WIKI / AI-SERVER-PCB-PRIMER

AI Server PCB — High-Layer-Count, Anylayer HDI, and the CCL Materials Squeeze

WIKI primerpcbcclai-infrastructurehdihlcglass-clothcopper-foilresin updated 2026-06-30

A first-principles primer on the printed circuit boards that route signals across NVIDIA GB200/GB300 racks, Google TPU pods, and AWS Trainium clusters — the layer of the AI build-out where 80% of the dollar volume sits in three Asian factories and 80% of the pricing power sits two layers upstream in glass cloth.

PART I — THE OPENING HOOK: WHY THIS MATTERS NOW

Two things happened in 2025 that quietly changed the shape of the AI infrastructure trade.

The first was that NVIDIA's NVL72 rack — the GB200 — shipped at scale. Each rack is roughly two tonnes of metal, copper, optics, and silicon. The 72 GPUs and 36 Grace CPUs that draw the headlines are the smallest fraction of the bill of materials. The largest single non-silicon line item is the printed circuit board: the GPU board, the Grace board, the NVSwitch tray, the compute tray backplane, the OAM module. A single GB200 rack consumes more advanced PCB area than an entire smartphone factory's annual output. And the PCBs in question are not the four-layer FR-4 boards you find in a thermostat — they are 30-50+ layer, anylayer-HDI, low-loss laminate boards with line widths under 50µm, drilled with 30,000+ holes per panel, fabricated to a yield where one missed via per board scraps the entire stack.

The second thing was that Nittobo, the world's most important fiberglass cloth maker, was reported to be pulling back from ultra-thin, high-end E-glass cloth (the grade that goes into AI server CCL) toward low-Dk (NE/NER/NEZ) and T-glass. The independently confirmed facts (TrendForce Nov 2025, DigiTimes Jun 2026) are an up-market mix-shift, price hikes of ~20-30% on high-performance and specialty grades through 2026, and actual capacity expansion (T-glass capex, NEZ for 2027, a Nan Ya JV). The stronger reading, that this is a structural withdrawal of capacity from the layer everyone downstream (Shengyi, Panasonic, ITEQ, Elite Material, Victory Giant, Unimicron, Ibiden) depends on, is an interpretation (Collyer Bridge / @QQ_Timmy), not a Nittobo-stated exit, and it is the load-bearing premise of this primer's upstream-squeeze thesis, so it is flagged as unverified pending a primary source. The thin-fabric price moves are documented; the framing that they reflect a capacity withdrawal is not. Taiwan Optical Electronics is forecasting a 2 million meters/month LDK2 supply gap by 2H26. Then in mid-April SABIC, which controls roughly 70% of the world's PPO resin supply, lost 25-30% of its Saudi capacity to a natural-gas constraint. PPO is the resin in low-loss CCL. So we now have a simultaneous shortage developing in both reinforcement (glass cloth) and matrix (PPO resin) of the laminate that goes into every AI server board.

The market is treating this as a downstream "AI demand" story and clustering investor attention on the named PCB fabricators (Victory Giant, Shennan, Unimicron, Ibiden). The fabricators are great businesses with real share gains. But the alpha is two layers upstream, in the glass cloth and resin tier where the supply curve has gone vertical and the consensus model still assumes commodity material. This primer maps the whole stack, identifies who is bottleneck and who is just a beneficiary, and ends with a tiered picks framework.

PART II — THE PROBLEM PCBs SOLVE

A modern computer is, at the bottom, a problem of moving electrons between transistors as fast and as quietly as possible. Inside a single chip the transistors sit micrometers apart and copper interconnects connect them through metal stack layers. But chips have to talk to other chips: GPU to GPU, GPU to memory, GPU to network, GPU to power. That communication happens through wires — and those wires, at the centimeter-to-meter scale, live on a printed circuit board.

A PCB is, at its simplest, a sandwich. Layers of copper foil are separated by layers of glass-cloth-and-resin insulation, glued together under heat and pressure, then drilled and plated to connect the right pieces of copper. The technology is over a hundred years old. Every electronic device you have ever touched contained one. So why is this suddenly a $30B+ growth industry?

Because the requirements have changed faster than the industry was set up to deliver. A smartphone PCB needs to route a few gigabits per second of signal across a 10cm board with a budget of pennies. An NVIDIA GB200 NVSwitch tray needs to route 1.8TB/s of NVLink signal across 50+ centimeters of board with insertion loss tighter than -1.5dB at 28GHz, on a stack of 30-40+ layers, fabricated to <1ppm defect rate, at a unit cost in the high four figures. Those are not the same problem. They are not even the same industry. The companies that win one tend not to win the other.

The problem PCBs solve has therefore split. The legacy PCB business (consumer, automotive, industrial) is a margin-compressed, capital-intensive, fragmented commodity. The advanced PCB business (AI server, networking switch, IC substrate, mil-aero) is a 5-10 player oligopoly with 30-40% gross margins, 18-month qualification cycles, and structural supply tightness in its inputs. Both are called "PCB" but they have less in common than people who don't work in the industry assume.

This primer is about the second one.

PART III — THE TECHNOLOGY FROM FIRST PRINCIPLES

What a PCB actually is

Take two sheets of glass cloth — woven fiberglass, like very thin window screen — and impregnate them with epoxy resin. Press a sheet of copper foil onto each side, heat-cure it, and you have a copper-clad laminate, or CCL. CCL is the raw material of every PCB. It comes in standard thicknesses and grades and is sold by the kilogram by makers like Shengyi, Elite Material, Panasonic, ITEQ, EMC, and TUC.

To turn CCL into a PCB you do five things, in this order:

  1. Pattern the copper. Photolithographically expose the desired trace pattern, etch away the rest. You now have a single-layer board.
  2. Stack and laminate. Sandwich multiple patterned layers between bonding sheets ("prepreg"), heat and press them together. You now have a multi-layer laminate.
  3. Drill. Make holes — thousands to tens of thousands — through the stack to allow electrical connection between layers. Done with mechanical drills (≥0.1mm) or laser drills (≤0.1mm). This is where Union Tool's micro-drills and Mitsubishi Electric's CO2 lasers earn their fees.
  4. Plate. Electroplate copper onto the hole walls (a "plated through-hole" or PTH) so signals can pass between layers. For HDI boards, plate microvias instead of through-holes — they only connect adjacent layers and are far smaller (50-100µm).
  5. Finish. Apply solder mask, silkscreen, surface finish (ENIG, immersion silver, OSP), test, depanel.

That's it. Conceptually, a PCB factory is a machine shop with an electroplating line bolted to a chemistry lab. Every PCB ever made follows this flow. The reason an AI server PCB is hard is not that the steps are different — they're the same — but that the tolerances at every step are 10-100x tighter than a smartphone board.

The four physical limits that drive everything

Once you understand the physics, the rest of the industry's structure falls out of it.

