Built for Vera Rubin, NVIDIA Spectrum-6 Arrives in Gigascale AI Factories

Details

NVIDIA has released Spectrum-6, the latest generation of its Spectrum-X Ethernet platform, delivering 102.4 terabits per second of switching capacity — twice the bandwidth of the previous generation. The switch is built specifically to support the Vera Rubin GPU platform and is designed to let hundreds of thousands of GPUs operate as a unified, synchronized computing resource for large-scale training and inference. The practical problem Spectrum-6 addresses is coordination overhead: at gigascale, slow or unreliable network links force GPUs to wait during collective communication operations, throttling overall throughput even when individual GPUs are fast. Spectrum-6 targets this with higher bandwidth, improved resilience, and what NVIDIA describes as end-to-end AI factory management. CoreWeave, Microsoft, Nebius, SpaceXAI, and Tesla are named as first adopters. CoreWeave highlighted the switch's role in keeping GPU clusters in tight synchronization during frontier model training. Nebius emphasized resilience for long-running jobs. The product anchors the Vera Rubin infrastructure stack alongside the NVL72 GPU superchip system.