Arista Networks: The Overlooked Network That Lets Thousands of AI Computers Work Together

Imagine thousands of computers working on the same enormous problem. Their processors are ready. Their calculations are moving along. Then one group needs information from another, and the next step cannot begin until that information arrives.

That pause is easy to overlook. The attention usually goes to the chips doing the calculations or the data centers housing them. But a large artificial intelligence system also needs a fast, carefully organized way for its computers to communicate. Without it, some of that expensive computing power can sit waiting.

This is the part of the AI infrastructure story that Arista Networks works on. The company reported third-quarter 2026 results on November 3, giving us a timely reason to look beyond the revenue number and understand what it actually builds. Arista makes networking equipment and software that connect computers, direct their traffic and help administrators manage large networks. Its role here is networking, rather than selling the accelerators or training the large models they run.

It is easy to overlook this layer because it is not what users see when they open an AI application. There is no chatbot window or clever answer to point to. The network is infrastructure in the plainest sense: it is the system that lets the other pieces exchange information. The distinction matters because faster processors alone do not solve every problem inside a distributed computing system.

The traffic system inside a data center

A data center is a facility filled with computing and storage equipment. It might run websites, business applications or cloud services. Some of the newest and largest facilities are built to handle artificial intelligence workloads.

The computers inside are often called servers. A server is a computer designed to provide computing resources, information or applications to other devices. Servers are commonly mounted in metal cabinets called racks. A rack keeps equipment organized and provides space for power, cooling and network connections.

Some servers contain specialized processors called accelerators. A graphics processing unit, or GPU, is one familiar example. GPUs are useful for AI because they can perform many mathematical operations in parallel. An accelerator is a chip, not a complete computer; it works as part of a server or other computing system.

The equipment that connects these machines is called a network switch. A switch receives information arriving over one connection and forwards it toward the appropriate outgoing connection. Ethernet switches direct frames using addressing information. Ethernet is a widely used standard for wired networking, and a frame is the basic unit of information carried across an Ethernet link.

Why AI makes the network matter

Large AI models can be trained across many GPUs, with each processor handling part of the calculations. The parts still have to be coordinated. GPUs exchange intermediate results and combine information before continuing.

NVIDIA's Collective Communications Library, or NCCL, provides communication operations for multiple GPUs in one computer and across multiple computers. Its tools include ways to share results, distribute information and send data between processors.

That communication takes time. If a network is congested, some processors may wait for information before continuing. Bandwidth describes how much information a connection can carry over time; latency is how long information takes to travel between two points. Both matter. A network can have high bandwidth and still suffer delays when traffic collects at a bottleneck.

AI networking therefore involves more than choosing a switch with a large capacity number. The network must distribute traffic across available paths and handle bursts of communication. The demands vary by workload and system design.

There is no single traffic pattern for every AI job. Some tasks require processors to exchange large amounts of information at once. Others send smaller messages more frequently. The network design has to account for those differences. A system that works well for one training arrangement may not be ideal for another, which is why customers pay attention to the combination of switches, software, cabling and computing hardware rather than one specification in isolation.

How leaf and spine networks connect the racks

A common data-center design is called leaf-and-spine. Leaf switches sit closest to the servers, often connecting equipment within individual racks. They are commonly installed at the top of a rack, which is why they are also called top-of-rack switches. Spine switches form the upper layer, connecting the leaf switches to one another.

In a standard two-tier design, every leaf connects to every spine. A server connected to one leaf can therefore reach a server on another leaf through a spine switch. This creates multiple paths through the network and makes it possible to expand by adding more leaf switches and servers.

The arrangement does not eliminate congestion. Capacity depends on the number and speed of links and how much traffic uses them. If demand exceeds the available capacity, the network is oversubscribed. Engineers can add links or switches as the system grows, but that expansion has to be planned. More hardware does not automatically make every path equally fast or every workload efficient.

Larger fabrics can be built from this leaf-and-spine design. Arista sells switching equipment for data-center networks.

Three ways AI networks grow

Networking companies use three terms for different ways of connecting computing resources.

Scale-up means connecting more accelerators within a tightly coordinated system, such as a server or rack, so they can work together closely.

Scale-out means adding more servers and racks, then connecting them so the larger group can share a workload.

Scale-across means connecting computing resources in separate data centers, sometimes in different locations.

A large AI installation may use all three. Communication within one rack is different from communication across a facility, while connections between data centers add distance and network topology to the challenge. Arista describes its networking strategy as spanning all three scales.

