OpenAI Unveils MRC Networking Protocol and Stargate Supercomputer Infrastructure Plan

As generative AI models scale, hundreds of thousands of GPUs must exchange data in sync. Network congestion or a single-path failure can slow training. OpenAI announced the Stargate Project on January 21, 2025, with plans to invest $500 billion in the United States over four years and deploy an initial $100 billion. Improving supercomputer network efficiency is therefore critical to expanding computing capacity.

On May 5, 2026, OpenAI published the Multipath Reliable Connection (MRC) specification, developed with NVIDIA, AMD, Microsoft, Broadcom and Intel, and submitted it to the Open Compute Project. MRC distributes data packets across multiple paths and actively routes around congestion or failures. It supports a two-tier network architecture for more than 100,000 GPUs and will serve as the networking design for Stargate infrastructure.