  1. Insertion loss. Every centimeter of trace attenuates the signal. At 28GHz (PCIe 6.0, NVLink 5) the loss budget across a 50cm trace is brutal — you need a CCL with dielectric constant (Dk) ~3.0-3.5 and dissipation factor (Df) <0.002. Standard FR-4 is Dk ~4.5, Df ~0.02 — completely unusable. The materials that hit AI-server requirements are Panasonic Megtron 6/7/7N/8, Rogers RO4000 series, Shengyi S7460, ITEQ IT-988GSE — and these all use low-Dk glass cloth (LDK1, LDK2, sometimes T-glass) plus PPO/PPE-based resin (SABIC's NORYL, Mitsubishi Gas Chemical BT). This is why the materials supply chain matters more than the fab.
  2. Layer count. More layers = more routing capacity = more parallel signal paths between chips. AI accelerators need 80-120+ HBM lanes and 18+ NVLink lanes per GPU; networking switches need 64+ 800G ports. You can't fit that on 8 layers. AI server motherboards now run 20-30 layers; NVSwitch trays and OAM baseboards run 40-50+; 5G base-station boards push 50+. Each layer roughly doubles fab complexity because lamination has to happen sequentially with intermediate inspection.
  3. Via density and aspect ratio. A 1mm-thick board with a 0.1mm via has aspect ratio 10:1 — at the limits of copper plating. AI server boards are pushing to 20:1 and 30:1, requiring HDI (High Density Interconnect) with stacked or staggered microvias, plus anylayer HDI where every layer can have microvias to any adjacent layer. Anylayer is what makes Apple's substrate-like PCBs (SLP) and NVIDIA's GB200 HDI baseboard work.
  4. Yield × area. Yield × board area = effective output. A fab that yields 95% on a 600mm² smartphone board can yield 60% on a 2,500mm² AI server board because defect probability scales with area. Every advanced PCB fab is fighting the yield curve. This is the reason the share leaders aren't just whoever has the most capacity — they're whoever has the highest yield, which is process know-how.

These four constraints — loss, layer count, via density, yield — are fundamental. They cannot be designed around. They can only be solved with better materials, more sophisticated processing, and accumulated craft. The companies that lead are companies that have invested in all three for 15-20 years.

Glossary

Term Definition
PCB Printed circuit board — the stack-up that holds and connects all the components
CCL Copper-clad laminate — the raw sheet material that PCBs are built from
Prepreg Partially cured resin-glass sheet used as the bonding layer between PCB layers
HDI High-density interconnect — PCB with microvias and laser-drilled connections
Anylayer HDI HDI where any layer can connect to any adjacent layer via microvia
HLC High-layer-count — typically ≥20 layers, AI server staple
Dk Dielectric constant — how much the material slows signal propagation
Df Dissipation factor — how much signal energy the material absorbs
PTH Plated through-hole — a drilled hole copper-plated to connect layers
Microvia Laser-drilled via, ≤150µm, connecting adjacent layers only
E-glass Standard electrical-grade fiberglass cloth — cheapest, highest-Dk (~6.9 @10GHz), most common reinforcement
LDK (Low-Dk) Low-dielectric glass cloth. Industry grades: NE-glass (Nittobo "Low Dk1", Dk ~4.7), NER ("Low Dk2"), NEZ ("Low Dk3", 2027). For high-frequency, low-loss PCBs
T-glass Low-CTE / high-modulus glass cloth — for IC substrates and large boards. A separate CTE axis, not the rung above NEZ
Q-glass (quartz) ~99.9% SiO2 fused-silica fiber; lowest-loss woven cloth (fiber Dk ~3.8, Df ~0.0002 @10GHz). Pairs with M9-class resin for Vera Rubin CCL; allocation-constrained, hard to weave and drill. Note: the quoted "Dk 2.3" is the woven-fabric value (fiber + air), not the fiber
PPO/PPE Polyphenylene oxide/ether — the high-performance resin matrix in low-loss CCL
BT resin Bismaleimide-triazine — substrate-grade thermoset resin
OAM OCP Accelerator Module — standardized GPU module form factor
UBB Universal Baseboard — the OAM-supporting motherboard
ABF Ajinomoto Build-up Film — substrate dielectric used between IC and PCB

PART IV — HOW PCB MANUFACTURING ACTUALLY WORKS

The flow, end to end

[Glass yarn]       [Resin]       [Copper foil]
       \             |              /
        \            |             /
       [Glass cloth weaving — Nittobo, Fulltech]
                     |
       [Prepreg / CCL lamination — Shengyi, Panasonic, ITEQ, Elite, EMC]
                     |
       [Inner-layer imaging / etch] — pattern copper on each layer
                     |
       [Layer stacking + lamination] — bond layers under heat/pressure
                     |
       [Drilling — Union Tool drills + Mitsubishi/Schmoll machines]
                     |
       [Desmear + plating] — clean and copper-plate vias
                     |
       [Outer-layer imaging / etch] — pattern outer copper
                     |
       [Solder mask + surface finish + silkscreen]
                     |
       [E-test + AOI + flying probe + impedance test]
                     |
       [Final inspection + depanel] → finished PCB

Every step has its own equipment, chemistry, and yield trap. The hardest are: inner-layer imaging (any defect in any of 30 layers scraps the panel after you've spent the cost of the whole stack), lamination (warpage scales with layer count and panel size), and drilling/plating (aspect-ratio limits and via reliability under thermal cycling).

Where Asian fabs win

The industry consolidated to East Asia 25 years ago and has not de-consolidated. Today >80% of advanced PCB capacity is in China, Taiwan, Japan, and Korea. The reasons are non-romantic:

  • Skilled labor at scale. Inner-layer imaging on a 30-layer board is roughly 30 separate process steps, each requiring a clean panel, a tightly-tuned chemistry, and a yield-conscious operator. Western fabs (Sanmina, TTM, AT&S) survive in mil-aero and medical because they have lower volume and higher pricing — they cannot win the AI server volume.
  • Materials proximity. Shengyi's CCL plant is hours from Victory Giant's PCB fab. ITEQ's CCL is hours from Tripod and Unimicron. Nittobo's glass cloth is hours from Panasonic CCL. Logistics matter because CCL is bulky and prepreg has shelf life.
  • Yield know-how. Advanced PCB yield is empirical, not theoretical. The fabs that have been making 30-layer boards since 2010 (Ibiden, Unimicron, Shennan) have process libraries the new entrants can't replicate.

The implication for investors: the advanced PCB winners list is short and stable. Disruption from below comes from Chinese fabs (Victory Giant, Shennan, Suntak, Aoshikang) climbing the layer-count ladder, not from Western fabs catching up.

Yield economics

A simplified model. Assume:

  • Smartphone board: 8 layers, 100% yield baseline, $5 ASP, $1.50 cost → 70% gross margin (idealised)
  • AI server motherboard: 30 layers, 60% effective yield, $1,200 ASP, $300 marginal cost @ 60% yield → ~75% gross margin for the boards that ship, plus 40% inventory write-off on scrap

The AI server board therefore generates ~5-7x the gross profit per panel-equivalent of a smartphone board even after accounting for yield loss, because the ASP scales faster than the cost. This is why every PCB fab in Asia is racing to qualify on AI server programs: the unit economics are radically better, and capacity tied up in AI work is unavailable for low-end commoditized work that's losing pricing power anyway.

PART V — KEY TECHNICAL METRICS INVESTORS SHOULD TRACK

Metric What it measures Why it matters Watch level
Layer count Number of copper layers Proxy for board complexity AI servers now 22-30L, NVSwitch 40-50L, target 60L+ by 2027
Material grade (Dk/Df, datasheet @10/12GHz) Signal loss properties Determines whether board can support next-gen interconnect Current mainstream Megtron 6/7; AI-grade transitioning to M7N/M8 with Df <0.002
Line/space (L/S) Trace width / gap HDI density Mainstream 50/50µm, AI server pushing to 30/30µm
Aspect ratio Hole depth / diameter Limit of via plating reliability 10:1 mainstream; 20:1 for HLC; >25:1 is bleeding edge
Yield (panel-level) % usable boards per laminated panel The single biggest cost driver Target ≥80% on HLC; new entrants 40-60%
Capex/revenue Capital intensity Rough proxy for tech investment 12-18% for advanced PCB; <8% for commodity
AI revenue mix % % revenue from AI/HPC PCB Determines exposure to the secular driver Victory Giant >70%; Shennan ~40-50%; Tripod ~30%; Unimicron ~20%
Glass cloth grade mix E-glass vs LDK vs T-glass consumed Proxy for material spend going to bottleneck tier LDK1/2 share rising as % of CCL bill; T-glass share rising for substrate

The two metrics with the most signal for the next 12 months are AI revenue mix % (because it filters which fabricators benefit from the demand spike) and glass cloth grade mix (because it identifies the material-tier squeeze).

