NVIDIA has announced that Meta and Oracle will enhance their AI data center networks using NVIDIA Spectrum-X Ethernet switches, marking a major step in building the next generation of hyperscale AI infrastructure.
Both companies are adopting Spectrum-X as an open, accelerated networking architecture designed to speed up large-scale deployments, boost AI training efficiency, and reduce time to insights. As one of the core parts that enables hyperscalers to link millions of GPUs into one massive computing system for training trillion-parameter models, Oracle plans to build giga-scale AI factories powered by NVIDIA’s Vera Rubin architecture and interconnected by Spectrum-X Ethernet, while Meta integrates Spectrum Ethernet switches into its networking stack for the Facebook Open Switching System (FBOSS) to improve scalability and efficiency.
The hardware combines Spectrum-X switches and SuperNICs in the first Ethernet system purpose-built for AI to achieve tremendous results, such as a record 95% data throughput efficiency using NVIDIA’s advanced congestion-control technology versus standard setups typically at around 60%. The platform’s Spectrum-XGS technology even supports scale-across networking, linking data centers across vast distances to form global AI super-factories.
By leveraging NVIDIA’s full-stack ecosystem from GPUs and CPUs to NVLink and AI-optimized software, Spectrum-X ensures seamless, predictable performance for massive AI workloads through adaptive routing, congestion control, and AI-driven telemetry.
