Nvidia’s MRC: When ‘just Ethernet’ isn’t enough for gigascale AI - SiliconANGLE

In the context of rapidly advancing artificial intelligence technologies, NVIDIA's latest Multipath Reliable Connection (MRC) technology marks a significant shift in Ethernet's role, moving it from a ... Read original →

Nvidia’s MRC isn’t just an incremental upgrade—it redefines the communication paradigm for AI infrastructure. By transforming Ethernet from best-effort to deterministic delivery, it forces switch ASICs, optical modules, and NIC vendors to overhaul their roadmaps. Microsoft and OpenAI’s co-development signals cloud giants’ strategic push to avoid InfiniBand lock-in while demanding more than vanilla Ethernet. This pressures AMD and Intel to accelerate DPU/SmartNIC alternatives and may trigger new U.S. export controls on high-performance interconnects, raising compliance barriers for Chinese AI builders. Within 18 months, MRC could become the de facto standard for AI factories, elevating Ethernet from a commodity to a premium, value-dense layer—and reshaping who captures profit in the datacenter stack.