Cisco: New 102.4 Tbit/s Switching Chip Silicon One G300 for AI Data Centers

Contents

At the start of Cisco Live EMEA in Amsterdam, Cisco on Tuesday unveiled its new in-house developed network chip, Silicon One G300. The chip delivers 102.4 Terabits per second of Ethernet switching capacity and supports 1.6 Tbit/s Ethernet ports via integrated SerDes (Serializer/Deserializer) at 200 Gbit/s.

With up to 512 ports, it is intended to supply AI clusters with up to 128,000 GPUs – using only 750 switches instead of the previous 2500. This is intended to enable a "flatter" network, allowing operators to move more AI compute resources to the center of the network to reduce latency, simplify network construction, and increase the efficiency of AI training and inference tasks.

Intelligent Collective Networking

Cisco explicitly positions the G300 beyond pure speed records. Under the term "Intelligent Collective Networking," the manufacturer combines three functions: a fully shared packet buffer of 252 MByte, path-based load balancing that is intended to react 100,000 times faster than software-based optimization, and integrated network telemetry at the session level.

In simulations, Cisco claims to have achieved 33 percent higher network utilization and 28 percent shorter job completion times compared to conventional packet distribution. This generates more tokens per GPU hour – a factor relevant to profitability for operators of AI data centers.

The G300 utilizes P4 programmability ("Programming Protocol-Independent Packet Processors" – a programming language for controlling the packet forwarding functions of routers and switches), which Cisco markets as "Adaptive Packet Processing." This allows new network functions to be retrofitted even after commissioning, without having to replace hardware. Cisco points to its predecessor, the G200, which later implemented support for the Ultra Ethernet Consortium (UEC) 1.0 – even though the chip was designed years before the specification was finalized.

New Systems, Liquid Cooling, and Optics

Based on the G300, Cisco is launching new Nexus 9000 and Cisco 8000 systems, each with 64 OSFP cages for 1.6 Tbit/s optics. The systems are available in air-cooled and fully liquid-cooled variants. The latter are intended to improve energy efficiency by around 70 percent and deliver the bandwidth for which six systems were previously required. In addition, there are new 1.6 Tbit/s transceivers and Linear Package Optics (LPO), where signal processing takes place entirely within the switch chip. This halves the power consumption of the optics modules.

Furthermore, Cisco is expanding its P200 portfolio: the 51.2 Tbit/s routing chip announced in October 2025 will receive new fixed and modular systems for Cisco 8000 and Nexus 9000 – designed for DCI (Data Center Interconnect) and core routing, including 800G-ZR/ZR+ optics for long-haul connections.

Nexus One as a Unified Management Layer

Cisco is consolidating all data center network technologies – from Nexus and SONiC-based hyperscale systems to ACI – under the "Nexus One" platform. This offers on-premises, cloud, and API-driven operating models and integrates Splunk-based observability as well as the new "Agentic Ops" capabilities, where AI agents continuously monitor and optimize network status, security, and configurations.

According to Cisco, the SDK for the G300 is already available. The first systems are expected to be delivered in the second half of 2026. The expanded P200 systems and the new optics modules are also expected this year. With this chip, Cisco is directly competing against Broadcom's Tomahawk 6 and Nvidia's Spectrum-X Ethernet Photonics.

(mma)