) protocol aims to solve the "all-or-nothing" fragility of current supercomputer clusters. By releasing the specification through the Open Compute Project (OCP), the group is attempting to standardize how thousands of GPUs communicate, a move that could reduce the reliance on proprietary networking technologies that have historically locked customers into specific hardware ecosystems.
The technical challenge MRC addresses is rooted in the sheer scale of modern AI training. When training a frontier model, data must be synchronized across tens of thousands of GPUs simultaneously. Under existing standards, a single failed network link or a congested switch can cause the entire training job to crash, wasting millions of dollars in compute time. MRC introduces "adaptive packet spraying," which allows a single data transfer to be split across hundreds of different paths through the network. If one path fails, the protocol reroutes data in microseconds, preventing the cascading failures that have plagued the industry’s largest clusters.
OpenAI confirmed that MRC is already operational within its most advanced infrastructure, including the NVIDIA GB200-powered supercomputers hosted by Oracle Cloud Infrastructure in Texas and Microsoft’s Fairwater systems. The protocol is specifically optimized for the latest 800Gb/s network interfaces, suggesting that the next generation of AI hardware will be built with this multipath capability as a core requirement. By integrating with SRv6 (Segment Routing over IPv6), the protocol also simplifies the network control plane, allowing operators to bypass hardware failures using static source routing rather than relying on complex, slow-to-converge dynamic routing protocols.
While the collaboration includes NVIDIA, the dominant force in AI networking via its proprietary InfiniBand technology, the push for an open standard like MRC signals a strategic shift. Broadcom and AMD have long advocated for Ethernet-based alternatives to InfiniBand, arguing that open standards foster a more competitive and scalable supply chain. However, some analysts remain cautious about the speed of adoption. Patrick Moorhead, Chief Analyst at Moor Insights & Strategy, has frequently noted that while open standards are essential for long-term industry health, NVIDIA’s vertical integration currently provides a performance "premium" that many top-tier labs are unwilling to sacrifice for the sake of interoperability.
The release of MRC through the OCP effectively places it in competition—or perhaps coordination—with the Ultra Ethernet Consortium (UEC), another industry body working to modernize Ethernet for the AI era. The distinction lies in immediate utility; while UEC is building a comprehensive new stack from the ground up, MRC is a targeted protocol already proven in production environments. This suggests a bifurcated market where the largest "hyperscalers" may deploy specialized protocols like MRC to solve immediate scaling pains while the broader enterprise market waits for the finalized UEC 1.0 specifications to arrive in hardware later this year.
The economic stakes of this networking overhaul are significant. As AI models grow toward the trillion-parameter threshold, the cost of networking components is representing an increasingly large slice of total capital expenditure. If MRC succeeds in making networks more resilient with fewer redundant components, it could lower the barrier to entry for second-tier cloud providers and sovereign AI initiatives. For now, the protocol serves as a critical patch for the industry's most pressing problem: ensuring that the world's most expensive computers do not sit idle because of a single broken wire.
Explore more exclusive insights at nextfin.ai.
