Cisco Unveils 100Tbps Switch for AI Workloads
Cisco has developed a new G300 switch capable of 100 terabits-per-second, designed to connect over 128,000 GPUs for large-scale AI operations. A Cisco Silicon One architect noted on the David Bombal Podcast that the network, not compute, is now the primary bottleneck in AI. The hardware features embedded "intelligent agents" for self-healing and traffic optimization, creating new, complex security surfaces for penetration testers to assess.
- The Cisco Silicon One G300 is actually a 102.4 Tbps switch chip that will power the company's next-generation Nexus 9000 and 8000 series systems, expected to ship in the second half of 2026. It is manufactured using TSMC's 3-nanometer process and can support up to 512 ports, enabling flatter, lower-latency network designs for AI clusters. - A key feature is "Intelligent Collective Networking," which uses a fully shared packet buffer and path-based load balancing to manage the bursty, unpredictable traffic of AI workloads. Cisco claims this technology can reduce AI job completion times by up to 28% and increase network utilization by 33%. - The G300 chip is programmable using the P4 language, allowing its packet-processing