The box and the software inside it

Arista's products combine physical switches with software that controls and manages them. The company's central network operating system is called EOS, short for Extensible Operating System.

An operating system manages how a device functions. EOS handles network operations and gives administrators tools to configure and monitor Arista switches. Arista describes EOS as Linux-based and programmable, with a single software image across its switching family. Linux is widely used in servers and other computing environments. Programmable means administrators can automate network operations rather than make every change manually.

Arista also sells CloudVision, a platform for monitoring and managing networks across multiple devices. Network administrators need to know whether equipment is working, where traffic is moving and whether something unusual is happening. Management software can bring that information together and help apply changes consistently. The switch moves information, EOS runs the switch, and management software helps customers operate the larger network.

Arista does not make every chip or assemble every switch

Arista's networking equipment uses merchant silicon, chips designed and sold by specialist semiconductor companies for use in other manufacturers' products. Its 2025 annual report says Arista is primarily reliant on Broadcom for its switching chips. Broadcom supplies key components; Arista designs networking systems around them, develops its own software and sells the finished products and related services.

Arista also relies on third-party contract manufacturers. These companies assemble equipment to Arista's designs and specifications. The model lets Arista focus on system design and software, but component availability and manufacturing capacity affect its ability to deliver products.

That places Arista between several parts of the AI supply chain. Chip companies design processors and switching silicon; server makers build computers around processors; cloud operators run data centers. Arista supplies networking systems that connect the equipment and manage the traffic between it.

The October AI architecture announcement

On October 7, Arista described new Ethernet-based rack-scale architectures for AI networks. The announcement named AMD, Arm, Broadcom, d-Matrix, Meta, Microsoft and Qualcomm as partners in the work.

One architecture, called Etherlink SU-144, is described by Arista as supporting up to 144 accelerators in a rack-scale network. The company says a cross-rack version can extend to 1,024 XPUs through connections between racks. XPU is a broad industry term for a processor used to accelerate computing tasks.

The aim is to bring Ethernet into tightly connected AI systems, where processors communicate within and between racks. The traditional image of Ethernet is a cable connecting one computer to a switch. Rack-scale designs push the technology into a more integrated arrangement, with the network built around the way accelerators are installed and connected. Arista describes physical approaches including an orthogonal chassis and a cabled backplane. A backplane connects components inside a system; in the cabled design, cables provide those internal connections.

The 144- and 1,024-processor figures are announced architecture capabilities, not measured customer throughput or proof of widespread production deployment. Actual performance depends on the configuration, workload and operating conditions.

Not just AI: the campus network

Arista also sells networking equipment and software for enterprise and campus environments. At a university or large office complex, laptops, phones and other devices need to connect across buildings. Wireless access points provide Wi-Fi, while switches connect wired equipment.

Arista's campus offering includes wired switches, wireless networking and management tools. The company's campus materials describe a unified management approach for wired, wireless and wide-area networks. A wide-area network connects sites over larger geographic distances, such as offices in different cities. These requirements differ from those of a giant AI data center, but the job is familiar: connect many devices and keep information moving reliably. For a campus IT team, a dropped connection or an unresponsive wireless network can be just as visible to its users as a data-center delay is to an AI operator.

A business with a longer history

The business began as Arastra in 2004 and adopted the Arista Networks name in 2008. Its early focus was high-speed Ethernet networking for data centers and cloud computing, markets that already needed to connect growing numbers of computers. The current AI demand did not create Arista's networking expertise from scratch; the company had spent years building switches and network software for large computing environments.

The scale, and the customer concentration

Arista reported third-quarter 2026 revenue of $3.66 billion, up 58.5% from a year earlier. Product revenue was $3.19 billion and service revenue was $467.6 million. The quarter ended September 30. The company also gave fourth-quarter revenue guidance of $3.9 billion to $4.0 billion, a management forecast rather than a completed result. Arista says product revenue primarily comes from switching and routing products and related network applications. Service revenue mainly comes from support contracts, which customers typically buy with equipment and may later renew. That support can continue long after the original switch is installed.

For 2025, one customer represented 26% of Arista's revenue and another represented 16%. The annual report does not name them in that disclosure. Large orders matter to the business, and the timing of purchases by a small number of very large customers can affect results from one period to another.

The network behind the machines

The most visible parts of artificial intelligence are the models, the chips and the data centers. Arista works on a less visible piece: the network connecting the computers and helping them exchange information.

Networking does not replace the processors or the software running on them. It is one of the systems they depend on. As computing jobs are divided among more machines, the connections between those machines become part of getting the work done. Arista's business is built around providing that connective layer.