Broadcom Showcases Tomahawk 6 and Jericho 4 to Tackle AI Data Centre Networking Bottlenecks

As artificial intelligence models grow larger, building powerful processors is only part of the challenge. Thousands of accelerators must exchange enormous amounts of data quickly and reliably, making the networks connecting them a critical part of AI infrastructure.

Broadcom is addressing this challenge with its Tomahawk 6 and Jericho 4 Ethernet switching platforms, which it highlighted alongside other networking technologies at the Open Compute Project (OCP) Global Summit held in San Jose, California, from October 12–15, 2026. The company’s portfolio targets different parts of large-scale AI networking, from high-bandwidth connections inside data centres to links between geographically separated facilities.

The development reflects a broader shift in AI infrastructure: as computing clusters expand, network capacity, power efficiency and traffic management are becoming increasingly important to the performance of the entire system.

Why AI Data Centres Need Faster Networks

Training and running advanced AI models requires processors to exchange data continuously. During training, accelerators coordinate calculations and share intermediate results. During inference, computing resources may need to communicate across multiple servers to generate responses or complete complex tasks.

As the number of accelerators increases, the volume of network traffic can rise substantially. A cluster may contain tens of thousands of processors, with each relying on the network to deliver data without unnecessary delays or congestion.

A network that cannot keep pace can leave expensive processors waiting for information instead of performing useful calculations. This reduces the efficiency of the overall system, even when the computing hardware itself is highly capable.

Traditional data-centre networking therefore faces new demands. Operators need greater bandwidth, predictable latency, effective congestion management and better energy efficiency while keeping systems manageable at scale.

Broadcom’s Ethernet switching technologies are designed to address these requirements.

Tomahawk 6 Targets High-Bandwidth AI Clusters

Tomahawk 6 is Broadcom’s high-capacity Ethernet switching platform, designed for the demanding communication requirements of large AI and cloud data centres.

The platform provides 102.4 terabits per second of switching capacity in a single chip. This represents the aggregate capacity to move network traffic through the switch, rather than the speed available to one individual connection.

Its architecture supports high-speed Ethernet links and advanced traffic-management features intended to help large computing clusters communicate efficiently. Broadcom has described the platform as suitable for AI networks supporting both scale-up and scale-out architectures.

Scale-up networking connects processors and accelerators more closely within a system or rack, while scale-out networking links computing resources across multiple servers and racks. Both are important as AI infrastructure moves beyond smaller clusters toward systems containing tens of thousands or more accelerators.

Tomahawk 6 also incorporates Broadcom’s Cognitive Routing 2.0 capabilities, including telemetry, congestion control and failure detection. These functions help network operators monitor traffic conditions and respond to changing workloads.

The company has also developed a version called Tomahawk 6-Davisson that integrates co-packaged optics. In this design, optical components are placed close to the switching silicon, potentially improving power efficiency and signal reliability compared with some conventional optical interconnect arrangements.

These technologies are intended to help operators increase network capacity without allowing power consumption, congestion or connection failures to undermine AI performance.

Jericho 4 Extends AI Networking Beyond One Data Centre

While Tomahawk 6 focuses on high-capacity switching, Jericho 4 addresses a different challenge: connecting AI computing resources across larger networks and multiple data centres.

As AI infrastructure expands, organisations may distribute computing capacity across facilities because of limitations involving electricity supply, available land, cooling infrastructure and operational requirements. Connecting these sites creates additional networking challenges, particularly when the systems must exchange data with high throughput and predictable performance.

Broadcom announced Jericho 4 in 2025 as an Ethernet fabric-routing platform designed to interconnect more than one million XPUs across multiple data centres.

The platform supports high-capacity HyperPort connections and congestion-management features intended to improve the movement of data across distributed computing infrastructure. Broadcom has specified support for connections extending beyond 100 kilometres, making the platform relevant to certain geographically distributed AI network designs.

Jericho 4 also incorporates deep buffering and Media Access Control Security (MACsec), which provides link-layer encryption for network traffic. These features are intended to help manage congestion and protect data as it travels between network components.

