
⚡ Silicon & Steel · The Board
PCB Story 2 of 4: the material

These are not bottles of milk, but bobbins of glass yarn.
Ever wondered what NVIDIA, Microsoft, Google, Amazon and Apple are all reported to be queuing for right now?
This time not chips. not memory. A cloth: a woven glass fabric, made mostly by one company in Japan. Story 1 showed that glass cloth is the slowest input to fix in the 2026 circuit board market, even though it is the smallest share of the cost.
This story explains why a fabric decides what a computer can do. Eight minutes. Here is the whole board.
📖 In plain words: the terms you need
Laminate (CCL). The raw sheet a board is cut from: glass cloth, hardened resin, copper foil.
Prepreg. Glass cloth soaked in resin, not yet hardened. The glue between a board's layers.
Dk (dielectric constant). How much a material slows down a signal passing near it.
Df (dissipation factor). How much of a signal's energy the material soaks up and turns into heat. Low Df is the expensive kind. Where to check it: laminate makers publish Dk and Df on each product's data sheet, for example Panasonic's Megtron series.
Low-Dk glass. Glass cloth made from a special recipe that slows and absorbs signals less. Used in fast circuit boards. NE glass is one type.
T-glass. A different special glass that barely expands when heated. Used in the small boards chips are mounted on, where warping is the enemy.
Package substrate. The small board a chip is mounted on, which then sits on the main circuit board.
Fibre weave effect. Timing errors caused when a copper line runs over glass yarn and its partner runs over resin.


♟️ The chess move
🎯 THE WHAT
At AI speeds, the board material decides how far a signal can travel, which helps decide where chips can sit. The best material depends on specialty glass cloth, and the dominant supplier is choosing customers rather than raising prices.
The playbook: When you are the only qualified source of an input the richest buyers cannot design out, you do not need to auction it. You allocate it, lock in the relationships, and let everyone else wait.

🔍 The breakdown
⚙️ THE HOW
1. A circuit board is not really a board
Start with sand, which becomes glass. The glass is melted and pulled into filaments thinner than a hair, twisted into yarn, and woven on a loom into fabric, like any other cloth. That fabric is soaked in resin, baked stiff and clad in copper. Story 1 has the step-by-step flowchart.
A toaster's board might have two layers. A car's engine controller, a handful. AI server boards can have dozens: Würth Elektronik puts their layer counts at three to five times conventional designs. Prismark reported that boards with 18 or more layers grew 40% in value in 2024, the fastest-growing part of the market.
Underneath the most advanced machine we build, there is a textile.
2. The material decides how far apart chips can sit
At the speeds modern chips run, the signal is carried by an electromagnetic field that spreads from the copper line into the glass and resin around it. That material absorbs some of the energy as it goes. That is what Df measures.
Ordinary FR-4 has a Df of roughly 0.02. A widely used low-loss material, Panasonic's Megtron 6, is about 0.002 at 1 GHz and about 0.004 at higher frequencies. So the expensive material absorbs roughly five to ten times less, depending on frequency.

Cheap material, short distances. Expensive material, longer ones. A property of baked plastic and woven glass helps decide the physical layout of a computer.
3. The loom problem
Glass slows signals more than resin does. In woven cloth, the glass is bunched into yarns with resin-rich gaps between them. Fast signals travel in matched pairs, and if one line sits over glass and its partner over resin, they arrive at slightly different times. Engineers call this the fibre weave effect (explainer).
One standard fix sounds like a joke: rotate the whole design on the manufacturing panel by about ten degrees, so every line crosses the weave at an angle. The IPC, now the Global Electronics Association, has published technical work on this. It reduces the problem rather than eliminating it.

The other fix is better glass: flatter, spread yarns and a different glass recipe. That is where the shortage lives.
4. Two special glasses, two different jobs
Low-Dk glass is for the circuit board itself. It slows and absorbs signals less, so AI servers and network switches need it. T-glass is for the package substrate under a chip. It barely expands when hot, so a big, hot AI package stays flat and its connections do not crack.

Nittobo is reported to hold about 90% of T-glass and 60% to 70% of the leading low-Dk segment. As AI packages grow, they soak up T-glass. Goldman Sachs has warned, as reported by TrendForce, that makers of BT substrates, a different package used widely in memory and smaller chips, could face double-digit T-glass shortfalls.
5. One weaver, and how it is behaving
Nittobo is expanding. Its current medium-term plan budgets about ¥120 billion of capital investment over the four years to March 2028, up from about ¥80 billion in the 2024 plan. New lines are going into Fukushima and Taiwan. New facilities arrive from 2027, and it aims for three times its March 2026 capacity by the year ending March 2029.

