AI infrastructure is at a turning point. While multibillion-dollar buildouts show no sign of slowing down, the focus is shifting from rapid deployment-a "just build" approach-to energy efficiency, operating costs, and total return on invested capital.

The economic incentives are clear. As AI evolves beyond predictable training workloads, networks must now handle the collision of synchronous, high-bandwidth GPU communication with the unpredictable traffic patterns of agentic AI collaboration and on-demand inference. In response, network clusters are scaling to gigawatts of power spanning buildings and regions, introducing latency and cost challenges-and ultimately impacting capital expenses and operating profit. Networks must be reliable, intelligent, and adaptive to dynamic workloads while optimizing the shortest paths and running at higher traffic loads without dropping packets.

As the global AI boom enters its fourth year of exponential growth, there's also a new dimension: mixing old and new infrastructure equipment. Unlike traditional workloads, massively parallel AI computing is not well-suited to being run over "mix-and-match" network equipment. Between extensive interoperability testing and the risk of forcing new deployments to revert to legacy functionality, operators are shouldering new hidden costs.

Building on four decades of experience, Cisco is taking a multigenerational approach to AI networking that prioritizes network efficiency and lower total cost of ownership across AI deployments.

It all starts with the silicon. At Cisco, we believe that the right silicon strategy goes beyond "speeds and feeds": it optimizes the entire lifecycle of AI data center deployments.

Introducing Cisco Silicon One G300

Cisco Silicon One G300 is our latest generation in high-performance switching optimized for AI cluster networks. It delivers 102.4 terabits per second (Tbps) of Ethernet switching capacity contained within a single device.

The G300 is enabling the next wave of industry-standard Ethernet switching for AI networks. It supports 1.6T Ethernet ports with on-chip integrated 200 Gbps SerDes, developed in-house at Cisco, for low power consumption and high performance with longer reach. It has high radix scaling for up to 512 ports, enabling a "flatter" network with the ability to connect more compute resources closer to the edge of the network. This allows operators to connect a larger number of GPUs at the closest physical proximity, reducing compute latency, simplifying the network, and maximizing the efficiency of AI training and inference workloads.

A total approach to network efficiency and TCO

Cisco Silicon One G300 was engineered to optimize total cost of ownership (TCO) and drive higher profitability for the network cluster. We achieve this through two strategic pillars: Intelligent Collective Networking and Future-Proofed Infrastructure.

Intelligent Collective Networking

Cisco Silicon One G300 introduces a suite of features we call Intelligent Collective Networking, designed to deliver maximum performance and reliability for massive-scale GPU clusters:

1. Fully shared packet buffer: We have embedded a massive 252MB of packet buffer directly into the silicon, allowing a packet from any port to occupy any available space. This expanded capacity provides up to 2.5x increased burst absorption than industry alternatives, absorbing traffic bursts to prevent performance dips and keeping the network running at peak intensity.
2. Path-based load balancing: This feature directs traffic across all possible network paths and reacts in hardware to instantaneous congestion events or network faults at speeds 100,000x faster than software-based tuning. This path selection automation eliminates the need for manual optimization, enabling higher GPU density and ensuring the network remains optimized even as traffic patterns shift in real time.
3. Proactive network telemetry: Silicon One G300 provides a rich set of programmable session-level diagnostics. This visibility helps customers proactively identify and address network faults and optimization opportunities.

Intelligent Collective Networking delivers measurable benefits. In simulations, the larger packet buffer achieved a 33% increase in network throughput. This allows for higher GPU interconnect traffic without the need to increase network capacity, build larger networks, or add more switches-resulting in lower CapEx per deployed GPU.

Furthermore, simulations show a 28% reduction in job completion time (JCT) compared to advanced packet spray implementations, significantly increasing AI compute efficiency.

Finally, by integrating telemetry and visibility directly into the switch, there is minimal software intervention needed at runtime. The network can seamlessly handle different workloads without reconfiguration or reoptimization.

The result? Increased profitability for AI data centers at scale.

Future-proofed infrastructure

Onboarding new data center equipment is a significant financial and operational challenge. Historically, deploying new hardware with advanced features forced network operators into a difficult choice: revert new equipment to the "lowest common-denominator" functionality or face the cost of a forced phase-out of older equipment.

Silicon One breaks this cycle with Adaptive Packet Processing. G300 is just the latest Silicon One chip to use P4 programmability to enable operators to evolve infrastructure without replacing hardware.

This programmability has two major benefits:

1. One hardware design can be optimized for many roles. Silicon One G300 can be used in the back end, front end, and in a disaggregated scale-across application spanning data centers. The result is fewer unique hardware SKUs, simplified inventory management, and a lower total development spend.
2. New features can be rolled out post-deployment. This shifts the paradigm from buying new equipment every cycle to extending the existing infrastructure while ensuring consistency across mixed-generation rollouts.

In the fast-moving AI landscape, where requirements change constantly, programmable silicon future-proofs investments while maintaining flexibility.

Cisco Adaptive Packet Processing has been proven in real-world systems. Some examples of new features enabled in our previous-generation G200 switch silicon:

1. Ultra Ethernet Consortium (UEC): The 1.0 spec was announced in 2025 and has been implemented in G200, even though the silicon was designed years before the spec was available.
2. Traffic steering in AI networks: Customers have different approaches to how they optimize traffic steering algorithms. Adaptive Packet Processing allows for custom enhancements to our network path-selection intelligence.
3. Network visibility and serviceability: We enabled hardware-accelerated flow tracing for in-field debug of observed network behaviors.

Each feature addition was possible without requiring silicon spin.

Simply put: don't rip and replace when you can upgrade in place.

Cisco Silicon One pushes new boundaries for the AI era

With Cisco Silicon One G300, we're rethinking AI network operations and lowering the total cost of ownership for AI at scale. The G300 software developer kit (SDK) is available. We are actively engaged on new customer designs and are targeting 2H2026 for first system availability.

This is just the latest innovation in the Cisco Silicon One portfolio, which now powers over 60 Cisco systems. In October, we introduced the Cisco 8223, the world's first 51.2T deep-buffer fixed router, powered by a single Cisco Silicon One P200 processor shipping to hyperscalers. Cisco Silicon One P200 delivers unmatched routing and switching performance, programmability, security, and power efficiency.

Today, we are announcing expansion of the Cisco Silicon One P200 portfolio with new Cisco 8000 and Nexus 9000 fixed and modular systems with flexible OS support. Built on Cisco Silicon One architecture, Cisco Silicon One P200 solutions power new use cases of DCI, Universal Spine, and Core and Peer Routing to address broad market coverage.

Stay tuned as we continue to push the boundaries of what is possible in networking silicon.

Learn more about Cisco Silicon One

Additional resources 

• IP Is Better Than Ever with Integrated Performance Measurement
• Cisco Announces New Silicon One G300
• Networking for the Agentic Era: Cisco Unveils New Innovations in Scale and Simplicity