Traditional Ethernet has officially hit a wall in the construction of modern LLM infrastructure. When you attempt to train next-generation models on clusters spanning hundreds of thousands of GPUs, standard congestion management methods effectively fall apart. According to a joint report from OpenAI, Microsoft, NVIDIA, AMD, and Broadcom, the process operates in lock-step: total computational speed drops to the level of the slowest link. Consequently, network connectivity has become a scarcer and more expensive resource than the compute cores themselves—a minor network glitch can cost millions of dollars while high-end hardware sits idle waiting for data.
Industry giants are tackling the problem of "tail latency," which previously paralyzed thousands of chips due to a single congested node, through the Multipath Reliable Connection (MRC) protocol. This is a new RDMA-based transport protocol that implements "packet spraying." As explained by the architects—including OpenAI’s Joao Araujo and Mark Handley—MRC aggressively balances loads across all available paths. Paired with the SRv6 protocol, this allows the system to literally flow around bottlenecks. Microsoft and AMD engineers estimate that this architecture transforms disparate hardware into a monolithic environment that maintains stability even during link failures.
The economics are straightforward: implementing MRC directly impacts Total Cost of Ownership (TCO). By moving to the two-tier Multi-plane Clos topologies advocated by NVIDIA and OpenAI, it becomes possible to efficiently utilize clusters exceeding 100,000 GPUs. This is no longer theoretical; Microsoft and OpenAI have confirmed that this specific stack powers the training of their latest models. Instead of simply buying more chips, companies are investing in intelligent networking to maximize the utilization rates of existing hardware.
However, full autonomy—where infrastructure is truly "self-healing"—remains out of reach. While MRC and SRv6 static routing allow the system to bypass network failures, a physical compute node failure still requires manual intervention and external coordination. This creates a deceptive illusion of reliability: while the network learns to be immortal, the hardware remains an Achilles' heel. The gap between automated error bypass and true fault tolerance is still significant.