Its Spectrum-6 Ethernet switch system is beginning to arrive in large AI data centers built around the company’s Vera Rubin platform. The system delivers 102.4 terabits per second of switching capacity, twice the capacity of the previous generation, according to NVIDIA.
The company frames networking as a central bottleneck for modern AI factories. Large training and inference jobs depend on thousands of accelerators exchanging data constantly, and a slow network link can stall far more expensive GPU capacity.
NVIDIA says CoreWeave, Microsoft, Nebius, SpaceXAI, and Tesla are among the first builders bringing Spectrum-6 into AI factories. CoreWeave, Microsoft, and Nebius are also listed as early providers deploying Vera Rubin-based infrastructure with Spectrum-6.
The announcement is aimed at cloud providers and companies building their own clusters, not individual developers. Its practical claim is that higher-bandwidth Ethernet can improve utilization, resilience, and token economics at extreme scale. The details that matter for customers will be availability, integration with existing networks, and real workload performance.