OpenAI released MRC, a new supercomputer networking protocol, through the Open Compute Project. The system improves resilience and performance within large-scale training clusters by managing multipath connections. This reduces training downtime caused by network failures. Engineers can now implement more stable infrastructure for massive model runs without relying on proprietary hardware locks.