PART VI — TECHNOLOGY VARIANTS — HEAD TO HEAD

Plated Through-Hole vs Microvia vs Anylayer

Approach How it works Where it's used Limits
PTH (mechanical drill) Drill through entire stack Older boards, high-power layers Aspect ratio cap ~10:1; large via real-estate cost
Microvia (laser drill) Connect adjacent layers only Mainstream HDI One-step microvia depth ~75µm; needs via stacking for multi-layer connection
Stacked microvia Microvias stacked vertically across multiple layers High-density designs (smartphone SLP, AI baseboard) Reliability under thermal cycle is a yield killer; demands tight via registration
Anylayer HDI Every layer can microvia to any adjacent Apple iPhone SLP, NVIDIA GB200 baseboards Most expensive; 5-10x cost of mainstream HDI; <10 fabs globally

Anylayer HDI is where the real moat sits. The industry estimate is fewer than ten fabs worldwide can deliver anylayer HDI at AI server volume and yield. Victory Giant, Unimicron, Tripod, Ibiden, Compeq, AT&S, and a few others. Pricing power lives here.

Standard CCL vs Low-Loss CCL vs IC Substrate

Material Application Dk (datasheet) Df (datasheet) Cost / sqm
FR-4 Consumer, low-end ~4.5 @1GHz ~0.020 1x baseline
Megtron 6 High-speed networking ~3.7 @10GHz ~0.005 @10GHz 4-5x
Megtron 7 / 7N AI accelerator boards ~3.6 @12GHz ~0.003 @12GHz 8-10x
Megtron 8 / Tachyon 100G Next-gen 224G SerDes ~3.3 @10GHz ~0.0018 @10GHz 12-15x
BT resin substrate IC substrate ~3.4 ~0.005 20-30x (substrate, smaller area)
ABF film High-end IC substrate n/a (dielectric film) n/a premium

(Datasheet Dk/Df at the vendor's quoted frequency. Df rises with frequency, so in-situ 28G/56G loss is higher than these reference points. The prior "@28GHz" figures were extrapolations and ran ~0.2-0.4 optimistic on Dk; corrected to published 10/12GHz datasheet values 2026-06-30.)

Material grade and supplier are tied. M7/M7N is dominated by Panasonic with Shengyi and EMC the credible alternates. M8 is essentially Panasonic alone today, with Shengyi catching up. The CCL grade transition is a multi-year share-shift event — every step up in grade reshuffles supplier rankings, because qualification on a new grade takes 12-18 months and the customer has to re-test the board with the new laminate.

Glass-cloth grade ladder — the reinforcement inside CCL

CCL is glass cloth plus resin, so the cloth grade is a direct lever on signal integrity, not just the resin. The reinforcement ladder runs from cheap, lossy E-glass up to quartz. Dk/Df are frequency-dependent; the values below are at ~10GHz from independent sources (Nittobo, AGY, Heraeus, peer-reviewed Scientific Reports 2024) and span vendor datasheets.

Grade Dk @10GHz Df @10GHz CTE (ppm/°C) Role / handling
E-glass (standard) 6.9 0.0070 5.4 Cheapest, easiest to weave/drill; unusable above ~25G (loss + worst glass-resin Dk contrast)
NE-glass (Nittobo "Low Dk1") 4.7 ~0.002–0.0035 (vendor spread) 3.3 Mainstream low-loss cloth for 56G/112G/224G; still a borosilicate, so it drills like E-glass
NER / NEZ (Low Dk2 / Low Dk3, 2027) below NE lower still low The roadmap rungs below NE; NEZ targets 2027 volume
L-glass (AGY / Western) 4.8 0.0030 3.9 Sibling tier to NE; non-Nittobo second source
T-glass (low-CTE) GHz Dk/Df unverified unverified 2.8 Chosen for CTE/warpage, not loss; a separate axis, primarily a substrate cloth — NOT the rung above NEZ
Q-glass / quartz ~3.7–3.8 (fiber) ~0.0002–0.0005 ~0.5 Lowest loss + lowest CTE; fiber Dk closest to resin (least fiber-weave skew). Gated by cost + hardness (damages drills/presses); top-end only (100GHz+ RF, 1.6T optics, flagship AI)

Why quartz beats NE-glass, and why it isn't mainstream yet. Quartz wins the physics on all three axes: its fiber Dk (~3.8) sits closest to the resin (~2.9–3.5), which minimizes the glass-to-resin Dk contrast that drives intra-pair (fiber-weave) skew at 224G PAM4; its Df is roughly an order of magnitude below NE-glass; and its CTE (~0.5) is far lower. But it is the highest-cost reinforcement (fused silica drawn at ~2000°C), and it is hard enough to wear out drills and lamination tooling (US Patent 8,394,731 exists to blend quartz with softer fiber to recover drillability). NE-glass and next-gen NEZ are qualified and in volume; quartz is still early-adoption. So quartz is the better material but not yet the better engineering or economic choice below the extreme top end. One honesty flag: the popular "quartz Dk 2.3" is the woven-fabric effective value (fiber plus air voids), not the ~3.8 fiber — compare 3.8 against the other grades.

Sources: Yang et al., Scientific Reports 2024 (s41598-024-79456-2; E/NE/L at 10GHz); Heraeus / Shenjiu quartz datasheets; Nittobo product pages; TrendForce. T-glass GHz dielectric numbers are not independently verified and are not tabled as datasheet values.

Mechanical drill vs Laser drill

Mechanical (Union Tool, Tungaloy, Kyocera Precision drills, Schmoll/Posalux machines) handles ≥0.1mm holes. Laser (Mitsubishi Electric, Via Mechanics, Schmoll CO2/UV) handles <0.1mm and is essential for HDI microvias. Both are growing. Mechanical scales with layer count (more holes per board). Laser scales with HDI density (more microvias per board). AI server boards consume both at unprecedented rates: 30,000-50,000 mechanical holes plus 200,000-500,000 microvias per panel.

PART VII — A LITTLE HISTORY

The advanced PCB industry was Japanese and American until the late 1990s. IBM, Hitachi, Hadco, Sanmina dominated mainframes, telecom, and military. The first big shift was the smartphone era (2007-2015): Apple's iPhone needed substrate-like PCB (SLP) at consumer volumes that Western fabs could not deliver. Production migrated to Taiwan (Unimicron, Compeq, Tripod) and Japan (Ibiden remained because of substrate exclusivity).

The second shift was the smartphone HDI commoditization (2015-2020): Chinese fabs (Shennan, Suntak, Aoshikang, Victory Giant) climbed the layer-count ladder rapidly, supported by domestic 5G build-out and Apple/Foxconn dual-sourcing. Margins compressed; Taiwanese names underperformed.

The third shift — the one we're in now (2022-) — is the AI server build-out. It is reshuffling rankings again, this time around three axes: (a) AI server PCB volume, where Chinese fabs with 30+ layer capability are taking share; (b) advanced HDI/anylayer, where Taiwanese and Japanese fabs retain a moat; (c) IC substrates (ABF), where Japanese (Ibiden, Shinko) plus Taiwanese (Unimicron, Nan Ya PCB) own >80% of the share.

