Cisco to launch Silicon One G300 chip for AI data centers

Cisco announced the upcoming release of its Silicon One G300 switch chip, designed to improve data transfer speeds in large data centers, with sales expected in the second half of 2026.

The chip, manufactured using Taiwan Semiconductor Manufacturing Co.’s 3-nanometer process, aims to facilitate communication between AI training and inference systems across hundreds of thousands of links.

Cisco said the chip incorporates features to manage data traffic spikes and reroute data automatically, potentially increasing AI-related processing speeds by up to 28%.

Cisco’s new offering emphasizes network efficiency and resilience in supporting AI workloads.

🔗 Source: Reuters

🧠 Food for thought

Implications, context, and why it matters.

Cisco’s new chip is about more than just speed

The G300 centers on P4 programmability (a way to customize how network devices handle traffic). Customers can add functions through software, which may keep the hardware useful as features and standards change 1.
Cisco ties the speed jump to a “collective networking engine” with a fully shared packet buffer plus a path-based load balancer. The company puts link utilization at 33% higher and training time up to 28% shorter than packet spraying (a technique that spreads traffic across many paths) 1.
The chip arrives with new N9000 and 8000 series switches plus 800 Gbps Linear Pluggable Optics (LPO). Cisco puts LPO pluggable power use at about 50% lower, and about a 30% switch power drop when paired with N9000 or Cisco 8000 systems 1.

This signals a strategic push deeper into hyperscale AI networking

Cisco wants stronger footing with hyperscalers (the biggest cloud and internet companies that run massive data centers). IDC figures cited by Data Center Knowledge place Cisco behind rivals such as Arista Networks and ODMs (original design manufacturers that build white-box networking gear) 2.
In fiscal Q1 2026, Cisco booked $1.3 billion in AI infrastructure orders from hyperscaler customers, as described by The Motley Fool 3.
The G300 fits a wider effort to push Ethernet as a higher-performance option to InfiniBand for AI clusters. Work includes the Ultra Ethernet Consortium (UEC) (an industry group working on faster, lower-latency Ethernet for AI data centers), while TrendForce ties momentum to Ethernet’s open ecosystem and lower costs 4.
Analysts connect Cisco’s unified Silicon One approach to a longer-term shift in enterprise IT and networking teams toward more hyperscale-style AI networking 5.