Meta and Oracle are adopting Nvidia's Spectrum-X Ethernet networking for their mega AI factories, standardizing Nvidia networking across largest GPU clusters.
At the Open Compute Project Summit, Nvidia announced that Meta Platforms Inc. and Oracle Corp. will adopt its Spectrum-X Ethernet networking platform, a purpose-built system designed for AI workloads that connects millions of graphics processing units into one unified fabric. Unlike traditional Ethernet solutions retrofitted for AI, Spectrum-X was engineered from the ground up to handle the communication patterns of large-scale AI workloads — massive all-to-all GPU synchronization, low-latency messaging and congestion-prone flows. The company claims 1.6 times higher networking performance for AI communication versus conventional Ethernet, a leap that directly translates to less GPU idle time and higher throughput during training and inference.
Meta's integration of Spectrum-X into its Facebook Open Switching System, or FBOSS, and Minipack3N switch marks a key moment for open networking. As Gaya Nagarajan, vice president of networking engineering at Meta, said, "Meta's next-generation AI infrastructure requires open and efficient networking at a scale the industry has never seen before." By merging Spectrum-X Ethernet with FBOSS, Meta is pairing open, programmable control planes with AI-optimized physical infrastructure to achieve predictable, congestion-free performance while preserving the flexibility of its disaggregated network model.
Oracle Cloud Infrastructure is taking a complementary approach focused on scale. Oracle is using Spectrum-X to build giga-scale AI factories powered by the upcoming Nvidia Vera Rubin architecture. According to Mahesh Thiagarajan, executive vice president of Oracle Cloud Infrastructure, "By adopting Spectrum-X Ethernet, we can interconnect millions of GPUs with breakthrough efficiency." Spectrum-X's scale-across capabilities (SpectrumXGS) allow Oracle to connect clusters across multiple data centers — even across countries — into one logical AI system.
Spectrum-X's architecture combines the Spectrum4 Ethernet switch, offering 51.2 terabits per second of throughput, with BlueField-3 SuperNICs and DPUs, which offload and secure network services so GPUs can focus purely on compute. It enables 400 gigabits per second RDMA over converged Ethernet for faster GPU-to-GPU communication and secure, multitenant environments. As part of Nvidia's full-stack strategy, Spectrum-X integrates with DOCA, Cumulus Linux, Pure SONiC, NetQ, AI Enterprise and AI Workbench for development, deployment and operational visibility.
Networking has quietly become the de facto operating system of the AI era. The industry consensus is forming around a new reality: AI's performance frontier has shifted from compute to connectivity. The network now plays the integrative role that operating systems once did — orchestrating, scheduling and synchronizing distributed resources across clouds, edges and data centers.