Cisco unveils AI networking silicon and Nvidia optics at OCP Summit

At the annual Open Compute Project Global Summit, Cisco positioned its new hardware as a response to a specific bottleneck: as frontier AI models grow, they need clusters of thousands of GPUs and other accelerators to act as one giant processor. Moving enough data between chips without inflating hardware costs or power use has become a central challenge for data center operators.

Cisco's answer is a "scale-across" architecture that links distributed AI data centers into a unified compute fabric, built on more coherent optics and deep-buffered systems. The company says this open approach lets operators swap components from different vendors. The updated Silicon One chip family sits at the center, with the flagship G300 now shipping at volume and the new G302 switch arriving before the end of the year.

The Nvidia collaboration on repluggable optics points to the same goal from a different angle. Cisco said the module packs 16 full-duplex lanes into a standard form factor and offers about four times the bandwidth density and 1.3 times the power efficiency of existing 200G optics. Cisco also introduced the Cisco 8001 out-of-band management system and a Live Protect security feature for the SONiC operating system, available next year.