Nvidia unveiled Spectrum-X Ethernet, a networking architecture designed to address traffic congestion in AI data centers. The system claims up to 1.6 times the network throughput of conventional Ethernet solutions, with failover time reduced from 1.08 seconds to 2.68 milliseconds, representing a 99% improvement. The Scale-In capability introduced at the Hot Chips conference in August 2026 constitutes the fifth pillar of AI networking, enabling multiplane topology scaling to over 512,000 endpoints in a two-tier configuration. Nvidia is betting that the AI market will standardize on Ethernet over InfiniBand, with direct competitors including Arista Networks and Broadcom in the high-speed Ethernet switching market.
Source: Read the original article
