OpenAI and Five Chip Giants Release MRC to Cut Wasted GPU Time in AI Training Clusters

2. OpenAI and Five Chip Giants Release MRC to Cut Wasted GPU Time in AI Training Clusters

OpenAI, AMD, Broadcom, Intel, Microsoft, and Nvidia jointly released Multipath Reliable Connection (MRC) on May 4, 2026, an open networking protocol designed to make large AI training clusters faster and more reliable. MRC addresses idle GPU time caused by network congestion and packet loss inside the high-speed fabrics that connect thousands of accelerators during training runs. OpenAI confirmed MRC is already running on its largest training supercomputers, meaning the protocol is not a research proposal but a production-grade specification being published for industry adoption.

The strategic weight here is in who signed it. This is not a unilateral OpenAI infrastructure paper. Pulling AMD, Broadcom, Intel, Microsoft, and Nvidia into a single open specification creates immediate pressure on cloud providers and hyperscalers who have built proprietary networking stacks. AWS's EFA, Google's Jupiter fabric, and Meta's RoCE-based infrastructure all solve the same congestion problem with closed approaches. MRC positions an open alternative backed by the dominant hardware vendors, which shifts the default toward interoperability and makes proprietary lock-in harder to justify to enterprise customers building multi-vendor clusters.

The broader pattern is an infrastructure standards race running in parallel with the model capability race. Whoever controls the networking layer shapes training economics, and training economics shape who can afford frontier model runs. Watch whether cloud providers adopt MRC natively or fork it, and whether Nvidia's backing here conflicts with its existing InfiniBand ecosystem investments. The next signal to track: whether MRC gets submitted to an open standards body like the IETF or stays a consortium specification, which will determine how fast it spreads beyond the founding five.

Source: @OpenAI on X