Cisco Silicon One G300: 102.4T AI switches due in 2026

On 10 February 2026, at Cisco Live EMEA in Amsterdam, Cisco announced Silicon One G300, a 102.4 Tbps Ethernet switch chip for AI back-end networks. It uses 200 Gbps SerDes, supports 1.6 Tbps Ethernet ports and up to 512 ports, and has a fully shared packet buffer that heise reported at 252 MB. Cisco positions it against Broadcom's Tomahawk 6 and NVIDIA's Spectrum-X Ethernet Photonics.

Cisco groups the chip's traffic features under the name Intelligent Collective Networking: path-based load balancing, session-level telemetry and P4 programmability. In Cisco's simulations against non-optimised path selection, these gave 33 per cent higher network utilisation and 28 per cent shorter job completion times. These are vendor figures and have not been independently tested.

Systems and optics

The chip goes into new fixed and modular Nexus 9000 and Cisco 8000 systems, each with 64 OSFP cages for 1.6T optics and offered in air-cooled and fully liquid-cooled versions. Cisco says the liquid-cooled systems, together with the new optics, improve energy efficiency by nearly 70 per cent, and that one system replaces six of the previous generation. The optics announced alongside include 1.6T OSFP modules for switch-to-NIC links and 800G linear pluggable optics (LPO), which Cisco says halve module power compared with retimed modules and cut total switch power by around 30 per cent on LPO-capable systems. Cisco also expanded systems based on its 51.2T P200 routing chip, first shown in October 2025, with 800G ZR/ZR+ coherent pluggables for data-centre interconnect.

The software development kit was available at launch. Cisco said the G300, the systems and the optics will ship in 2026, and heise reported the first systems for the second half of the year; the release carries the usual note that timelines may change. Among the customers quoted in the announcement was du, the UAE operator. Two days later, on its fiscal second-quarter call, Cisco said it had shipped its one-millionth Silicon One chip and booked $2.1 billion of AI infrastructure orders from hyperscalers in the quarter.

What it means for buyers

For most enterprises in the region, G300 systems matter as a 2027 planning item rather than a current purchase: first shipments target hyperscalers, neoclouds and sovereign AI projects. Buyers planning GPU clusters should ask for confirmed ship dates per SKU, decide early between air and liquid cooling because it affects the facility design, and specify optics by type (1.6T OSFP, 800G LPO or retimed) since LPO depends on host-side support. Where a project needs capacity in 2026, 51.2T systems with 800G optics remain the available option. InTechnologies can quote Cisco Nexus systems and optics with current lead times.