As data centers pack in more GPUs for AI training and inference, network switching silicon is emerging as a critical bottleneck. According to Data Center Knowledge, vendors including Cisco, Broadcom, and Nvidia are converging on 51.2 to 102.4 Tbps class Ethernet switches designed to keep bursty AI traffic moving without stranding expensive compute.
Sameh Boujelbene, Vice President of Research at Dell’Oro Group, framed the stakes bluntly: a company can spend billions on GPUs, but without a fabric that delivers predictable bandwidth and low latency, it ends up with an expensive collection of stranded chips rather than a functioning AI supercomputer.
Three Ways AI Infrastructure Scales
AI clusters generally grow along three dimensions. Scale up adds more compute per server or rack through high-bandwidth intra-node links. Scale out connects more racks together to run larger models or more concurrent jobs. Scale across links multiple data centers over optical networks so clusters in different locations can work together. Each dimension demands different network characteristics, from tight intra-rack latency to deep buffering and telemetry across wide-area links.
Cisco’s Silicon One G300 and P200
Cisco is addressing these needs with two new Silicon One chips, first announced in February 2026 and detailed further at Cisco Live in June. Broad availability is expected before the end of the year.
- G300: Targets scale-out networking between racks of GPUs inside a data center, delivering 102.4 Tbps aggregate bandwidth via 512 lanes at 200 Gbps each. It includes real-time telemetry, identity-aware forwarding, and congestion avoidance designed to instantly reroute packets around bottlenecks.
- P200: Focused on scale-across connectivity between data centers over optical networks, offering 51.2 Tbps via 512 x 100 Gbps links paired with external high-bandwidth memory for deep buffering. Cisco says this lets optical fiber run close to capacity for extended periods. Select customers already have early access, with 28.8 Tbps switches arriving in Q3 2026 and 51.2 Tbps models by year end.
Because this class of silicon generates significant heat, Cisco is building liquid cooling into its top-end Nexus 9000 switches, using cold plates similar to those used on GPUs. Cisco offers a mix of fully liquid-cooled, partially liquid-cooled, and air-cooled switch options depending on customer needs.
Hyperscalers First, Enterprises Later
Boujelbene expects hyperscalers, neoclouds, and sovereign AI cloud providers to adopt this silicon first, since most enterprise data centers do not yet have workloads that justify it. Over time, the technology is expected to filter down into more mainstream environments. Her guidance to enterprises planning serious GPU deployments is to treat networking as part of the compute system design from the start, rather than an afterthought.
Competing Silicon
Cisco is not alone in this space. Broadcom’s Tomahawk 6 and Nvidia’s Spectrum-6 are pursuing similar scale-out and scale-across targets, reflecting an industry-wide shift toward treating the network fabric as a first-order design decision for AI infrastructure.