← Back to feed

On May 5, 2026, OpenAI published a technical specification for Multipath Reliable Connection (MRC) through the Open Compute Project, co-authored with NVIDIA, Microsoft, AMD, Broadcom, and Intel. The protocol is already running in production inside OpenAI’s Stargate supercomputer, Microsoft’s Fairwater data center, and Oracle Cloud’s Abilene facility.

MRC addresses a specific failure mode that becomes increasingly costly as training clusters grow: when a single GPU connection uses one network path, any congestion, failed link, or overloaded switch delays the entire job. At 100,000+ GPU scale, the probability of at least one path failure during a training step approaches certainty.

The architecture change

Conventional RDMA over Converged Ethernet (RoCEv2) treats a network interface as a single 800 Gbps link. MRC breaks that into 8 parallel 100 Gbps paths operating simultaneously. This produces a switch radix advantage that compounds at cluster scale:

• Old design: 64 ports at 800 Gbps per switch
• MRC: 512 ports at 100 Gbps per switch
• Result: 131,000 GPUs reachable with 2 switch tiers instead of 3-4

Two switch tiers versus four means fewer hops, fewer failure points, and significantly lower power consumption per GPU. The simplification is roughly equivalent to removing one full layer of spine switches from every large AI factory.

Adaptive packet spraying

The second core innovation is congestion handling. Traditional RoCEv2 paths all send traffic down the single available route — congestion on that path backs up the entire flow. MRC implements adaptive packet spraying: each data transfer is split across hundreds of parallel paths simultaneously. Packets carry their complete memory address, so the destination can reconstruct the original data even if packets arrive out of order. When any path shows congestion, the sender reroutes to an idle one in real time without waiting for software intervention.

OpenAI’s head of industrial compute, Sachin Katti, noted that “MRC’s end-to-end approach enabled us to avoid much of the typical network-related slowdowns and interruptions and maintain the efficiency of frontier training runs at scale.”

Routing overhaul

MRC also replaces dynamic routing protocols (BGP) with IPv6 Segment Routing. Under the new design, the sender specifies the complete path at transmission time — static source routing. This eliminates an entire class of routing failures caused by BGP convergence delays and removes the BGP state machines from the data plane entirely.

The specification was released to the Open Compute Project rather than kept proprietary. OpenAI’s stated rationale: shared standards in infrastructure layers reduce complexity for the whole industry. If major cloud providers and AI infrastructure vendors adopt MRC, it could standardise how large-scale GPU clusters handle networking in the same way RoCEv2 standardised RDMA over Ethernet.

Production deployment

Stargate, the primary training cluster for GPT-5.5 and future OpenAI frontier models, uses MRC as its base networking protocol. The scale reported: “100,000+ GPU clusters.” Microsoft Fairwater (hundreds of thousands of NVIDIA GB200s, opened ahead of schedule in Q1 2026) and Oracle Cloud Infrastructure’s Abilene data center run MRC-based fabrics.

The specification is available through the Open Compute Project. First-generation hardware support is delivered via NVIDIA Spectrum-X Ethernet, though the OCP release is designed to enable adoption across other switch and NIC vendors.