The Counterintuitive Networking Decisions Behind OpenAI’s 131,000-GPU Training Fabric

To prevent single slow data transfers from stalling massive AI training jobs, OpenAI's 131,000-GPU supercomputers use a radical networking architecture. Called Multipath Reliable Connection (MRC), the system abandons traditional routing protocols and instead splits each GPU's connection into eight independent, parallel network planes. MRC then sprays individual data packets across hundreds of different random paths simultaneously, rather than pinning a transfer to a single route. This counterintuitive design ensures that the failure or congestion of any one link has a negligible impact, providing predictable performance and resilience at an unprecedented scale.