Explain It in 30 Seconds
A computer can finish only the work for which it has the necessary data. When many computers or accelerators cooperate, exchanging that data becomes an important part of the job.
Network interfaces connect individual systems to a network. Switches move traffic among connected devices. Larger designs organize those components into networks with enough capacity, suitable behavior and a practical way to handle failures.
Broadcom supplies silicon and supporting capabilities for these jobs. Equipment manufacturers and operators still need to turn components into complete, configured and supported networks. Tomahawk and Jericho are product families inside that infrastructure portfolio, rather than names for one universal finished data center. [S8][S10]
Where It Sits in the Company
Ethernet sits between compute resources and the wider systems with which they exchange information. It can serve AI infrastructure as well as other data-center and communications requirements.
Broadcom calls part of this approach merchant networking: component technology can be used by organizations building their own equipment and systems. That differs from requiring every customer to adopt one vertically integrated computing platform.
The company's custom-compute work provides another connection. A customer designing processors also needs to connect them. The OpenAI collaboration explicitly links accelerator programs and networking, showing how Broadcom can participate in more than one layer of the same infrastructure project. [S7][S13]
The Problem a Network Must Solve
A network must move useful information among endpoints, but peak link speed is only one part of success. Traffic can arrive unevenly. Several senders may want to reach one receiver. A busy path may delay one job while another path is less occupied.
For distributed computing, waiting has an economic cost. Expensive processing resources can remain underused while required information is in transit. The relevant outcome is therefore not only how fast a port can transmit but how efficiently the whole workload progresses.
Failures create another requirement. A production network needs to detect problems, preserve useful connectivity where possible and help operators diagnose what happened. A large fabric with impressive nominal capacity can still be difficult to use if its behavior is unpredictable.
How the Pieces Work Together
The network interface gives a host an entry point. It must fit the server's hardware, software and communication needs. From there, switch equipment directs traffic toward the intended destination.
The switch chip is an important ingredient, but the completed switch also requires board design, ports, power, cooling, firmware and control software. The operator then chooses a topology, establishes policies and integrates monitoring.
Links connect those elements. Depending on distance, speed and design requirements, a system can use different electrical or optical approaches. These choices affect power, reach, cost and serviceability rather than adding capacity without tradeoffs.
Broadcom's portfolio spans several of these functions. The architecture still needs to be evaluated as a connected system, not as the sum of independent component headlines. [S7][S9][S48]
Scale-Up, Scale-Out and Connections Across Facilities
Different communication problems appear at different scales. Closely coordinated processing resources may require tightly integrated connections. Larger groups need a fabric that connects many systems. Connecting facilities adds further questions about distance, reliability and operating constraints.
Those categories should not be collapsed into one claim that a faster switch makes every network equally suitable. The intended traffic and system boundary determine which characteristics matter.
Broadcom positions Tomahawk 6 for high-capacity AI network applications and Jericho4 for distributed computing across data centers. Those are product-positioning statements. They do not establish that every customer has deployed the maximum advertised configuration. [S8][S10]
Why Congestion Management Matters
If several streams compete for a constrained part of the network, adding more compute at the edge may not solve the bottleneck. The network must manage how traffic shares available paths and how endpoints respond when demand exceeds immediate capacity.
The problem resembles coordinating access to a shared transport system, but the operating details are digital and highly time-sensitive. Buffering, routing choices, feedback and failure detection influence what the applications experience.
Broadcom's Tomahawk announcements describe features intended to address these challenges. A useful evaluation still asks how the features behave with the customer's traffic, topology and software. Stated design intent and independently observed workload performance are different kinds of evidence. [S8][S9]
Before and After a Network Upgrade
Before an upgrade, a customer may be constrained by link capacity, the number of network stages, operational complexity or the behavior of particular workloads. The constraint should be identified rather than assumed from the age of the equipment.
After an upgrade, the customer expects more useful communication capacity or better operation. However, surrounding interfaces, optics, software or endpoint behavior can become the next bottleneck. Replacing one component is not the same as rebuilding a balanced system.
The commercial consequence is that purchasing decisions involve a transition plan. Equipment must be validated, integrated and operated alongside existing resources. A roadmap may promise future capacity while the installed network continues to serve current work.
