OpenAI open-sources MRC, a networking protocol for AI training clusters

OpenAI has released MRC — Multipath Reliable Connection — a new supercomputer networking protocol designed to improve resilience and performance in large-scale AI training clusters. The company published the specification through the Open Compute Project (OCP), making it available for anyone building AI infrastructure to adopt.

The release targets one of the least visible but most consequential bottlenecks in frontier AI development: the network that connects thousands of accelerators into a single training system. As models grow, training depends less on any individual chip and more on whether the cluster's interconnect can move data reliably between GPUs at scale. A protocol that improves both resilience and performance in those networks addresses a direct engineering constraint on how large a model a cluster can train, and how efficiently it does so.

The name describes the mechanism. "Multipath" means the protocol can use more than one route through the network between endpoints. "Reliable Connection" signals that it guarantees delivery across those routes. In a large AI training cluster, where a single failed link or component can stall or crash a training run, the ability to route traffic around failures while maintaining a reliable connection is the core value proposition OpenAI is putting forward.

The choice of OCP as the release channel matters. OCP is the open hardware and infrastructure community founded originally by Facebook, and it has become the standard venue through which hyperscalers publish data center technologies for industry-wide adoption. By submitting MRC there rather than keeping it internal, OpenAI is inviting other operators, hardware vendors, and network engineers to implement, evaluate, and extend the protocol. That is the same open-specification route the industry has used before for rack, power, and interconnect designs.

The move also fits a broader pattern among frontier labs. The infrastructure layer of AI — networking, scheduling, failure recovery — has become a competitive differentiator as much as model architecture, and labs have increasingly pushed pieces of that stack into the open. OpenAI, which operates some of the largest training clusters in the industry, is contributing a protocol born of that operational experience.

For network engineers and AI infrastructure teams, the immediate significance is practical: MRC is now a published, implementable option for the reliability and performance problems that dominate large training runs. Its real-world impact will depend on adoption — whether hardware and software vendors build support for it, and whether other large-scale operators find it outperforms existing approaches in production clusters.