NVIDIA Unveils AI Switch Chip Spectrum-6! Analysts: Could Double Token Generation Capability

GPU leader NVIDIA unveiled its next-generation Spectrum-6 Ethernet switch chip on Tuesday, May 21, positioning it as the core component of its large-scale AI infrastructure based on the Vera Rubin platform, signaling a shift where networking transitions from a supporting role to a central 'actor' in AI factories.

The Spectrum-6 chip delivers a staggering 102.4Tbps of switching capacity per chip—double that of the previous generation—enabling connections across hundreds of thousands of GPUs and CPUs to support large-scale AI training and inference.

Spectrum-X Ethernet switches equipped with this chip offer 1.6x higher AI network performance compared to standard Ethernet, maintaining 95% network efficiency even in deployments exceeding 100,000 GPUs. Spectrum-6 supports both pluggable optical modules and co-packaged optics (CPO) technology, and employs liquid cooling for thermal management. The chip is now in full mass production and is rapidly being adopted across gigascale AI factories worldwide.

China Galaxy Securities noted that Spectrum-X Ethernet Photonics is the world's first large-scale 200G/lane CPO switch, validating the value of optical interconnects in next-generation data centers and expected to drive rapid growth in upstream optical chip demand.

Kaiyuan Securities analyzed that as AI model parameters expand, computing power growth faces diminishing returns due to single-chip power consumption and interconnect bottlenecks, making networking the new 'ceiling' for compute performance. For example, in NVIDIA's NVL72 architecture, the switch-to-GPU ratio has increased from the traditional 1:21 to 1:2, highlighting the rising importance of networking within computing clusters.

According to the latest IDC data, NVIDIA surpassed traditional networking giants Broadcom and Cisco for the first time in Q1 of fiscal 2026, becoming the global revenue leader in data center Ethernet switches.

Initial adopters including CoreWeave, Microsoft, SpaceXAI, and Tesla have confirmed deployment of NVIDIA's Spectrum-6.

CoreWeave stated that this technology will provide customers with the critical bandwidth and flexibility needed to train advanced models.

NVIDIA emphasized that 'AI performance is fundamentally a networking problem,' and system-level co-architectures are becoming the key path to overcoming single-chip performance limits.