NVIDIA Adds Multi-Rail Connectivity to Spectrum-X Ethernet Fabric
NVIDIA has added Multi-Rail Connectivity to its Spectrum-X AI-native Ethernet networking fabric, targeting gigascale AI infrastructure deployments, according to the NVIDIA Blog.
The update to Spectrum-X, which NVIDIA describes as an open, AI-native Ethernet fabric, introduces MRC — a feature designed to improve network performance and efficiency for large-scale AI training and inference clusters. Multi-Rail Connectivity allows GPU servers to simultaneously utilize multiple network paths, increasing the available bandwidth for the data-intensive communication required during distributed AI workloads.
The update carries implications for the broader AI infrastructure market, where networking has emerged as a critical bottleneck as clusters scale to tens of thousands of GPUs. Spectrum-X competes in a space historically dominated by NVIDIA’s own InfiniBand technology, which has been the preferred networking fabric for the largest AI supercomputers, including those operated by OpenAI, Meta and other leading AI labs.
NVIDIA has positioned Spectrum-X as a complementary Ethernet-based option that gives hyperscalers and cloud providers more flexibility in designing AI data center networks. Ethernet’s open standards and broad ecosystem support make it attractive to operators who want to avoid proprietary lock-in while still achieving the high throughput and low latency required for AI workloads.
The MRC addition addresses a key challenge in scaling AI clusters: ensuring that networking capacity grows proportionally with compute capacity. As AI providers deploy increasingly dense GPU configurations — often using NVIDIA’s own Blackwell and Hopper architectures — the network must keep pace to avoid becoming a performance bottleneck.
The update has direct implications for U.S. hyperscalers, cloud providers and AI labs that are investing billions in domestic AI infrastructure. Companies including Microsoft, Google, Amazon and Meta have all announced major data center expansions, and the choice of networking fabric is a key architectural decision that affects both performance and cost at scale.
NVIDIA’s dual approach — offering both InfiniBand through its Quantum platform and Ethernet through Spectrum-X — gives the company coverage across the full spectrum of AI networking deployments, from purpose-built AI supercomputers to general-purpose cloud infrastructure being adapted for AI workloads.