Challengers exist. Taiwan Glass was certified for low-Dk cloth, reportedly only the third company to get there. Fulltech has been certified too. Asahi Kasei announced in April it would enter AI glass cloth. Nan Ya Plastics says it will weave about a fifth of Nittobo's specialty fabric by 2027.
Certification is slow, because laminate makers, board makers and their customers each have to prove a new cloth performs identically.
The most revealing fact is about price. Nittobo raised T-glass prices, then said in August it had no further increases planned, unless energy costs force a rethink. Instead, by all reports, it is deciding who gets supplied.

The scarcity at the top of the chain is being managed by rationing, not auctioning. That is why, as Story 1 showed, ordinary buyers feel it as waiting and quotas, not only as price.
6. Where it goes next: quartz
The newest ultra-low-loss laminate grades, sometimes called M9, need glass that absorbs even less. The industry's answer is quartz cloth, woven from nearly pure silica. Fulltech began certifying M9 quartz cloth this summer.
High-purity silica is the same family of material used for the crucibles that grow silicon wafers. It is now heading into the boards those chips sit on. A scarcer material, not a more common one.
"The most advanced machine on earth is built on woven cloth, and the supplier of the best cloth is choosing its customers."

📊 The numbers
As of September 2026.
Measure | Value |
|---|---|
Nittobo share of T-glass (reported) | ~90% |
Nittobo share of leading low-Dk glass (reported) | ~60–70% |
Nittobo capacity target | 3× March 2026 level, by FY ending March 2029 |
Nittobo planned capital investment, FY2024–27 | ~¥120bn, raised from ~¥80bn |
AI server board layer counts vs conventional | 3–5× (Würth Elektronik) |
18+ layer boards, 2024 growth (Prismark) | +40% value, +59% area |
FR-4 dissipation factor | ~0.02 |
Megtron 6 dissipation factor | ~0.002 (1 GHz) to ~0.004 |
Typical fibre weave rotation | ~10° |
Glass cloth share of laminate cost (TrendForce) | ~19% |

💼 The investor read
Unlike many chokepoints in the chip supply chain, most of this part of the board chain is publicly listed.
Nittobo (TSE: 3110) holds the most concentrated position in the series. It has chosen allocation over price, and the tripling that relieves the shortage will also, eventually, remove the scarcity.
Nan Ya Plastics (TPE: 1303) gains a role as Nittobo's weaving partner, inside a large, diversified group, so the exposure is diluted.
Taiwan Glass (TPE: 1802) and Asahi Kasei (TSE: 3407) are the challengers. Watch customer certification, not capacity announcements. An unqualified plant produces inventory, not revenue.
Watch the substrate side separately. T-glass shortfalls hit chip packages, not circuit boards directly. Reporting that mixes the two misreads who is exposed.
The investor read is analysis of industry structure, not investment advice. Nothing here is a recommendation to buy or sell any security. I hold no positions in the companies covered and take no payment from them. Do your own work.

🧾 What I'm assuming
Nittobo's shares are reported estimates, not company disclosures. The ordering is reliable; the exact percentages are not.
The allocation list is reported by trade press, not confirmed by the companies.
Df depends on frequency, supplier and test method. Five to ten times is the right order of magnitude; newer grades go lower.
Loss is not only the material. Copper roughness and trace geometry matter too.
Fibre weave rotation is one fix among several, and not every high-speed board uses it.
Capacity timing is company guidance, and plans move.
"Rationing, not auctioning" is my reading of Nittobo's stated pricing and reported allocations, not its words.

🕳️ The rabbit hole
🧭 THE WHERE
2. Altium: how the fibre weave effect influences signal integrity, the engineering behind the ten-degree trick [10 min]
3. The Real AI Shortage Is Memory, our own piece on the other AI bottleneck that sits in the package [7 min]

📬 The series
PCB Story 1: Why Your PCB Quote Went Up
The price on your desk, and why buyers outside AI are waiting in a queue.
PCB Story 2: Your AI Cluster Runs on Cloth (you are here)
Why a woven glass fabric decides what an AI computer can do.
PCB Story 3: Nobody Is Getting Rid of the Circuit Board
Why the board keeps losing work to the chip package, and gets harder to make anyway.
PCB Story 4: The Map
Who makes the board, where, and how the map came to look this way.

One question to sit with: which input in your own product has one qualified supplier, and what happens the day it starts choosing customers?
Silicon & Steel Intelligence Desk · The Board. If I get something wrong, tell me: [email protected]. Coffee invite open to everyone, free or paid.

