OpenAI Rewired AI Networking. The Fix Breaks an Old Orthodoxy.
OpenAI just rewired one of the hardest problems in AI infrastructure — and the fix breaks with decades of networking orthodoxy.
When you connect thousands of GPUs in a single training run, packets get lost. That's not a software bug — it's physics. Ethernet was designed for links where occasional packet loss is normal, and network equipment handles it by simply retransmitting what gets dropped. At AI supercomputer scale, however, retransmission delays compound fast enough to stall training jobs costing hundreds of thousands of dollars per hour. The industry standard fix is a mechanism called Priority Flow Control, or PFC — a layer that essentially tells the network to pause and wait when congestion spikes, preventing the drops that would otherwise cascade into training interruptions. It is the foundational assumption behind every hyperscale lossless-Ethernet fabric deployed in the last fifteen years.
OpenAI's new networking protocol, called Multipath Reliable Connection, throws that assumption out. According to the OpenAI blog, MRC extends RDMA over Converged Ethernet — a technology that lets GPUs communicate directly across the network without CPU involvement — to spray packets across hundreds of paths simultaneously rather than pinning each transfer to a single route. If a path congests, MRC swaps it for another. If a path fails, MRC stops using it immediately and probes for recovery.
The architectural implication that matters most: MRC disables PFC entirely and runs Ethernet in best-effort mode. According to Microsoft, this avoids the global pause behavior that can devastate tail latency — the kind of microsecond-scale delays that, at 131,000-GPU scale, translate directly into wasted compute hours and dollars. A single switch with 64 ports running at 800 gigabits per second can instead be configured to connect 512 ports at 100 gigabits per second. That lets a network fully connect roughly 131,000 GPUs with just two tiers of switching hardware, according to OpenAI, dramatically reducing the number of hops each packet travels.
The 131,072-GPU figure is not theoretical. OpenAI says MRC is already deployed across all of its largest NVIDIA GB200 supercomputers, including its site with Oracle Cloud Infrastructure in Abilene, Texas, and in Microsoft's Fairwater supercomputers. Microsoft's own architecture description details a two-tier multiplane topology enabling 100,000-plus GPU scale with graceful degradation — losing a network interface card port reduces bandwidth but does not crash running jobs, and flapping links between switching tiers often go unnoticed by applications. NVIDIA has deployed MRC in its Blackwell generation hardware and supports the protocol on its Spectrum-X Ethernet platform.
The collaboration behind MRC spans the industry's major chip and systems makers. OpenAI developed MRC in partnership with AMD, Broadcom, Intel, Microsoft, and NVIDIA, and AMD has published its own post detailing production validation of the protocol at scale with a leading cloud provider. OpenAI released MRC as an Open Compute Project contribution, making the specification publicly available — a notable move for a company whose infrastructure announcements typically stay vendor-controlled.
What the company posts collectively confirm is this: disabling PFC at 131,000-GPU scale is the architectural gamble, and it works by distributing traffic so evenly that the congestion PFC was designed to prevent never forms in the first place. MRC sidesteps the lossless-Ethernet orthodoxy entirely rather than engineering around its failure modes.
The InfiniBand question is the counterforce. NVIDIA's own blog post positions MRC alongside its existing InfiniBand portfolio rather than as a replacement — NVIDIA says Spectrum-X combines Spectrum-4 Ethernet with BlueField-3 DPUs and NVLink, while noting that MRC deploys on the same platform alongside existing networking choices. That framing suggests the protocol is additive for NVIDIA's customer base rather than a clean displacement of InfiniBand. For buyers evaluating AI training infrastructure today, the practical question is not whether MRC beats InfiniBand in the abstract, but whether the operational simplicity of an Ethernet-only fabric — no lossless networking overhead, no vendor lock-in — justifies the transition cost at their scale.
The second-order implications go in two directions. For networking vendors: if MRC becomes the de facto standard for large-scale Ethernet fabrics, the suppliers of proprietary lossless networking solutions face pressure even inside NVIDIA's own ecosystem. For cloud buyers and enterprise AI operators: an open OCP specification for AI networking, contributed by OpenAI and backed by AMD, Broadcom, Intel, Microsoft, and NVIDIA, means the underlying plumbing of frontier AI training is no longer locked inside proprietary vendor roadmaps. That shifts leverage toward buyers who can now spec OCP-compliant components from multiple suppliers rather than accepting whatever their preferred cloud provider offers.
Financial Times and Toms Hardware have reported on OpenAI's shifting compute posture — the company has effectively abandoned first-party Stargate data centers in favor of leasing compute through bilateral deals with Oracle and others, a context in which MRC's efficiency gains carry direct cost weight. MRC has been validated at OpenAI, Microsoft, and Oracle deployments. Independent adoption beyond that partner ecosystem remains unconfirmed. Whether the broader industry converges on this design — or whether competing approaches to the same problem win out — is what to watch next.