OpenAI Unveils MRC Networking Protocol with AMD, Broadcom, Intel, Microsoft, Nvidia

What happened

OpenAI announced the release of Multipath Reliable Connection (MRC), a new networking protocol developed in collaboration with AMD, Broadcom, Intel, Microsoft, and Nvidia over the past two years. MRC was released through the Open Compute Project (OCP) as an open specification. The protocol addresses network congestion that occurs during large-scale AI model training by distributing a single data transfer across hundreds of network paths simultaneously, rather than forcing traffic through a few congested routes. MRC is built into the latest 800 Gb/s network interfaces and extends the existing RDMA over Converged Ethernet (RoCE) standard. It enables hardware-accelerated remote direct memory access for GPUs and CPUs, allowing data to route around network failures in microseconds. The protocol has already been deployed in production environments, including OpenAI's supercomputers using Nvidia GB200 GPUs at Oracle Cloud Infrastructure in Abilene, Texas, and Microsoft's Fairwater supercomputers. These systems are used to train frontier models powering ChatGPT and Codex. MRC also underpins OpenAI's planned Stargate supercomputer, which aims to deploy 10GW of AI compute by 2029.

Why it matters

For Intel, this collaboration positions the company as a key partner in developing critical infrastructure for next-generation AI workloads, even as it competes with Nvidia and AMD in the GPU and accelerator markets. Intel's participation in the MRC consortium demonstrates its ongoing relevance in AI networking and data center technologies, particularly through its networking hardware and chipsets that support large-scale AI training environments. As AI model sizes and training clusters continue to scale into the hundreds of thousands of GPUs, networking bottlenecks have become a major constraint on performance and efficiency. Network delays can cause expensive GPU resources to sit idle, disrupting multimillion-dollar training runs. MRC addresses this critical infrastructure challenge by improving network resilience, reducing congestion, and enabling more efficient use of GPU compute resources. The open release of MRC through OCP also signals a broader industry shift toward collaborative, open standards in AI infrastructure, rather than proprietary, closed solutions. This could influence how future AI data center architectures are designed and how Intel positions its own networking and infrastructure offerings.

Bigger picture

The MRC announcement reflects the growing complexity and scale of AI infrastructure as companies race to train ever-larger models. Traditional networking approaches struggle to keep pace with the unique demands of AI workloads, which require tight synchronization and massive data throughput across tens or hundreds of thousands of GPUs. Nvidia's Spectrum-X Ethernet platform, which supports MRC, is part of a broader trend toward AI-native networking, where the network is treated as an integral part of the compute pipeline rather than generic plumbing. The collaboration among typically competing firms—AMD, Broadcom, Intel, Microsoft, and Nvidia—underscores the shared recognition that network congestion is a universal bottleneck that no single vendor can solve alone. At the same time, MRC is one of several emerging Ethernet variants for AI, including the Ultra Ethernet Consortium's efforts, suggesting a future with multiple coexisting protocols tailored to different hyperscaler needs rather than a single dominant standard. This pluralistic landscape gives infrastructure providers and model developers more flexibility but also increases complexity in evaluating and deploying AI networking solutions.

What to watch

Investors should monitor how broadly MRC is adopted beyond OpenAI, Microsoft, and Oracle, and whether other hyperscalers and AI infrastructure providers integrate the protocol into their own training environments. Watch for Intel's specific contributions and product announcements related to MRC-compatible networking hardware, as well as competitive responses from other networking and silicon vendors. The rollout of OpenAI's Stargate supercomputer and its reliance on MRC will be a key test case for the protocol's scalability and performance at extreme cluster sizes. Additionally, developments within the Ultra Ethernet Consortium and how it relates to or competes with MRC will shape the future landscape of AI networking standards. Finally, track how MRC's open-source nature influences cross-vendor collaboration and whether it becomes a de facto standard or remains one option among many in a fragmented market.

This article was generated by Quantli AI using publicly available news sources.