On August 14, NVIDIA announced that its Spectrum-X Ethernet silicon photonics switch had officially entered mass production, delivering cutting-edge, scalable networking infrastructure tailored for AI data centers. Leveraging Co-Packaged Optics (CPO) technology, this switch is the result of a collaborative effort involving TSMC, SPIL, Lumentum, TFC Communication, and Foxconn. The switch offers remarkable improvements, including a 75% reduction in the number of lasers required, an 80% drop in power consumption, and a tenfold increase in the mean time between failures (MTBF).

Early adopters of this innovative switch include CoreWeave, Lambda Labs, and Oracle. NVIDIA highlighted that as AI computing scales up, the primary bottleneck for AI servers is transitioning from raw computational power to the capacity for high-speed interconnections.