Why this matters: the names that won the smartphone era are not necessarily the names that win the AI era, and the names that lost the commoditization phase are now winning back. Victory Giant went from #7 with 1.7% global AI/HPC PCB share in 2024 to #1 with 13.8% share in H1 2025. That is a 10x share gain in 12 months. It does not happen in mature industries. It does happen during regime shifts.

PART VIII — THE VALUE CHAIN MAP

                       AI SERVER PCB VALUE CHAIN
                       =========================

LAYER 1 — Raw materials
  Glass yarn:        Nittobo, Taiwan Glass, Owens Corning, AGY, Unitika
  Resin (PPO/PPE):   SABIC (~70% global), DIC, Asahi Kasei
  Resin (BT):        Mitsubishi Gas Chemical, DIC
  Copper:            Standard commodity copper

  ↓

LAYER 2 — Glass cloth / Copper foil (★ BOTTLENECK TIER)
  Glass cloth:       Nittobo (3110.T), Fulltech (1815.TWO), Taiwan Glass,
                     Unitika (3103.T), Asahi Kasei
  Copper foil:       Mitsui Mining & Smelting (5706.T), Furukawa Electric (5801.T),
                     Iljin Materials (KS:020150), Lien Yu (Chang Chun Group)

  ↓

LAYER 3 — CCL / Prepreg (★ BOTTLENECK TIER)
  Premium CCL:       Panasonic, Shengyi (600183.SH), Mitsubishi Gas Chemical
  Standard CCL:      Elite Material (2383.TW), ITEQ (6213.TW), TUC,
                     EMC (1909.TW), Doosan (KS:000150)

  ↓

LAYER 4 — PCB fabricators (the "named" winners)
  AI server PCB:     ★ Victory Giant (002476.SZ / 2476.HK), Shennan Circuits
                       (002916.SZ), Suntak (002815.SZ), Aoshikang
                       (002913.SZ), Tripod (3044.TW), Unimicron (3037.TW),
                       Compeq (2313.TW), GCE (2368.TW), Nan Ya PCB (8046.TW),
                       Ibiden (4062.T), MEC (4971.T)
  IC substrate:      Ibiden, Shinko Electric (6967.T), Unimicron, Nan Ya PCB,
                     Kinsus (3189.TW), Simmtech (KQ:222800)

  ↓

LAYER 5 — Drilling / Equipment / Consumables
  Drill bits:        Union Tool (6278.T), Tungaloy (IMC/Berkshire),
                     Kyocera Precision (6971.T)
  Drill machines:    Mitsubishi Electric (CO2 laser), Via Mechanics (UV),
                     Schmoll, Posalux
  Test/AOI:          Camtek (CAMT), Orbotech (NDSN — Nordson Test & Insp)

  ↓

LAYER 6 — System integrators / OEMs
  AI server ODM:     Foxconn (2317.TW), Quanta (2382.TW), Wiwynn (6669.TW),
                     Inventec (2356.TW), Supermicro (SMCI)

  ↓

LAYER 7 — End customers
  Hyperscalers:      NVIDIA (NVDA), Google (GOOGL), Amazon (AMZN),
                     Microsoft (MSFT), Meta (META), Tesla (TSLA)

Where the profit pools sit

Layer Revenue pool '26E Gross margin range Concentration Where the alpha lives
1 — Raw materials $5-8B 25-35% Medium (SABIC dominant in PPO) SABIC outage = priced
2 — Glass cloth / Cu foil $8-12B 30-50% High — top 3 glass cloth = ~70% Yes — Nittobo/Fulltech bottleneck
3 — CCL $25-35B 25-35% (premium) / 15-20% (standard) Medium-high Yes — premium CCL squeeze
4 — PCB fabricators $80-110B 25-40% (advanced) / 12-18% (commodity) Medium Partial — share-shift play
5 — Equipment / consumables $5-8B 35-45% High (Union Tool, Mitsubishi) Already priced
6 — ODM $300-400B 4-8% High (Foxconn dominant) Low margin, no pricing power
7 — End customer n/a n/a n/a Where the demand comes from

The single most under-priced layer is Layer 2 — high-end glass cloth. Total revenue pool is small (~$8-12B) but margin is expanding fast and the top three players control 70%+. Layer 3 (premium CCL) is the next-best alpha layer because the PPO and glass cloth tightness flows directly into CCL maker margin.

Layer 4 (the named PCB fabs) is where the dollar volume sits but where the market has already started pricing in the AI bull case. Victory Giant, Shennan, Tripod have all run hard. The fabs are still investable but as a tactical share-shift trade rather than a pricing power trade.

PART IX — PRODUCT & SEGMENT CATALOG

Segment A: AI accelerator OAM/UBB baseboards

  • What it is: The board that hosts NVIDIA HGX/B200/B300/GB200 modules, AMD MI300X, Google TPU. Universal Baseboards (UBB) per OCP standard.
  • Tech: 22-30 layers, anylayer HDI required, Megtron 7/7N material, ~600 × 400mm panel, ASP $2,000-5,000.
  • Who makes it: Victory Giant (★ majority share), Tripod, Unimicron, Shennan, Ibiden.
  • Customer: NVIDIA system board direct + via Foxconn/Wiwynn.
  • Volume: ~10-15M panels/yr by 2026E.

Segment B: NVSwitch / network switch trays

  • What it is: The high-speed switching fabric that connects GPUs in NVL72 rack and similar. Handles 1.8TB/s NVLink bandwidth.
  • Tech: 30-50+ layers, anylayer HDI, Megtron 8 material, lowest-loss requirement of any board type, ASP $5,000-15,000.
  • Who makes it: Tripod (★ rumored to be primary), Compeq, Victory Giant (gaining share per Collyer Bridge research).
  • Volume: 1-2 trays per rack × ~100K racks = 100-200K trays/yr.
  • Note: The April 2026 supply story flagged by Global Tech Research — NVIDIA evaluating PCB-based replacement for Amphenol's overpass connector — suggests NVSwitch tray PCBs absorbing more of the rack's high-speed routing. Bullish for advanced PCB volume per rack.

Segment C: Compute tray / CPU baseboards

  • What it is: Boards hosting Grace, Xeon, EPYC CPUs alongside accelerators. Less demanding than accelerator/switch boards.
  • Tech: 16-22 layers, mainstream HDI, Megtron 6 material, ASP $500-1,500.
  • Who makes it: Broader pool — Shennan, Suntak, GCE, Nan Ya PCB, Tripod.
  • Volume: Higher than switch tray but lower per-unit value.

Segment D: Power / mid-board converter PCBs

  • What it is: 48V → 0.8V step-down boards on the GPU/accelerator board. High-current, multi-phase power conversion.
  • Tech: 8-16 layers, heavy copper (3-6oz), thermal-conductive laminate, ASP $200-500.
  • Who makes it: Specialized fabs — Shengyi (in-house), Tripod, Aoshikang.

Segment E: IC substrates (ABF)

  • What it is: The interposer-like substrate that connects the silicon die to the PCB. Different industry segment but adjacent — covered in companion primer on packaging glass substrates.
  • Tech: 4-2-4 to 8-2-8 build-up layers, ABF dielectric, T-glass core, ASP $50-300.
  • Who makes it: Ibiden (★ leader, ~30% share), Shinko, Unimicron, Nan Ya PCB, Kinsus, Simmtech.
  • Volume: Multiple substrates per accelerator package; >100M units/yr.

Segment F: Optical module / co-packaged optics PCB

  • What it is: The PCB inside 800G/1.6T optical transceivers and emerging CPO modules.
  • Tech: 12-20 layers, Megtron 7/7N, very tight impedance control.
  • Who makes it: Compeq, Tripod, Senior PCB.

