Cisco is escalating competition in AI data center networking with the launch of its Silicon One G300, a 102.4-terabit-per-second Ethernet switch chip geared to support AI clusters. The new silicon places Cisco squarely against rivals Broadcom and NVIDIA as enterprises and hyperscalers build networks to support tens of thousands of GPUs.
Like comparable offerings from Broadcom and NVIDIA, the chip delivers 102.4 Tbps of total throughput and supports up to 512 lanes of 200-gigabit-per-second connectivity. That capacity enables far denser AI fabrics, reducing the number of switches needed to connect hefty GPU clusters. It also helps simplify network configurations that have grown unwieldy as AI models scale.
Handling AI Traffic
Cisco touts the chip’s ability manages congestion and adapt to the unpredictable traffic patterns generated by distributed AI workloads. Training and inference jobs often produce sudden bursts of traffic that can stall GPUs if networks are overwhelmed.
The G300 introduces what Cisco calls an intelligent collective networking architecture. Instead of dividing packet buffers by port, the chip uses a fully shared buffer, allowing traffic to be absorbed wherever capacity is available.
In parallel to this buffer, a path-aware load balancing system monitors flows across the network and dynamically redirects traffic around emerging bottlenecks. Cisco claims this approach improves link utilization by roughly one-third and can shorten AI job completion times by more than a quarter in certain scenarios.
Regardless of the accuracy of that claim, the emphasis reflects a broader shift in AI networking away from static packet distribution toward more adaptive, system-wide coordination.
Extending the Lifespan of Switching Infrastructure
Programmability is another aspect of Cisco’s sales pitch. The G300 is built around the P4 programming language, allowing network admins to update functionality through software rather than replacing hardware. In theory, this flexibility extends the useful life of expensive switching infrastructure as protocols evolve and new AI networking standards emerge.
The chip will support new systems across Cisco’s data center portfolio, including updated Nexus 9000 and Cisco 8000 platforms. These systems drive high-density AI deployments and will be offered in both air-cooled and fully liquid-cooled configurations.
Echoing vendors across the data center industry, Cisco says liquid cooling can deliver substantial gains in power efficiency, a key consideration as AI clusters boost data center energy consumption.
Alongside the switches, Cisco is rolling out new optical modules to support higher link speeds. The company introduced 1.6-terabit pluggable optics for short-reach connections, as well as lower-power 800-gigabit linear pluggable optics that route signals to the switch silicon. Cisco says the combination can reduce overall switch power consumption by about 30 percent.
Cisco is not fully committing to co-packaged optics, a technology championed by some rivals that integrates optics directly into the switch package. While Cisco has demonstrated the technology in labs, executives say commercial deployment will depend on customer demand.
Competition for AI Network Market
With the G300, Cisco is signaling that it will compete across the full AI networking stack, from silicon to optics to software. Whether that integrated approach can meaningfully disrupt NVIDIA’s and Broadcom’s momentum remains to be seen, but the battle for AI network control is clearly intensifying.
