NVIDIA CPO Switch Enters Mass Production

According to NVIDIA, its Spectrum-X Ethernet silicon photonics switch has officially entered full-scale mass production as of early August 14 local time. The product adopts co-packaged optics (CPO) technology and is described by NVIDIA as “the world’s first 200G/lane CPO Ethernet switch system to enter mass production.”

The newly mass-produced system integrates CPO technology directly with the switching chip, fundamentally changing the signal conversion approach used in traditional pluggable optical modules. By placing the optical engine and switch chip on the same substrate, the CPO architecture significantly shortens the optical-electrical signal transmission path, delivering major improvements in power efficiency, reliability, and data transmission performance.

According to NVIDIA, compared with conventional solutions, the Spectrum-X CPO Ethernet switch can achieve up to five times higher network power efficiency, five times longer AI application uptime, and a tenfold increase in mean time between failures (MTBF). These improvements are designed to address the growing demands of large-scale AI computing clusters, where network performance and energy consumption have become critical challenges.

The development of NVIDIA’s CPO switch also highlights close collaboration across the semiconductor and photonics ecosystem. Foxconn is responsible for system assembly, Lumentum provides laser chip manufacturing, Siliconware Precision Industries (SPIL) handles wafer-level and chip-level packaging and testing, TFC Optical Communication supports laser module subassembly, and TSMC provides advanced silicon photonics manufacturing capabilities. This cooperation covers key stages from photonic chip production and advanced packaging to system integration, demonstrating the importance of supply chain coordination in bringing CPO technology into commercial deployment.

As AI data centers continue expanding with larger GPU clusters, internal network communication requirements are increasing rapidly. Traditional pluggable optical modules are facing limitations in power consumption, bandwidth density, and reliability. Optical interconnect technology is becoming a key solution for overcoming communication bottlenecks in next-generation AI infrastructure.

NVIDIA’s move to mass-produce CPO-based Ethernet switches marks a significant transition of this technology from research and validation into large-scale commercial application. The technology is expected to provide AI data centers with more energy-efficient, reliable, and high-performance networking infrastructure to support the continued growth of artificial intelligence workloads.