PART X — INDUSTRY STRUCTURE & DYNAMICS

TAM and growth — Goldman Sachs January 2026 revision

The most authoritative sell-side print on this market was Goldman Sachs in January 2026, summarized in STF Research's "GS Aggressively Revises Up AI PCB/CCL TAM Targets" (7 Jan 2026, mirrored at KB/raw/substack-archive/stf-research/2026-01-07-gs-aggressively-revises-up-ai-pcb-ccl-tam.md). The new figures:

  • AI PCB TAM: $26.6B in 2027 (up from prior $17.4B) — implied 140% CAGR 2025–2027.
  • AI CCL TAM: $18.3B in 2027 (up from prior $8.0B) — implied 178% CAGR 2025–2027.
  • Growth accelerates in 2027 vs 2026 — that is, the second derivative is positive, not topping out.

The driver of the revision is NVIDIA Vera Rubin (VR200 / VR300), which replaces bridge cables with PCBs/CCLs at the midplane and backplane. Per the STF analysis:

  • Midplane content per GPU rises to ~$171–$256 (107% / 57% jump vs GB300 NVL72), starting 2H26.
  • Backplane content per GPU rises to ~$781–$1,563 (5x / 4.5x jump vs GB300), starting 2H27.

The non-obvious wrinkle is yield erosion. As layer counts move 24-28 (2025) → 40+ (2027) and HDI build-up moves 4+N+4 → 6+N+6, industry yield falls from ~73% to ~62%. Lower yield = more raw CCL consumed per finished panel. CCL TAM grows faster than finished PCB TAM (178% CAGR vs 140%) precisely because the boards are getting harder to make and more material is being scrapped along the way.

Translating to broader scope: total advanced PCB market (AI + non-AI advanced) was estimated at $8-10B in 2024, growing to $22-28B in 2026E and toward $45-60B by 2028E. The 5-year CAGR for the AI-specific slice is the headline number, ~140%.

The broader PCB market (including consumer, auto, industrial) is ~$80-100B and growing high single digits. The AI/HPC slice is therefore disproportionately important to growth even though it's still a minority of total revenue.

Cyclicality

Legacy PCB is cyclical (consumer + auto cycles). AI/HPC PCB is currently secular — demand is being set by hyperscaler capex, not end-consumer purchasing. The risk factor is hyperscaler capex moderation in 2027+ rather than near-term cyclicality.

The CCL and glass cloth tiers are driven by lagging supply additions: every cycle of demand growth in PCB triggers a 12-24 month delayed addition of cloth/CCL capacity. The current cycle is on the front edge of that lag — capacity is tight, prices are rising, expansions are announced but not yet shipping.

Barriers to entry by layer

Layer Capital intensity Tech intensity Customer qualification Total barrier
Glass cloth High High (decades of know-how) 12-18 months Very high
CCL (premium) High Very high (resin formulation) 12-18 months Very high
CCL (standard) High Medium 6-12 months High
PCB fab (advanced) Very high Very high (yield know-how) 12-24 months Very high
PCB fab (commodity) Medium-high Medium 6 months Medium
Drilling consumables Medium High (process IP) 6-12 months High
Drill machines High High 24 months High
ODM High Low-medium n/a Medium

The non-obvious takeaway: barrier height in PCB does not map to capital intensity. Glass cloth has lower capex per ton than a CCL line, but qualification cycles and tribal know-how make it the highest-barrier tier. This is exactly what makes it a structural pricing-power layer.

Switching costs

Switching a CCL grade in a qualified PCB design takes 6-12 months of testing and re-qualification with the end customer. Switching a PCB fab takes 12-24 months for an AI accelerator board. Switching a glass cloth maker can take 12-18 months because the CCL maker has to re-qualify the cloth, then the PCB fab has to re-qualify the CCL. These are real, measurable, multi-quarter switching costs. This is what gives each layer pricing power once a customer is locked in.

PART XI — REGULATORY & GEOPOLITICAL LANDSCAPE

Export control exposure

  • US BIS export controls on advanced semiconductors (October 2022, October 2023, December 2024) do not directly target PCBs but do constrain the end customers who buy them. Chinese-domestic AI accelerators (Huawei Ascend, Cambricon, etc.) increasingly use Chinese-domestic PCB fabs (Victory Giant, Shennan, Suntak) — a beneficiary of the US-China decoupling.
  • Conversely, NVIDIA H20/B30 / region-specific Blackwell chips are still US-controlled but use the same PCB stack — Victory Giant supplies both NVIDIA mainline and Chinese-domestic, providing a structural hedge.

Tariffs

  • Section 301 tariffs on China-origin PCB shipped to US customers are non-trivial. Victory Giant's Thailand acquisition (APCB, 2024) and Vietnam plant ($520M in Bac Ninh, started March 2025) are explicit decoupling moves to secure US-bound supply through non-China origin. This is a strategic play, not just a cost play.
  • Taiwan-origin PCBs (Tripod, Unimicron, Ibiden Taiwan) have favorable trade treatment.

Industrial policy

  • China's "Made in China 2025" and follow-on subsidies funnel preferential capex financing to domestic PCB fabs and CCL makers. Shengyi has been an explicit beneficiary; Victory Giant's Huizhou expansion has received local-government support.
  • Japan's METI subsidies favor advanced packaging and substrate (Ibiden, Shinko) but not commodity PCB.
  • Taiwan's policy is hands-off; the cluster sustains itself via private capex.

Supply chain concentration

  • Glass cloth: Japan dominant (Nittobo) → Taiwan secondary (Fulltech, Taiwan Glass). Single-country concentration risk.
  • CCL: Japan/China/Taiwan tri-pole — relatively diversified.
  • PPO resin: SABIC (Saudi Arabia/US) ~70% — single-supplier risk demonstrated by April 2026 outage.
  • Premium PCB fab: 80%+ in China + Taiwan + Japan.

The geopolitical risk profile of advanced PCB is dominated by single-supplier resin and glass cloth concentration, not by fab location. The 2026 SABIC PPO outage was a dress rehearsal for what a real disruption could look like.

PART XII — INDUSTRY MAP — KEY COMPANIES

Glass cloth and copper foil

Company Ticker Position Mkt Cap (USD est) Pure-play? Notes
Nittobo 3110.T Glass cloth tech leader; up-market mix-shift to low-Dk (NE/NEZ) + T-glass ~$3.5B Diversified Mix-shift + ~20-30% price hikes confirmed; "structural E-glass withdrawal" is unverified interpretation
Fulltech 1815.TWO Taiwanese fiberglass — 30-40% E-glass + 60-70% LDK ~$1.9B Yes Direct beneficiary of Nittobo mix-shift; re-rated ~3x (was sub-$1B)
Taiwan Glass 1802.TW Diversified glass; PCB cloth one segment ~$0.7B Diversified Smaller AI exposure
Unitika 3103.T Specialty fiber, niche glass cloth ~$0.3B Diversified Mentioned by Collyer Bridge as alt play
Mitsui Mining & Smelting 5706.T Copper foil + adjacent ~$3.0B Diversified High-end copper foil for HDI
Furukawa Electric 5801.T Copper foil + cables ~$5.0B Diversified