Who Uses It and Who Pays
Equipment designers use Broadcom components as part of their systems. Infrastructure operators use the resulting networks. Software and operations teams depend on the network's behavior even if they never purchase the chip directly.
That makes the customer chain important. The organization selecting an architecture can differ from the entity placing a component order. Public discussion of a famous operator does not automatically identify the contractual purchaser or reveal the price of a design win.
The end users of a cloud service sit farther downstream. They pay for an application or computing service, while the provider bears the responsibility for assembling the infrastructure that delivers it.
What the Portfolio Depends On
The portfolio depends on semiconductor implementation, packaging, supported interfaces and integration into equipment. It also depends on software that can configure and observe the network.
The physical interconnect adds its own requirements. A switch design needs usable links to other equipment. As communications rates rise, power and signal behavior can influence architecture and packaging choices. Broadcom's co-packaged-optics work addresses one approach to that problem, with tradeoffs that need assessment beyond a bandwidth label. [S11][S48]
At deployment scale, the operator needs power, cooling, staff capability, spare equipment and procedures for change. These are part of the economic reality of a network even though they are not all supplied by the switch-chip vendor.
What Depends on the Network
Distributed applications depend on timely, reliable communication. Storage services, virtualized infrastructure and AI workloads can each impose different patterns, so one workload's favorable result should not be generalized automatically to another.
Operators also depend on being able to understand network behavior. If a problem appears only under a particular traffic pattern, the ability to observe and reproduce it can be as important as headline capacity.
The dependency therefore includes both performance and operability. A network that supports more devices but becomes unmanageable does not automatically improve the service delivered to customers.
Why It Matters to Broadcom
Networking lets Broadcom participate in infrastructure growth even when the compute architecture varies. More powerful processors can increase the demand for communication, while the company's merchant approach can serve multiple system builders.
This does not mean every increase in compute spending produces proportional Broadcom revenue. Customer architecture, component selection, timing and supply execution determine the actual outcome. The opportunity is a role in solving a recurring infrastructure problem, rather than a guaranteed claim on all future AI expenditure. [S7]
Key Milestones
Broadcom introduced Tomahawk 6 in June 2025 and announced production-volume shipments in March 2026. The later milestone provides a clearer statement of commercial maturity than an introduction alone. The disclosed switch capacity is 102.4 terabits per second; it is not a measurement of every network built around the chip. [S9][S8]
Jericho4's August 2025 announcement addressed connectivity across facilities. Broadcom's October 2025 Davisson announcement connected the switching roadmap to co-packaged optics, while retaining early-access qualifications in its availability discussion. [S10][S11]
Together these events show a portfolio evolving across capacity, topology and physical connectivity. Introduction, sampling and production announcements establish different stages; they do not establish widespread deployment of every advertised system.
Common Misconceptions
“A switch chip is a complete switch.”
It is a major component. A usable switch also requires the surrounding physical design and operating software, followed by validation in the intended network.
“Faster ports guarantee a faster application.”
Only if the relevant constraint is improved. Endpoints, traffic patterns, other network stages and application behavior can limit the result.
“Ethernet makes all vendor implementations equivalent.”
A shared standard supports an ecosystem, but architecture, integration, features and software still differ. Interoperability also needs testing in the actual configuration.
Common Questions
Are Tomahawk and Jericho the same thing?
Does Broadcom build the whole AI service?
Its networking components contribute to the infrastructure. The operator still needs compute, software, facilities and a service that users can access.
Why are optical links discussed with Ethernet switches?
The switch needs physical connections to other equipment. Optical technology can address reach and bandwidth requirements, while introducing its own design, cost and support considerations. [S48]
Evidence
- Broadcom's infrastructure strategy establishes the merchant networking and custom-compute relationship. [S7]
- Dated Tomahawk announcements distinguish introduction from production-volume status. [S9][S8]
- The Jericho4 announcement establishes its intended distributed-network role. [S10]
- The optical documentation explains the connection between switching and physical-link integration. [S11][S48]
Connected Entities
The related compute layer is explained in Broadcom Custom AI Accelerators, while Broadcom Optical Connectivity explains the physical-link layer. The shared semiconductor industry page provides the wider manufacturing and supplier context.