Industry explainer · U.S. view · Semiconductors · Switches · Data centers

AI Networking Industry Explained: The U.S. View

An AI cluster is only as fast as the network linking its accelerators. Training a large model means thousands of chips exchanging data constantly, so switches, switch chips, cables and optics have become a major part of AI spending. Networks work at three levels: scale-up connects accelerators inside a rack, scale-out connects racks across a cluster, and scale-across links whole data centers. U.S. companies lead the chain. Broadcom, which designs switch chips and custom AI accelerators, reported AI semiconductor revenue of $16.7 billion in the quarter to August 2, 2026, up 221%, and expects $21.7 billion the next quarter. Arista, which builds Ethernet switches, passed $3 billion in quarterly revenue for the first time in the second quarter of 2026, up 37.7%, and guided about $3.3 billion next. Nvidia sells its own networking alongside its accelerators. Speeds are moving to 1.6 terabits per second per port, and power use is becoming a design limit, pushing new approaches such as linear and co-packaged optics.

中文版（中国视角）：一文看懂高速连接与交换机产业链（铜缆、连接器） →

Broadcom: AI semiconductor revenue $16.7B in the quarter to Aug 2, 2026, up 221%
Arista: Q2 2026 revenue $3.04B, up 37.7%, its first quarter above $3B
Speed: Arista introduced 1.6 Tbps AI fabric platforms in 2026

IN ONE LINEAI clusters scale on the network; switch silicon is the core.

How the chain works

Switch chip designers supply the silicon and switch makers build the systems. Networks connect accelerators inside a rack (scale-up), across racks (scale-out) and between data centers (scale-across, dashed line), using copper over short distances and optics over longer ones.Swipe sideways on a phone to see the full chart.

Scale-up, scale-out and scale-across

	Scale-up	Scale-out	Scale-across
Connects	Accelerators in a rack	Racks in a cluster	Clusters in different buildings or sites
Distance	Shortest; copper common	Longer; optics common	Longest; optics
2026 signal	Arista offers liquid-cooled scale-up options	Arista Etherlink switches up to 100 Tbps	Arista outlines scale-across designs
What to watch	Open standards vs proprietary links	1.6T port adoption	Multi-site AI training

Why the network matters for AI

In AI training, accelerators split the work and constantly share results. If the network is slow, expensive chips sit idle. That makes switches and links a large part of AI cluster budgets, and the network design determines how big a cluster can usefully grow.

Arista's chief executive Jayshree Ullal described customers as now seeing "networking as the central nervous system for infrastructure." In its second-quarter 2026 release Arista outlined designs for scale-up, scale-out and scale-across networks, a multi-planar leaf-spine layout, and an open protocol, Multipath Reliable Connection, to keep large AI clusters busy.

Silicon: custom chips and switch chips

Broadcom sits at the center of the chain because it designs both switch chips and custom AI accelerators for cloud companies. In its fiscal third quarter ended August 2, 2026, Broadcom's revenue rose 86% to $29.6 billion, and AI semiconductor revenue rose 221% to $16.7 billion, which it attributed to very strong demand for custom AI accelerators and networking. It guided fourth-quarter AI semiconductor revenue to $21.7 billion, up 236%.

Nvidia sells networking as part of its AI platforms. Its Vera Rubin platform, in full production by August 2026, includes Spectrum-6 switches that support both pluggable and co-packaged optics.

Systems: Ethernet switches and power

Arista reported second-quarter 2026 revenue of $3.036 billion, up 37.7% from a year earlier and 12.1% from the first quarter, with a GAAP gross margin of 62.9%, and guided about $3.3 billion for the third quarter. Its 7060XE7 Etherlink switches offer up to 100 Tbps of system bandwidth, and it introduced 1.6 Tbps AI platforms, including liquid-cooled options.

Power is now a design limit. Arista says its linear pluggable optics cut interconnect power by about 60% compared with traditional pluggables. Every watt saved in the network is a watt available for accelerators in a power-constrained data center.

How the U.S. view differs from the China view

The Chinese version of this page looks at the components China makes for AI networks, such as copper cables, connectors and optical modules, and at Chinese switch makers building domestic AI clusters with locally available chips.

From the U.S. side the high-value pieces, switch silicon, custom accelerators and switch systems, are mostly American. The competition is among U.S. approaches: open Ethernet-based designs versus proprietary platforms bundled with accelerators.

2026: networking grows faster than ever

As AI clusters grew in 2026, networking spending grew with them. Broadcom's AI semiconductor revenue more than tripled year on year, and Arista passed $3 billion in a quarter for the first time.

Switch speeds moved to 1.6 Tbps, and vendors focused on power savings with linear and co-packaged optics.

Cluster size is increasingly limited by the network and by power, not only by the accelerators themselves.

What this shows

The network decides how large an AI cluster can usefully be.
Switch silicon and custom accelerators are concentrated in a few U.S. companies.
Power efficiency is becoming the main design goal for AI networks.

Common misreadings

MythAI spending is all about GPUs.

RealityBroadcom's AI revenue, from custom accelerators and networking, rose 221% to $16.7B in one quarter.

MythNetworking is a mature, slow-growth business.

RealityArista's Q2 2026 revenue rose 37.7% year on year.

MythOnly one vendor's network works for AI.

RealityArista sells Ethernet AI fabrics, while Nvidia bundles its own Spectrum-6 switches with Vera Rubin.

MythNetwork power use is negligible.

RealityArista says linear optics cut interconnect power by about 60%.

Cluster size, standards and power: scenarios

Scenario	Main effect	How it passes through
Clusters keep growing	Switch and optics demand	More ports and faster speeds
Open Ethernet standards win share	Switch vendors	More competition with proprietary platforms
Co-packaged optics adopted	Switch design	Lower power, new supply chain
Multi-site training spreads	Scale-across	Demand for long-distance links
Power limits tighten	Data center design	More focus on efficient networking
AI capex slows	Orders	Network spending falls with accelerators

Supply and demand by segment

Segment	Supply rigidity	Concentration	New demand	Policy role	What to watch
Switch silicon	High (design cycles)	Very high	Yes	High (export controls)	New chip launches
Switch systems	Medium	High	Yes	Medium	Revenue, guidance
Copper cables and connectors	Medium	Medium	Yes (scale-up)	Medium (tariffs)	Rack designs
Optics	High (lasers)	High	Yes	Medium	1.6T adoption

Supply rigidity: how long new capacity takes and whether money can speed it up. Concentration: whether output sits in a few countries or companies. The more rigid and concentrated, the more prone to shortages.

How to track it

Indicator	What it tells you	Where to find it
Broadcom AI semiconductor revenue	Custom accelerators and networking	Broadcom quarterly results, free
Arista revenue and guidance	Ethernet AI networking demand	Arista quarterly results, free
Hyperscaler capex	AI cluster budgets	U.S. view: AI compute
Optical modules	Links inside the network	U.S. view: optical modules

Sources

Last updated 2026-10-18.

This page explains the structure and supply-demand logic of an industry for educational purposes. It does not refer to any specific security and is not investment advice. Figures are dated and sourced; they may have changed since, so check the latest official data.