NVIDIA's CPO switches are now in full production—this is a major milestone in the optical interconnect field. Yesterday, NVIDIA announced that Spectrum-X Ethernet Photonics has entered full-scale production, supporting the next-generation, large-scale network infrastructure for AI factories. Spectrum-X Ethernet Photonics is the silicon photonics (Silicon Photonics)/Co-Packaged Optics (CPO) switch technology within the NVIDIA Spectrum-X Ethernet networking platform, primarily designed for large-scale AI factories and high-performance AI data center scenarios. It integrates the silicon photonics engine and switch ASIC directly into the same package, replacing traditional pluggable optical modules, and is specifically designed for network interconnects in ultra-large-scale GPU clusters. Core application scenarios: 1) Large-scale AI training and inference clusters/AI factories: Supports interconnection for tens of thousands to hundreds of thousands, or even millions of GPUs, providing high bandwidth, low latency, and high effective bandwidth communication between GPUs. It is particularly suitable for generative AI, large model training, and distributed inference. This technology enables "GW-scale" AI infrastructure while improving overall energy efficiency and reliability. 2) Internal networks of ultra-large-scale data centers (commonly referred to as scale-out networks): Addresses issues with traditional Ethernet under ultra-large-scale AI workloads, such as high power consumption, multiple points of failure, and poor signal integrity. Compared to traditional pluggable transceiver solutions, it achieves approximately 5x network energy efficiency improvement, reduces the number of lasers by about 4x, improves Mean Time Between Failures (MTBF) by about 10x, and extends the continuous runtime of AI applications. MTBF is particularly important—one major consideration for cloud providers adopting CPO is the failure rate and interval between failures. 3) Cross-data center scaling (Scale-cross scenarios, combined with Spectrum-XGS): Can also work with related technologies to connect multiple dispersed data centers (even those far apart) into a unified, large-scale AI super factory. This is NVIDIA's next-generation network infrastructure specifically built for "AI factories," which require massive GPU high-speed interconnects, extreme energy efficiency, and reliability. It has now entered full production, with initial customers including CoreWeave, Lambda, and Oracle. Supply chain: - TSMC is responsible for silicon photonics manufacturing; - SPIL handles chip-level packaging and testing; - Lumentum provides laser chip manufacturing; - TFC handles laser module subcomponent assembly; - Foxconn provides communication switch system assembly. One of the reasons why optical interconnects were so promising back in early August was seeing NVIDIA's Senior Vice President Gilad Shainer announce at a tech forum that CPO had started production. Now, we finally see the official announcement confirming full-scale production!