CCL makers

Company Ticker Position Mkt Cap (USD est) Pure-play? Notes
Shengyi Tech 600183.SH China #1 CCL; rising in premium grades ~$25B Yes Direct beneficiary of Chinese AI PCB; Megtron 7/8 alternate
Elite Material 2383.TW Taiwan #1 CCL ~$10B Yes Mainstream + low-loss
ITEQ 6213.TW Taiwan high-end CCL ~$4B Yes Strong AI server exposure
TUC (Taiwan Union Tech) 6274.TW Mid-tier Taiwan CCL ~$3B Yes
EMC (Elite Material China) 1909.TW China-focused mid-tier ~$2B Yes
Panasonic Electric Works (subsidiary) Megtron series — premium AI CCL n/a (parent ¥3T) No (parent diversified) Single most important premium CCL brand globally
Mitsubishi Gas Chemical 4182.T BT resin + substrate CCL ~$3.5B Diversified
Doosan Corp KS:000150 Korean CCL (Doosan Electronics) n/a No

PCB fabricators — AI server / advanced

Company Ticker HQ Mkt Cap (USD est) AI mix % Specialty Notes
Victory Giant 002476.SZ / 2476.HK China ~$28-30B >70% AI server, anylayer HDI #1 global AI/HPC PCB share H1 2025 (13.8%); +600% 2025
Shennan Circuits 002916.SZ China ~$15-18B 40-50% AI server, networking, IC substrate Diversified, broader exposure than VG
Suntak 002815.SZ China ~$5-7B ~30% AI server, automotive
Aoshikang 002913.SZ China ~$4B ~30% AI server
Tripod 3044.TW Taiwan ~$8B ~30-40% NVSwitch tray (rumored primary); CSP boards
Unimicron 3037.TW Taiwan ~$10B ~20-25% IC substrate + advanced PCB Substrate tilts mix
Compeq 2313.TW Taiwan ~$3B ~25% HDI, optical
GCE (Gold Circuit) 2368.TW Taiwan ~$3B ~30% AI server compute boards
Nan Ya PCB 8046.TW Taiwan ~$2.5B ~15% IC substrate + standard PCB
Ibiden 4062.T Japan ~$10B ~20% (PCB only) IC substrate (★) + advanced PCB Substrate is 60% of revenue
MEC 4971.T Japan ~$2B ~15% Specialty CCL chemical + small fab Indirect play

Equipment & consumables

Company Ticker Position Mkt Cap (USD est) Notes
Union Tool 6278.T Drill bits — 30-40% global share ~$1.9B Already covered in vault deep-dive; love business, hate price
Kitagawa Seiki 6327.T Vacuum press for CCL/multi-layer lamination — world #1 small-mid cap Hidden champion per STF Research — Q1 FY2026 order intake ¥5.4B (10yr high); GM 27.2% → 31.9% YoY; PCB share of backlog 68% to 85% YoY; dominant / world-#1 press share, STF cites necessity for M9/Q-Glass tolerances (a literal "sole-qualified supplier" is inference, not confirmed)
Mitsubishi Electric 6503.T CO2 laser drill machines ~$30B Diversified
Via Mechanics n/a (subsidiary) UV/CO2 laser drill n/a Hitachi sub
Camtek CAMT AOI inspection — PCB + advanced packaging ~$5B Strong AI server exposure
Orbotech (Nordson) NDSN PCB direct imaging + AOI ~$15B Diversified

Solder mask & advanced packaging dielectric

Company Ticker Position Notes
Taiyo Holdings 4626.T Solder mask + dry film + FPIM — Rubin substrate exclusivity Per STF Research (Jan 31, 2026): Hopper → Blackwell → Rubin substrate area expands 3,025mm² → 5,625mm² → 8,100mm². B200 dry film attach rate jumps 50% → 90%; Rubin = 100% dry film. Zaristo dry film carries 30-50% ASP premium vs liquid ink. FPIM (Fine Pitch Insulating Material) co-developed with imec for Rubin's <2µm RDL on the HBM4 logic base die — STF claims earnings power could double current implied price over 3 years.

PART XIII — COMPANY DEEP-DIVES — THE LEADERS

Victory Giant Technology (002476.SZ / 2476.HK) — separate full deep-dive forthcoming

  • What they do: Pure-play advanced PCB. AI server + HPC PCB + automotive HDI.
  • Why they win: First Chinese fab to break into NVIDIA's primary AI accelerator board supply at AI-grade quality. Capacity expansion 2-3x from 2023-2026 with Thailand (APCB acquisition) and Vietnam ($520M Bac Ninh) supplementing Huizhou. >70% AI revenue mix is the highest in the peer set.
  • 2024 financials: Revenue CN¥10.73B, NI CN¥1.15B (net margin ~10.7%), Total Assets CN¥19.18B, Equity CN¥8.93B.
  • Bull case: Highest AI mix + share gainer + Nvidia supplier + decoupling-hedged via Thailand/Vietnam.
  • Bear case: Concentration risk on Nvidia; A/H spread tight after HK debut; +600% 2025 leaves little margin of safety.
  • See: companion deep-dive at KB/wiki/002476/002476.md.

Shennan Circuits (002916.SZ)

  • What they do: Most diversified Chinese PCB peer. AI server PCB + networking + IC substrate + automotive.
  • Why they win: Long-standing relationships with Huawei, ZTE plus AI server new wins. Substrate business adds optionality.
  • Bull case: Diversified beneficiary of both Chinese-domestic AI (Ascend) and Western AI (Nvidia) supply chains. Substrate adds high-margin layer.
  • Bear case: Less concentrated AI exposure than Victory Giant; share gain story is less explosive.

Tripod Technology (3044.TW)

  • What they do: Taiwanese PCB fab with strong NVSwitch tray exposure (rumored primary supplier per Collyer Bridge). Also CSP boards, optical PCB.
  • Why they win: Taiwanese cluster proximity, deep relationships with Foxconn/Quanta/Wiwynn, capability for highest-layer NVSwitch boards.
  • Bull case: Direct play on NVSwitch tray volume per rack; potential beneficiary of Amphenol overpass replacement story.
  • Bear case: Heavy Nvidia concentration; limited disclosure on customer mix.

Unimicron (3037.TW)

  • What they do: Taiwanese leader in advanced PCB + IC substrate. Substrate is the larger and higher-margin business.
  • Why they win: One of <5 substrate makers globally with ABF capability at AI accelerator scale.
  • Bull case: Substrate tightness story (separate from PCB); secondary play on AI accelerator share growth.
  • Bear case: Dilutes "pure PCB" exposure; substrate cycle has its own dynamics.

Ibiden (4062.T)

  • What they do: Japan's premier IC substrate maker (~30% global share); also advanced PCB.
  • Why they win: Substrate moat. ABF incumbent on Intel + AMD + Nvidia accelerators.
  • Bull case: Cleanest substrate pure-play in public markets; cyclically depressed entering AI ramp.
  • Bear case: Substrate pricing has been weak; Japanese cost structure higher than Taiwanese.

Shengyi Tech (600183.SH)

  • What they do: China's #1 CCL maker. Pushing into premium grades (Megtron-equivalent).
  • Why they win: Vertically integrated CCL with growing share in premium grades; low-cost domestic competitor to Panasonic/Mitsubishi.
  • Bull case: Direct exposure to PPO/glass cloth tightness; pricing power expanding as premium mix climbs.
  • Bear case: Premium grade qualification still maturing; some Western customers resistant.

Panasonic — Megtron CCL business (parent 6752.T)

  • Why it matters: Single most important AI server CCL brand. Megtron 7/7N/8 is the de facto standard.
  • Why it's not investable directly: Parent is a ¥3T conglomerate with massive non-PCB exposure (battery, appliance, auto). The Megtron business is high-margin but small in the parent context. Use as analytical anchor only.

