NVIDIA’s Spectrum-X Ethernet platform is set to revolutionize AI data center performance, with Meta and Oracle adopting the technology to meet the ever-growing demands of large-scale AI models. Spectrum-X, touted as the ‘nervous system’ for connecting millions of GPUs, is designed to facilitate the transformation of data centers into high-performance ‘AI factories,’ according to NVIDIA CEO Jensen Huang.
Oracle plans to integrate Spectrum-X into its Vera Rubin architecture, aiming for a significant boost in GPU interconnectivity and overall AI processing power. Meta, similarly, will incorporate the switches into its Facebook Open Switching System (FBOSS) to enhance its AI infrastructure and support increasingly complex models.
NVIDIA’s MGX modular system provides partners like Meta and Oracle with the flexibility to combine CPUs, GPUs, storage, and networking components, optimizing for specific AI workloads. This modular design, coupled with Spectrum-X Ethernet’s adaptive routing and telemetry-based congestion control, allows for efficient linking of millions of GPUs while maintaining peak performance. The Spectrum-X platform, encompassing Ethernet switches and SuperNICs, represents a substantial advancement over standard Ethernet solutions, specifically tailored for the unique demands of AI. Vera Rubin architecture is anticipated to be available in the latter half of 2026. NVIDIA is also collaborating on power components and data center designs to support the 800-volt DC transition.
Ultimately, NVIDIA envisions Spectrum-X as a key enabler of more efficient and accessible AI infrastructure, designed from the ground up to handle the most demanding AI workloads and unlock the full potential of large-scale AI systems.