OpenAI Releases New Network Protocol for AI Training Clusters
SAN FRANCISCO — OpenAI on Wednesday introduced a new supercomputer networking protocol called Multipath Reliable Connection, or MRC, designed to improve the resilience and performance of large-scale AI training clusters, the company announced on its blog.
The protocol, which OpenAI has released as an open standard through the Open Compute Project, addresses a bottleneck in AI infrastructure: keeping massive GPU clusters running efficiently during the weeks- or months-long training runs required to build frontier AI models.
MRC works by enabling multiple network paths between nodes in a training cluster, according to OpenAI. When a single network link fails or degrades — a common occurrence in clusters with tens of thousands of GPUs — the protocol can reroute traffic across alternate paths without interrupting the training job. Traditional protocols like TCP were not designed for the unique demands of AI supercomputers, where even brief network disruptions can cascade into costly training interruptions.
The decision to contribute MRC to the Open Compute Project, an industry consortium originally founded by Facebook to share data center hardware designs, signals OpenAI’s interest in shaping the broader AI infrastructure ecosystem. OCP contributions are freely available to any organization, meaning hyperscalers, cloud providers and AI startups alike could adopt the protocol for their own training clusters.
The release comes as the AI industry faces mounting infrastructure challenges. Companies including OpenAI, Google, Meta and Anthropic are racing to build ever-larger training clusters, with some planned facilities expected to house hundreds of thousands of GPUs. Network reliability at that scale is an engineering challenge, and failures that go undetected or unmitigated can result in lost compute time.
For the U.S. AI sector, the open release of MRC could accelerate domestic infrastructure development. American hyperscalers and AI companies building next-generation training facilities stand to benefit from a protocol purpose-built for AI workloads, potentially reducing the engineering burden of designing custom networking solutions in-house.
