The $8 Billion Network Factory Hiding Inside the AI Boom

The loudest objects in the artificial-intelligence boom are the chips. The quieter ones are the switches: flat metal boxes packed with ports, fans and specialized silicon, moving torrents of data among thousands of accelerators. Accton Technology makes those boxes. It also helps decide how they are designed, tests whether they will survive, manufactures them at scale and, increasingly, supplies the optics and cooling that keep them useful. The Hsinchu, Taiwan company has spent most of its 38 years being essential without becoming famous.

That low profile is partly the product. Accton is an original design manufacturer and a joint design manufacturer - ODM and JDM in the industry’s dense alphabet. A cloud operator, telecom supplier or network brand brings requirements. Accton’s engineers turn them into electrical, mechanical and software designs, navigate certification, assemble supply chains and produce finished systems that may wear somebody else’s badge. Its wholly owned Edgecore Networks subsidiary handles the more visible branded business, selling open wired and wireless products through distributors and integrators.

This arrangement made Accton a backstage company. AI has made the backstage unusually busy. Consolidated revenue reached NT$248.32 billion in 2025, about US$8.03 billion, up nearly 125 percent from the prior year. Network application equipment generated 62 percent of sales, while switches contributed 35 percent. By mid-2026, the company said first-half sales had risen another 60 percent year over year, although those monthly figures were unaudited.

The translator in the rack

Networking hardware lives at an awkward intersection. Chipmakers such as Broadcom create the merchant silicon. Open communities define common designs and interfaces. Network operating systems, including SONiC, give the hardware instructions. Customers bring workloads, budgets, thermal limits and deployment habits. None of those ingredients alone is a product that a technician can slide into a rack at 2 a.m.

Accton’s expertise is translation. It can combine a customer’s specification with a choice of switch silicon and CPU, engineer signal paths that behave at enormous speeds, design airflow or liquid loops, integrate firmware and software, pass regulatory tests, then repeat the result across high-volume lines. The company describes its durable advantages in less glamorous language: design for manufacturing, signal integrity, power integrity, mechanical design, component sourcing, reliability validation and on-time delivery. These are the disciplines that separate a working prototype from 10,000 working units.

The moat is not the metal box. It is the practiced journey from a demanding specification to a box that keeps working.YesPress analysis

Customers use that capability in two ways. Hyperscalers and major equipment vendors can specify customized systems without owning every factory step. Telecom carriers can deploy open cell-site gateways, aggregation routers and virtual optical-line terminals that separate hardware from software. Enterprises, campuses and smaller service providers can buy Edgecore systems through the channel. The common problem is vendor lock-in: a traditional supplier bundles hardware, operating system and support. Open networking lets the buyer mix layers, change software and scale with more standardized hardware.

Why 800G changes the factory

An AI training cluster is a group project conducted at machine speed. Accelerators constantly exchange partial results, and a slow or congested connection leaves expensive processors waiting. The network must move data quickly, distribute it evenly and recover predictably. Accton’s recent distributed disaggregated chassis design uses separate packet-forwarding and fabric systems to build a scheduled, lossless Ethernet network. The company says the architecture can grow to 32,000 GPUs at 800 gigabits per second.

Faster ports create a chain reaction. More electrical bandwidth raises power draw and heat. Copper traces become harder to manage, pulling optical links closer to the chip. Fans reach practical limits, making cold plates, liquid distribution and immersion more attractive. A vendor that once optimized a switch now has to think about the whole rack. That explains Accton’s spread into 800G optical interfaces, linear pluggable optics, silicon photonics, advanced fan control, open-loop liquid cooling and immersion tanks.

A white box with a longer résumé

Accton was early to open networking. In 2014, its AS5712-54X became the first network design accepted by the Open Compute Project. The following year it contributed open 40G and 100G designs and manufactured a commercial version of Facebook’s Wedge switch through Edgecore. In 2024 it returned with an 800G distributed fabric built around Broadcom’s Jericho3-AI and Ramon3 silicon.

The continuity matters. “White box” once implied a generic, low-cost device with the cleverness stripped out. Accton’s trajectory suggests a different reading. Standardized hardware can still require difficult engineering, and openness can become more valuable as the system grows. A buyer can preserve choice in the network operating system and silicon while relying on one group to integrate, validate and manufacture the stack.

Nexvec, introduced in 2025, is the logical extension. It combines disaggregated networking with composable compute and storage, then adds the Nous controller for planning, deployment and ongoing fabric management. The pitch is aimed at enterprises that like the economics and flexibility of open infrastructure but do not want to become their own integration lab. In effect, Accton is trying to package openness without packaging the homework.

“Every partnership must lead to shared success.”Jun Shi, writing on Accton’s open-infrastructure strategy

Where Accton wins - and where it can get squeezed

Accton sells engineering and execution wrapped around hardware. ODM programs bring product design and manufacturing revenue. JDM adds closer co-development. Edgecore brings branded hardware, software and professional services. The emerging “Manufacturing as a Service” language makes the factory network itself part of the offer, useful to customers managing volatile AI demand and geopolitical pressure on supply chains.

This is a competitive, price-sensitive market. Hardware generations turn quickly. Large customers have bargaining power and may split orders among suppliers. A rival can win the next switch program, an integrated vendor can justify a premium with software and support, or a hyperscaler can move more design work inside. Accton’s own annual reporting acknowledges margin pressure in standardized switches and routers.

Its response is to make each engagement harder to substitute: co-design with chip and software partners, add system-level integration, broaden into optics and cooling, and operate factories across a wider footprint. The company’s 2025 revenue mix hints that this is already more than a slide-deck strategy. Application equipment overtook conventional switches as the largest product line, while AI became the stated growth engine offsetting slower areas.

From the network to everything touching it

Accton now sits in an unusual market position. It is more specialized in networking than the broad electronics manufacturing giants, but less vertically closed than the major branded network vendors. It can offer a hyperscaler custom hardware, a telecom operator a disaggregated platform and an enterprise a packaged Edgecore or Nexvec solution. The binding idea is open infrastructure built with enough manufacturing discipline to behave like a finished product.

The experiments are widening. In 2026, Accton joined Taiwanese and Japanese partners in an all-photonics network demonstration involving NCHC, Chunghwa Telecom and NTT. It also signed a memorandum with Synaptics, Gallopwave and HYTech to explore onboard edge AI and navigation for drones. Neither project guarantees a large new business. Both show the direction: move useful compute closer to where data is generated, then provide the connectivity and integration around it.

There is a pleasing circularity here. Accton started in 1988 with twisted-pair hubs and network cards, the connective tissue of office computing. Today it is working on 102.4-terabit switches, 1.6-terabit optical links and cooling systems for AI racks. The units have changed spectacularly. The job has not. Someone still has to turn a fast-moving set of standards, chips and customer wishes into dependable machinery.

That is why Accton is useful as a map of the AI infrastructure market. The value is migrating outward from the accelerator. Networks, optics, power, thermals, software and supply chains all determine how much useful work a cluster produces. The company is betting that customers will prefer one experienced partner across more of those boundaries, so long as the architecture remains open. It is a practical bet, expressed in ports, fans, factory lines and shipping dates.