Cisco Silicon One G300 and 1.6T Optics Power AI Networking
3 min readCisco is putting itself at the center of the AI networking buildout with a refreshed Silicon One lineup, new air- and water-cooled switches, and a stack of pluggable optics aimed squarely at the gigawatt-class GPU clusters being stood up across hyperscalers, neoclouds, and sovereign cloud operators.
The centerpiece is the Silicon One G300, a 102.4 Tbps switching chip with on-chip 200 Gbps SerDes, 1.6T Ethernet ports, and radix scaling up to 512 ports. Cisco is pairing it with two fixed 64-port boxes: the 3RU air-cooled 8133, and the 2RU 8132, which is the company’s first commercially available water-cooled switch.
What the G300 Actually Does for AI Clusters
The G300 includes a new feature Cisco calls Intelligent Collective Networking, which combines shared packet buffers, path-based load balancing, and proactive telemetry to handle the bursty, all-to-all traffic patterns of large AI training and inference jobs. Senior VP Nick Kucharewski told Network World the chip’s larger packet buffer delivered a 33% throughput improvement in simulation, and a 28% reduction in job completion time compared with advanced packet-spray implementations.
The practical translation: more GPUs reachable per switch, fewer hops between them, lower latency on training collectives, and lower capex per deployed GPU. Cisco frames the G300 as the leaf of its three-tier AI networking strategy, with the existing G200 (51.2 Tbps) sitting at the spine and aggregation layer and the G100 (25.6 Tbps) at the edge.
Optics Are Doing the Heavy Lifting
Alongside the silicon, Cisco rolled out a deeper optics portfolio for AI scale-out and data center interconnect. The new 8122X-64EF-O is a 64x800G switch that runs SONiC and supports Cisco’s 800G Linear Pluggable Optics, which strip out the DSP to cut power on short reaches. Cisco’s existing P200-powered 8223 router gets a new 28.8T line card that pushes total system bandwidth to 518.4T in the 18-slot chassis, and pairs with 800G ZR and ZR+ coherent pluggables for data center interconnect links over 1,000 kilometers.
The market is validating the strategy faster than analysts expected. In Cisco’s most recent quarter, AI-related optics orders nearly quadrupled to roughly $950 million, and Acacia coherent optics crossed $1 billion in quarterly orders for the first time. Cisco raised its full-year AI infrastructure order forecast from $5 billion to $9 billion.
Why It Matters
AI networking has quietly become a category of its own, and switch fabric plus optics is where the next round of margin in AI infrastructure will be earned. Nvidia owns the GPU and its NVLink fabric, but everything that moves data between racks, halls, and data centers is contested ground. Cisco is betting that customers who want Ethernet-based scale-out, merchant silicon, and a clean separation of optics from switching ASICs will pick its boxes over a fully bundled Nvidia stack.
Watch two things from here. First, whether the 8132 water-cooled switch shows up in real hyperscale deployments, which would mark the moment liquid cooling moves from GPU racks to the switch layer. Second, whether Cisco’s Linear Pluggable Optics gain traction over traditional DSP-equipped modules; if they do, the power and cost curve for AI scale-out networks shifts in Cisco’s favor.