Cisco Silicon One G300 Delivers 102.4 Tbps for AI-Scale Data Centers

Cisco announced the Silicon One G300 at Cisco Live EMEA in Amsterdam on February 10, 2026. The new chip delivers 102.4 Tbps of switching capacity across 64 ports at 1.6 Tbps per port, positioning it squarely against Broadcom's Tomahawk 6 series for the AI data center networking market. The G300 is designed for gigawatt-scale AI clusters handling training, inference, and agentic workloads simultaneously.

The most striking practical claim is in GPU fabric density. Cisco states that a G300-based deployment can interconnect up to 128,000 GPUs using just 750 switches, compared to roughly 2,500 switches required with the prior generation. That reduction matters at hyperscale: fewer switches mean fewer failure domains, less cabling, lower power draw from networking gear, and reduced operational overhead.

Intelligent Collective Networking

The G300 introduces what Cisco calls Intelligent Collective Networking, a combination of a fully shared packet buffer, path-based load balancing, and proactive network telemetry. AI training traffic is inherently bursty—large all-reduce operations can saturate links momentarily before going idle—and existing fixed-buffer switch ASICs can drop packets during these bursts, stalling entire jobs waiting for retransmits. Cisco's shared buffer architecture is designed to absorb those bursts across the full 102.4 Tbps capacity. In internal testing with simulated workloads, Cisco reports 33% higher network utilization and a 28% reduction in AI job completion time compared to non-optimized path selection.

The chip supports 1.6T OSFP optics for switch-to-NIC links, alongside 800G, 400G, and 200G connectivity options. Cisco also announced 800G Linear Pluggable Optics that reduce optical module power consumption by 50%, contributing to an overall 30% reduction in switch power draw compared to the previous generation. The architecture is described as highly programmable, allowing post-deployment upgrades to support emerging protocols and network roles.

Liquid-Cooled Systems and Availability

New Cisco Nexus 9000 and Cisco 8000 systems powered by the G300 are being offered in 100% liquid-cooled configurations. Cisco claims the combination of liquid cooling and the new optics enables customers to achieve nearly 70% improvement in energy efficiency. A single G300-powered system delivers bandwidth equivalent to six prior-generation systems, which means a meaningful reduction in rack space and cooling infrastructure for equivalent networking capacity.

The G300 silicon, the new powered systems, and the associated optics are all scheduled to ship in the second half of 2026. The announcement places Cisco in direct competition with Broadcom and NVIDIA for the AI fabric market at a moment when hyperscalers, neoclouds, and sovereign cloud operators are all scaling up GPU clusters aggressively.