Network as a Primary Determinant
Sameh Boujelbene, Vice President of Research at Dell’Oro Group, highlighted the evolving role of networking in AI infrastructure. "In AI data centers, the network is becoming a primary determinant of cluster efficiency," Boujelbene stated. She added that companies can spend billions on GPUs, but without a fabric that delivers predictable bandwidth and low latency, the result is an expensive collection of stranded chips, not an AI supercomputer.AI infrastructure typically scales in three ways: 'scale up' by packing more compute per server or rack; 'scale out' by adding more racks and combining resources; and 'scale across' by connecting multiple data centers over optical networks. Each scaling dimension places distinct demands on the network, from intra-rack latency to wide-area buffering.
Cisco's Silicon One Solutions
Cisco is rolling out new Silicon One chips to meet these AI networking demands. The company positions its Silicon One G300 for scale-out within the data center, designed to manage backend networking between racks of GPUs. This chip offers an aggregate bandwidth of 102.4 Tbps via 512 lanes at 200 gigabits per second (Gbps) per lane, integrating real-time telemetry, identity-aware forwarding, and traffic visibility. Nick Kucharewski, Senior Vice President and General Manager, Cisco Silicon One, noted that the G300 redirects packets instantaneously to avoid delays and enable GPUs to do more work, even during AI traffic bursts.For scale-across requirements, Cisco’s Silicon One P200 focuses on inter-data center connectivity over optical networks. The P200 provides 51.2 Tbps via 512 x 100 Gbps links and uses external high-bandwidth memory (HBM) for deep buffers, crucial for sustaining high-throughput transfers over wide-area links. Early P200 systems are accessible to select customers, with 28.8 Tbps switches featuring the P200 expected in Q3 2026, and 51.2 Tbps models due before the end of the year.
Liquid Cooling for High-Density Switches
Given the significant heat generated by AI-class processing, Cisco’s high-end switches incorporate liquid cooling. Cold plates are mounted directly to the chips and connected to the rack or data center’s liquid cooling system. This approach is applied to the G300, P200, and other heat-intensive components in Cisco’s Nexus 9000 and other high-end Silicon One switches. Kucharewski confirmed that Cisco offers a range of switches, including fully liquid-cooled, partially liquid-cooled, and air-cooled options, to meet diverse customer needs.
Competitive Landscape and Adoption
Broadcom and Nvidia are also developing similar high-speed networking ASICs. Broadcom's Tomahawk 6 targets scale-out and scale-up AI networks, delivering up to 102.4 Tbps on a single chip, with features for load balancing and congestion management. Nvidia’s Spectrum-6 is another 102.4 Tbps Ethernet networking ASIC, designed specifically for accelerating AI and cloud workloads within the Nvidia Spectrum-X Ethernet Platform.Initial adoption of these high-end switches is anticipated from hyperscalers, neocloud providers, and sovereign AI cloud operators, as they have the immediate need for such advanced capabilities. Dell’Oro Group's Boujelbene emphasised that enterprises, while not needing an immediate refresh, must integrate AI networking as a core design decision for serious GPU deployments, rather than treating it as an incremental upgrade.