Fulltech (1815.TWO) — the bottleneck candidate

  • What they do: Taiwanese fiberglass yarn and woven cloth maker. 30-40% E-glass, 60-70% LDK.
  • Why they win: Direct beneficiary of Nittobo's up-market mix-shift away from high-end E-glass. Long history in ultra-thin fabric, exactly the spec going short.
  • Bull case: EPS NT$6-8 in 2026 / ~NT$10 in 2027 per Taiwanese sell-side estimate cited by Collyer Bridge (unverified); 10% T-glass revenue target by 2026 adds optionality. After the ~3x re-rate to ~$1.9B, "undervalued" is no longer obvious on size.
  • Bear case: Single-product company at the supply tier; demand dependent on a single end use; small market cap = volatile.
  • Conviction: Bottleneck-tier name worth tracking, but the cheap-entry case is largely spent. Mkt cap ~US$1.9B (TWD 62.2B at TWD 106.5, 2026-06-30, yfinance), already re-rated ~3x off the old sub-$1B base, so "under-priced small-cap" no longer applies on size; structural pricing power intact. Verify entry before chasing.

Nittobo (3110.T)

  • What they do: Japanese fiberglass yarn + cloth + insulation + chemicals. Tech leader in glass cloth.
  • Why this matters: Up-market mix-shift to LDK/T-glass plus price hikes (confirmed); a literal structural E-glass capacity exit is unverified interpretation. Either way they are moving up-market, not getting smaller.
  • Bull case: Margin expansion as the high-end mix shift completes and dominates. Exposure to T-glass for IC substrate (separate growth driver).
  • Bear case: Classic "leaving the easy market" story — execution on premium grades has to deliver. Diversified parent dilutes pure-play exposure.

Union Tool (6278.T)

  • Already covered: See KB/wiki/6278/6278.md. Love the business, current price re-rated 5x. Watch for pullback.

PART XIV — EMERGING PLAYERS & DISRUPTORS

  • Aoshikang (002913.SZ) — China-domestic, climbing fast on automotive + AI server crossovers.
  • Wus Printed Circuit (Kunshan) — private Chinese fab gaining HLC qualification.
  • Bomin Electronics — private Taiwanese fab specializing in HLC NVSwitch-style boards.
  • Senior PCB (300476.SZ) — Chinese fab pushing into optical module PCB.
  • Korean CCL (Doosan, LG Innotek) — capable but commercially limited outside Korean OEMs.
  • Indian / SEA fabs — material capacity but no qualified AI server presence yet. Potential 2027-2030 story tied to government industrial policy.

The disruption threat to incumbent winners is modest in the next 24 months because qualification cycles take that long. From 2027+, expect Chinese mid-tier fabs (Suntak, Aoshikang, Wus) to consolidate share against Taiwanese mainstream PCB peers. Top tier (Victory Giant, Shennan, Tripod, Unimicron, Ibiden) is more defensible.

PART XIV-B — BOTTLENECK HUNTING — WHERE DOES PRICING POWER LIVE?

Layer Smallest pure-play Mkt cap Concentration Bypassable? Market priced-in?
Glass cloth (LDK/E-thin) Fulltech (1815.TWO) ~$1.9B High (Nittobo+Fulltech+Taiwan Glass ~70%) No (12-18mo qual) Re-rated (was sub-$1B)
Glass cloth (T-glass) Nittobo (3110.T) $3.5B Very high (Nittobo dominant) No Partly priced
PPO resin SABIC (private/listed parent 2010.SR) n/a Very high (SABIC ~70%) No Partly priced (April outage)
Premium CCL Panasonic (6752.T) parent $30B+ High (Pana + Shengyi + MGC) Partial Partly priced
Standard CCL Elite (2383.TW), ITEQ (6213.TW) $10B / $4B Medium Partial Priced
Anylayer HDI fab Tripod (3044.TW), Compeq (2313.TW) $8B / $3B Medium-high (top 5 ~70%) No (24mo qual) Mostly priced
AI server PCB (HLC) Victory Giant (002476.SZ) ~$28B Medium (top 5 ~60%) Partial Priced post-rerate
Drill bits Union Tool (6278.T) $1.9B High (top 3 ~60%) No Priced (re-rated 5x)
Drill machines Mitsubishi Electric (6503.T) $30B High (top 2 ~70%) No Priced
ODM Foxconn (2317.TW), Quanta (2382.TW) $90B / $30B High No Priced

Top bottlenecks ranked

  1. Glass cloth — LDK/ultra-thin E-glass. Fulltech is the smallest pure-play, riding the Nittobo mix-shift (the structural-exit reading is unverified), demand acceleration from NVIDIA NVSwitch + Google TPU, projected 2H26 LDK2 supply gap of 2M meters/month+. Market cap ~$1.9B, re-rated ~3x off the old sub-$1B base, so the cheap-entry window has largely closed. High-quality bottleneck, no longer the cheap small-cap it was.
  2. Vacuum press equipment — Kitagawa Seiki (6327.T). As Vera Rubin moves to M9 + Q-Glass with tighter processing windows, second-tier presses cannot hold tolerance on $30-50K boards. Q1 FY2026 order intake at 10-year high (¥5.4B), gross margin expanding +470bps YoY. Small/mid cap, no English sell-side coverage. Strong alpha conviction.
  3. Solder mask dry film + FPIM — Taiyo Holdings (4626.T). Rubin substrate is 100% dry film vs Hopper's mostly-liquid; Zaristo carries 30-50% ASP premium; FPIM is a Taiyo + imec monopoly on Rubin's HBM4 RDL. STF Research suggests earnings power could double current implied valuation over 3 years.
  4. PPO resin. SABIC outage in April 2026 was a demonstration. No pure-play public alternate exists. Watch for substitute resin chemistries (Asahi Kasei NORYL competitors) gaining share.
  5. Premium CCL (Megtron 7N/8/M9 grade). Panasonic dominant; Shengyi catching up; pricing power will expand as glass cloth and resin tightness flows through. Shengyi (600183.SH) is the closest investable proxy.

Closing question

"What is the $100M-$1B Mkt cap name with unique exposure to the AI PCB material bottleneck that the market hasn't priced?" Fulltech (1815.TWO) was that name; after a ~3x re-rate to ~$1.9B it has outgrown the sub-$1B window, so the open question now points to second sources (Taiwan Glass 1802.TW) or further upstream. Nittobo (3110.T) remains the higher-quality but more diversified play.

Follow the capex

NVIDIA's GB200/GB300/Vera Rubin capex commitments are the demand pulse. Trace 4 layers of beneficiaries:

  1. Direct recipient: PCB fabricators with AI accelerator wins (Victory Giant, Tripod, Shennan).
  2. Near-competitors / alt-suppliers: Fabs absorbing overflow (Unimicron, Compeq, Suntak).
  3. Upstream components / tools / substrates: CCL makers (Shengyi, ITEQ, Panasonic), drill bits (Union Tool), substrate (Ibiden, Unimicron).
  4. Second-derivative capex cycle: Glass cloth (Nittobo, Fulltech), PPO resin (SABIC alternates), copper foil (Mitsui, Furukawa), drill machines (Mitsubishi Electric, Via Mechanics).

The named layer is priced; the second-derivative layer is not. The repeating LITE/COHR pattern from the 2024 NVDA EML preallocation suggests Fulltech-style names (small cap, single-feature exposure to an upstream squeeze) will be where the multibagger lives.

