OpenAI, Tech Giants Unveil Open-Source Protocol to Speed AI Supercomputers
SAN FRANCISCO — OpenAI and five major U.S. technology companies this week released an open-source networking protocol designed to eliminate bottlenecks in AI supercomputers, according to The Decoder.
The protocol, called MRC, was built in collaboration with AMD, Broadcom, Intel, Microsoft and NVIDIA. It sends data across hundreds of simultaneous paths between GPUs, unlike conventional networking approaches that rely on single or limited routing channels.
A key technical feature is a reduction in network switch layers. Traditional AI supercomputer architectures require three or four layers of switches to connect GPUs at scale. MRC reduces that requirement to just two layers, according to The Decoder, which translates to lower power consumption and reduced infrastructure costs.
OpenAI has already deployed MRC on its Stargate supercomputer, the AI infrastructure project announced in January 2025 as a joint venture with SoftBank and Oracle. The deployment marks a real-world test of the protocol beyond laboratory settings.
Why It Matters
As AI models grow larger and training runs demand more compute, the networking fabric connecting thousands of GPUs has become a primary constraint. Data must move between processors at high speeds, and each additional switch layer introduces latency, power draw and hardware cost. By flattening the network topology, MRC addresses a primary scaling challenge in AI infrastructure.
The decision to release MRC as open source comes as proprietary networking solutions face scrutiny for potentially slowing AI development. All five partner companies — AMD, Broadcom, Intel, Microsoft and NVIDIA — are U.S.-based or U.S.-listed, reflecting the domestic concentration of AI infrastructure development.
Strategic Implications
The six-company collaboration spans direct competitors. AMD and NVIDIA compete in the GPU market, while Broadcom supplies networking chips that connect those processors. Intel and Microsoft contribute semiconductor and cloud infrastructure capabilities, respectively. The companies developed a shared open standard for the protocol.
For the broader AI industry, MRC could lower the barrier to building large-scale training clusters by reducing the networking hardware and power required. Data center operators, cloud providers and national AI initiatives may benefit from a standardized, open protocol that works across multiple hardware vendors.
The protocol’s deployment on Stargate — a project with stated national competitiveness goals — has drawn attention from U.S. policymakers who have emphasized domestic AI infrastructure as a strategic priority. An open networking standard developed by American companies could strengthen that position internationally.
