NVIDIA Launches Spectrum-6 AI Switch Chip

According to NVIDIA, the company has launched its next-generation Spectrum-6 Ethernet switch chip, a key networking component for AI infrastructure based on the Vera Rubin platform. The release marks a major shift in AI data center architecture, with high-performance networking moving from a supporting role to becoming a core element of next-generation AI factories.

The Spectrum-6 delivers a single-chip switching capacity of up to 102.4Tbps, doubling the bandwidth of its predecessor and enabling connectivity for hundreds of thousands of GPUs and CPUs. Designed for large-scale AI training and inference workloads, the chip powers NVIDIA’s Spectrum-X Ethernet switches, which deliver 1.6 times higher AI networking performance than traditional Ethernet solutions and maintain up to 95% network efficiency even in deployments exceeding 100,000 GPUs.

The Spectrum-6 supports both pluggable optical modules and co-packaged optics (CPO) technology, while adopting liquid-cooling solutions to address the increasing thermal challenges of advanced AI infrastructure. The chip has entered full-scale mass production and is expected to accelerate deployment across global gigawatt-scale AI factories.

China Galaxy Securities noted that NVIDIA’s Spectrum-X Ethernet Photonics is the world’s first CPO switch platform to achieve large-scale mass production with 200G-per-lane technology, demonstrating the growing importance of optical interconnects in next-generation data centers and driving demand for optical semiconductor components.

Open Source Securities also highlighted that as AI models continue to expand in size and complexity, computing performance gains are increasingly constrained by single-chip power consumption and interconnect bottlenecks. Networking infrastructure has become a critical factor limiting AI computing scalability. Taking NVIDIA’s NVL72 architecture as an example, the ratio of switch chips to GPUs has increased from the traditional 1:21 configuration to 1:2, reflecting the growing importance of high-speed networking in AI computing clusters.

According to the latest data from IDC, NVIDIA surpassed traditional networking leaders such as Broadcom and Cisco in the first quarter of fiscal 2026, becoming the top company worldwide in data center Ethernet switch revenue.

Early adopters of NVIDIA Spectrum-6 include CoreWeave, Microsoft, SpaceX AI, and Tesla. CoreWeave stated that the technology will provide customers with the high bandwidth and flexible networking capabilities required for training advanced AI models.