The distinction between the two products matters. Tomahawk 6 provides high-bandwidth switching for AI clusters, while Jericho 4 is designed to extend Ethernet fabrics across larger, distributed environments. Together, they form parts of a broader networking portfolio rather than two interchangeable products.

Why Ethernet Matters in the AI Infrastructure Race

Ethernet is already widely used in enterprise networks and data centres. Its established ecosystem, familiar management tools and support for interoperable equipment make it an attractive foundation for organisations seeking to expand AI infrastructure.

However, the communication patterns generated by AI workloads can be particularly demanding. Large training jobs may involve frequent exchanges of information among accelerators, while distributed inference systems must coordinate computing resources across servers and facilities.

Broadcom is working to extend Ethernet’s capabilities for these environments through higher-capacity switches, networking interfaces, optical technologies and traffic-management features.

The company’s participation in the OCP ecosystem also fits a wider industry effort to develop more open and modular data-centre infrastructure. The Open Compute Project brings together technology companies and infrastructure operators to collaborate on hardware designs and specifications.

At the 2026 summit, Broadcom highlighted technologies including Tomahawk 6, Tomahawk Ultra, Jericho 4, AI Ethernet network interface cards and co-packaged optics. These components address different requirements across the networking stack, from connecting individual systems to supporting large-scale AI clusters.

An open networking approach does not automatically guarantee lower costs or better performance in every deployment. Actual results depend on equipment selection, software, network topology and how effectively the components work together.

The Efficiency Challenge Behind Larger AI Systems

Increasing bandwidth is only one part of the challenge. AI data centres also face growing demands for electricity, cooling and physical space.

Networking hardware consumes power, and the equipment required to move data between processors can become a significant part of a large facility’s infrastructure. Optical connections, switching architectures and traffic-management techniques can therefore influence both performance and operating costs.

Co-packaged optics is one approach to improving this balance. By bringing optical components closer to switching silicon, the design aims to reduce some of the electrical and optical constraints associated with moving data at increasingly high speeds.

The practical benefits depend on the system design and deployment conditions. Integration, serviceability, thermal management and the reliability of optical components remain important engineering considerations.

Similarly, high switching capacity does not mean every AI application will become faster by the same amount. Improvements depend on whether network communication is actually limiting the workload, as well as on the performance of processors, memory, storage and software.

What Broadcom’s Networking Push Means for AI Development

Broadcom’s showcase illustrates how the AI hardware market is expanding beyond processors and accelerators. As companies build larger computing clusters, they also need networking systems capable of moving data efficiently across those resources.

Tomahawk 6 targets high-capacity Ethernet switching, while Jericho 4 addresses connectivity across distributed AI infrastructure. Their roles reflect the growing importance of treating networking as a core part of AI system design rather than a secondary component.

The technologies do not remove every bottleneck associated with building large AI facilities. Power availability, cooling, computing hardware, software optimisation and the economics of operating data centres will continue to shape how quickly infrastructure can expand.

Nevertheless, improvements in networking can help operators use computing resources more effectively and support larger deployments. As frontier AI models demand increasingly complex computing environments, the ability to connect processors reliably and efficiently is becoming as important as the processors themselves.

FAQs

1. What are Broadcom Tomahawk 6 and Jericho 4?

Tomahawk 6 is a high-capacity Ethernet switching platform designed for large AI and cloud clusters. Jericho 4 is a fabric-routing platform designed to connect computing resources across multiple data centres.

2. How much switching capacity does Tomahawk 6 provide?

Broadcom specifies 102.4 terabits per second of aggregate switching capacity for Tomahawk 6. This is the total switching capacity of the chip, not the speed of a single network connection.

3. What is Jericho 4 used for?

Jericho 4 is designed to extend Ethernet networking across distributed AI infrastructure. Broadcom has described it as capable of supporting networks connecting more than one million XPUs across multiple data centres.

4. What is co-packaged optics in AI networking?

Co-packaged optics integrates optical components close to switching silicon. The approach aims to improve high-speed data transmission efficiency and reliability as network bandwidth requirements increase.

5. Do Broadcom’s networking chips make AI models faster?

They can help reduce networking bottlenecks in suitable systems, but performance improvements depend on the workload, network design, computing hardware and software. Faster switches alone do not guarantee faster AI training or inference.