OpenAI Releases Open Networking Protocol to Speed Up Frontier AI Training.

OpenAI has released Multipath Reliable Connection (MRC), a new open-source networking protocol designed to make large-scale GPU training clusters faster and more reliable. The company developed MRC in partnership with AMD, Broadcom, Intel, Microsoft, and NVIDIA, and has already deployed it across its largest supercomputers used to train frontier models.

The protocol addresses a critical bottleneck in AI training: network performance. When training runs at the scale of OpenAI's Stargate project, the network connecting thousands of GPUs becomes as important as the chips themselves. If data doesn't flow efficiently between processors, expensive hardware sits idle. If network links fail, training can stall or require costly restarts that waste compute time and delay model releases.

MRC spreads training data across hundreds of parallel network paths simultaneously, rather than relying on a single route between GPUs. When a network link or switch fails, the protocol automatically reroutes traffic without interrupting the training run. This approach helps keep GPUs actively processing data instead of waiting for network bottlenecks to clear.

The result is measurably faster training with less wasted compute. OpenAI says the protocol is now running in production across all of its largest training infrastructure, including the company's supercomputer site with Oracle Cloud Infrastructure in Abilene, Texas, and Microsoft's Fairwater supercomputers.

OpenAI published the MRC specification through the Open Compute Project Foundation, making it available for other organizations to implement. The protocol's development involved five major hardware and cloud infrastructure companies, suggesting industry appetite for a common standard as training clusters grow larger.

The move reflects OpenAI's broader compute strategy of rethinking infrastructure from the ground up. Rather than simply buying more GPUs, the company is optimizing each layer of the training stack, from networking protocols to data center design, to extract more performance from existing hardware.

For context, OpenAI has been systematically expanding its infrastructure partnerships over the past year. The company previously launched a safety bug bounty focused on identifying AI abuse risks, but this networking protocol release signals a shift toward fundamental infrastructure optimization.

As AI models grow larger and training runs consume more compute, efficiency gains at the infrastructure layer compound quickly. A protocol that reduces idle GPU time by even a small percentage can translate to millions of dollars in saved compute costs and weeks shaved off training timelines.

The open-source release also positions OpenAI as a contributor to industry-wide infrastructure standards, not just a consumer of compute. With Stargate and other large training initiatives pushing the boundaries of scale, solving networking challenges benefits the entire AI research community.

The MRC specification is available now through the Open Compute Project Foundation.