---
title: "The Boring Green Board That AI Data Centers Cannot Get Enough Of"
date: 2026-11-16
category: Chips & Compute
site: NeuroAI
canonical: https://neuroai.site/a/na-chips-ai-server-pcb-bottleneck
language: en
---

# The Boring Green Board That AI Data Centers Cannot Get Enough Of

> AI servers need boards with 20-plus layers and exotic laminates. Chinese PCB profits surged in 2025; Japanese suppliers have since pushed through 30% price rises.

Inside every AI server there is a slab of fibreglass and copper that nobody photographs. It carries no compute, holds no weights, and does not appear in any TOPS chart. It is also the component that turned from commodity into constraint in under two years, and the reason a group of Chinese manufacturers reported profit growth that looks like a software company's.

## What a printed circuit board actually does

A printed circuit board (PCB, 印制电路板) is the wiring layer of the physical world. Chips sit on it; the board routes power and signals between them through copper traces buried in resin-and-glass layers. An AI training server board can carry twenty-two layers or more, each one etched with impedance-controlled traces that must deliver a clean signal at speeds where a centimetre of extra routing is a measurable error.

Three things make AI hardware unusually demanding. The first is density: a GPU module with eight HBM stacks and a switch chip needs thousands of connections in a small area. The second is signal integrity: at 112 and 224 gigabit-per-second lane speeds, the dielectric material itself becomes a design parameter, because loss in the laminate eats the signal before it arrives. The third is heat and mechanical stress: large boards, heavy copper, and repeated thermal cycling.

That is why the interesting competition is not in the board, but one layer below it, in the copper-clad laminate (CCL, 覆铜板) — the resin, glass cloth, and copper foil sandwich that the board is etched from. Ordinary CCL is a commodity. The grades used in AI accelerators and 800G switches are not. Only a handful of suppliers worldwide can make ultra-low-loss material at volume, and qualification cycles are measured in quarters.

## The 2025 numbers

Chinese board makers had an extraordinary year, and their filings show exactly where the money came from.

- **Wus Printed Circuit (沪电股份)**, one of the main suppliers of high-layer-count AI server boards, reported 2025 revenue of RMB 18.945 billion (≈US$2.63 billion / HK$20.6 billion), up 42.00%, and net profit of RMB 3.822 billion (≈US$531 million / HK$4.16 billion), up 47.74%.

- Its data communications segment contributed RMB 14.656 billion (≈US$2.04 billion), about 77.4% of revenue. Boards with 22 or more layers brought in roughly RMB 5.2 billion (≈US$722 million), around 27.5% of group revenue, and PCB gross margin rose to 36.91%.

- **Victory Giant Technology (胜宏科技)** reported 2025 revenue of RMB 19.292 billion (≈US$2.68 billion / HK$21.0 billion), up 79.77%, and net profit of RMB 4.312 billion (≈US$599 million / HK$4.69 billion), up 273.52%.

- **Shennan Circuits (深南电路)** posted revenue of RMB 23.647 billion (≈US$3.28 billion / HK$25.7 billion), up 32.05%, with net profit of RMB 3.276 billion (≈US$455 million / HK$3.56 billion), up 74.47%.

- Upstream, laminate maker **Shengyi Technology (生益科技)** earned net profit of RMB 1.473 billion (≈US$205 million / HK$1.60 billion), up 343.76% on revenue of RMB 28.431 billion (≈US$3.95 billion / HK$30.9 billion).

Those margins are not normal for a board shop. Historically, board manufacturing meant single-digit net margins, brutal price negotiation, and chasing cheap labour. A 36.91% gross margin on the PCB business means the product stopped being a commodity for a while.

## Why prices went up, and who dared to raise them

The pricing power started at the top of the chain. According to reporting by Securities Times (证券时报), Japan's Resonac told customers it would raise prices on copper-clad laminates and bonding film by more than 30% effective 1 March 2026, and Mitsubishi Gas Chemical followed with increases of about 30% across high-end board materials from 1 April. When two of the largest suppliers of a critical input both move by a third within a month, the increase travels down the chain whether or not the board maker wants it to.

The mechanism is easy to miss: AI demand did not just increase the quantity of boards, it changed which boards are made. Capacity that would have produced conventional laminates got reallocated to high-speed grades, tightening supply in both. Copper foil followed the same pattern: the very low-profile foil needed for low-loss laminates is a small slice of total foil output, and it became the scarcest slice.

Capacity responded. An industry tally cited by Securities Times put announced high-end PCB plans above RMB 40 billion (≈US$5.6 billion / HK$43.5 billion) by March 2026, aimed at AI server boards, high-frequency material, IC substrates, and automotive boards. Victory Giant alone approved a 2026 investment plan of up to RMB 20 billion, with up to RMB 18 billion of that in fixed assets.

## The part that should make you cautious

Every one of those plans is a future supply curve. Board capacity arrives in roughly 18 to 24 months, in large discrete chunks, because plants are not built by halves. The 2023-2024 memory market showed what happens when a demand spike meets capacity that lands just after the peak: prices collapse and record margins become write-downs.

There is also a concentration problem. Wus derived about 82% of 2025 revenue from outside China, and its customer list is the top tier of networking and AI infrastructure companies. That is a strength while those customers expand and a single point of failure when one re-plans a rack architecture. Automotive boards, at RMB 3.045 billion (≈US$423 million) and 16.1% of revenue, are the intended second leg, but they are not yet a stabilizer.

And there is the material question. Chinese board makers are strong in fabrication and are gaining ground in laminates, but the highest-performance resin systems and specialty copper foil still concentrate in a few Japanese and American suppliers. The 30% increases demonstrated where the bargaining power sat at the start of 2026.

## Honest limitations

- **The price increases are supplier announcements reported by trade and financial media**, not an audited index. Announced list-price changes and realized transaction prices can differ substantially.

- **Company figures come from 2025 annual reports as summarized in Chinese financial press.** Segment splits such as "22-layer and above" depend on how each company classifies its products and are not standardized across the industry.

- **I have not verified the RMB 40 billion investment tally independently.** It is an industry estimate compiled by Securities Times, and announced plans are routinely revised downward.

- **Board count is a crude proxy for difficulty.** Layer count correlates with capability, but material grade, trace geometry, and back-drilling matter as much or more.

- **This is not investment analysis.** Nothing here accounts for valuation, and the margin figures describe a cyclical peak, not a steady state.

## What readers can do now

- **If you are specifying hardware, ask for the laminate stack-up, not the layer count.** The dielectric loss grade determines whether your link budget survives; the layer count is marketing.

- **If you follow the AI supply chain, watch laminate price announcements as an early indicator.** They move before board makers' margins do, in both directions.

- **If you are planning a purchase of AI infrastructure, expect board lead times to matter more than chip lead times in a tight quarter.** Systems integrators can substitute compute; they cannot substitute a qualified high-speed board without requalifying the whole design.

---

Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-chips-ai-server-pcb-bottleneck
Free to quote with attribution and a link to the original.
