OpenAI releases MRC networking protocol

Hot take: this is infrastructure becoming a product. OpenAI is turning hard-won supercomputer networking work into an open standard, which matters more than the protocol name itself.

• –MRC sprays a single transfer across hundreds of paths and uses SRv6 source routing to route around failures in microseconds
• –OpenAI says it is already running in production on its largest clusters, including OCI in Abilene and Microsoft Fairwater, so this is battle-tested, not theoretical
• –The main upside is fewer stalled training jobs and less wasted GPU time during congestion, link flaps, and switch maintenance
• –If the OCP contribution gets traction, it could shape how future AI fabrics are built at hyperscale
• –The protocol angle also signals how frontier labs are increasingly building their own stack when commodity networking is not enough