PART XV — SECULAR TAILWINDS & HEADWINDS

Tailwinds

Tailwind Mechanism Magnitude Durability
AI accelerator volume growth Each new GPU generation = new PCB design = repeat qualification cycle Very large 5-10 years
Layer count escalation Every new GPU/switch generation adds 2-4 layers Large 3-7 years (until photonic interconnect displaces some PCB routing)
Material grade transition Megtron 6 → 7 → 7N → 8 = ASP step-up at fab + CCL maker Medium-large 3-5 years per transition
Geographic diversification (China decoupling) Customers paying for non-China origin = higher margin Thailand/Vietnam Medium 5-10 years
Hyperscaler vertical integration Custom silicon (Trainium, TPU, Maia) = multiple custom PCB designs each Medium-large 5-10 years
Glass cloth and resin tightness Margin expansion at materials tier Large 2-3 years (until capacity catches up)

Headwinds

Headwind Mechanism Likelihood Impact
Hyperscaler capex normalization (2027+) After current build-out, capex growth flattens Medium-high Medium
Photonic / co-packaged optics displacement Some high-speed PCB routing moves into optics Medium Small near-term, larger by 2030
Glass-core packaging substrate If Intel GlassCore takes share from organic substrate, IC substrate makers (Ibiden, Unimicron) lose growth Medium-low Substrate-specific; not direct PCB
China-on-China competition Domestic Chinese fabs commoditize each other on price High Medium for mainstream; low for advanced HDI
Tariff escalation US 232/301 tariffs on Asian PCB Medium Margin compression for non-China-origin Asian fabs

The single biggest risk to the bull case is hyperscaler capex normalization. The single biggest opportunity beyond AI is custom hyperscaler silicon — every Trainium/TPU/Maia generation is a distinct PCB design, distinct qualification, distinct revenue stream.

PART XVI — TECHNOLOGY ROADMAP

Year Layer count (AI server) HDI build-up Material grade Interconnect speed Key shift
2022 16-22L 3+N+3 Megtron 6 28G NRZ / 56G PAM4 Hopper / H100 ramp
2024 22-28L 4+N+4 Megtron 6/7 56G/112G PAM4 Blackwell B100/B200
2026 28-32L 4+N+4 → 5+N+5 Megtron 7/7N 112G/224G PAM4 GB200/GB300
2027 40+L 6+N+6 M9 + Q-Glass 224G PAM4 Vera Rubin VR200/VR300
2028E 40-50L 6+N+6 M9 / next-gen 224G+ Vera Rubin Ultra
2030E 50-60L+? 7+N+7? Glass-core hybrid? Photonic-PCB hybrid CPO mainstream

Source for the 2027 line: Kitagawa Seiki coverage in STF Research (March 2026). The transition to M9 and Q-Glass ultra-low-loss materials narrows the processing window dramatically — vacuum-press tolerances and lamination cycles for M9/Q-Glass are tighter than anything in production today, which is why Kitagawa's order intake has hit a 10-year high (¥5.4B in Q1 FY2026, gross margin 27.2% → 31.9% YoY). When the materials get harder to handle, the equipment that handles them becomes a bottleneck.

The HDI build-up notation n+N+n describes "n build-up layers, N core layers, n build-up layers". 4+N+4 means 4 layers of HDI build-up on each side of the core. 6+N+6 means six on each side. Each step up in n requires another full sequential lamination cycle and another generation of yield risk. Going from 4+N+4 to 6+N+6 is a much bigger production shift than the layer count alone suggests.

The transition to watch in 2027-2028 is the integration of photonic interconnect into the PCB stack. Co-packaged optics, in-package silicon photonics (LightMatter, Ayar Labs, etc.), and on-board optical interconnect will displace a fraction of the highest-loss high-speed traces with optical waveguides. That changes the PCB content per rack — somewhat smaller PCB area, but at higher complexity and ASP. Net effect: probably neutral-to-slightly-positive for advanced PCB demand, definitely positive for the highest-grade laminate (because the surviving electrical traces are even more performance-critical).

The longer-term risk is glass-core packaging substrate displacing organic ABF substrate (covered in companion primer). That's a substrate story, not a PCB story, but it does affect Ibiden / Unimicron mix.

PART XVII — ADJACENT INDUSTRY CONVERGENCE

  • IC substrates — directly adjacent. ABF substrate makers (Ibiden, Shinko, Unimicron, Nan Ya PCB, Kinsus) are PCB-adjacent companies whose substrate business is more glass-cloth-and-resin-intensive than their PCB business. Substrate tightness flows back to glass cloth.
  • Co-packaged optics / silicon photonics — emerging convergence layer. SiPh modules (Coherent, Lumentum, Marvell, Innolight) sit on/in PCB and may displace some PCB routing.
  • Power electronics — 48V → 0.8V power delivery on AI accelerator boards is converging with PCB design. SiC / GaN power devices on PCB and in-rack Power Shelf design (Forgent FPS, Delta Electronics) are PCB-adjacent.
  • Burn-in test — covered in vault burn-in primer; PCBs of test sockets and interface boards are an adjacent niche.
  • Probe cards — covered in vault probe card primer; uses HLC PCB internally for high-pin-count tester interfaces.

The most under-appreciated convergence is the migration of high-speed routing between PCB, substrate, and optics. Whoever wins each generation depends on physics + cost + ecosystem readiness. The PCB layer has more demand than market consensus models because some of this routing was projected to move to optics by now and hasn't.

PART XVIII — CYCLE POSITIONING

We are early in the up-cycle. Indicators:

Signal Current Reading
CCL utilization (premium grade) >85% Tight
Glass cloth utilization (LDK/E-thin) >90% Very tight
PCB fab utilization (advanced) 80-90% Tight
PCB fab utilization (commodity) 60-70% Soft (irrelevant for AI thesis)
Drill bit consumption All-time high Tight
PCB ASP (advanced HLC) Rising Pricing power flowing
Lead times (CCL, premium) 8-14 weeks (vs 4-6 normal) Tight
Inventory days (premium CCL) Below seasonal Tight
Capacity announcements Multiple, but not yet shipping Lagged

Cycle position: mid-2nd inning of a 2-3 year up-cycle in the materials tier; 4th-5th inning in the named PCB fabricator tier. Materials capacity comes online slower (12-24 month lag) so that tightness lasts longer. PCB fab capacity comes online faster (12-18 months) so the named-fabricator pricing power wears off earlier.

Trough/peak signals to watch:

  • Trough: CCL utilization drops below 70%, lead times return to 4-6 weeks, glass cloth spot prices stabilize.
  • Peak: Hyperscaler capex guidance flattens, rack-per-quarter shipments plateau, second-tier Chinese fabs achieve advanced HDI qualification at scale.

PART XIX — HOW TO INVEST

Conviction-ranked picks

Rank Company Ticker Layer Thesis Risk Timeframe
1 Fulltech 1815.TWO Glass cloth Bottleneck-tier pure-play; Nittobo mix-shift + LDK gap; sell-side EPS NT$6-8/10; ~$1.9B mid-cap, re-rated ~3x (no longer cheap small-cap) Single-product, AI cycle exposure, valuation re-rated 12-24 months
2 Kitagawa Seiki 6327.T Vacuum press equipment Dominant / world-#1 press supplier for M9/Q-Glass (sole-qualified is inference); 10-yr-high orders; +470bps GM YoY; no English coverage = mispricing Small-mid cap; cyclical with PCB capex 12-24 months
3 Taiyo Holdings 4626.T Solder mask + FPIM Rubin = 100% dry film; FPIM imec monopoly; STF claims 2x current implied price Substrate cycle exposure 18-36 months
4 Shengyi Tech 600183.SH Premium CCL Direct beneficiary of glass cloth + PPO tightness; premium grade share gain; Chinese AI domestic catalyst China-listed liquidity; FX 12-24 months
5 Victory Giant 002476.SZ / 2476.HK AI server PCB #1 global AI/HPC PCB share; Nvidia anchor; geographic decoupling hedge; HK listing brings Western capital Concentration, post-rerate valuation 12-18 months
6 Tripod 3044.TW NVSwitch tray PCB Rumored primary supplier; Amphenol overpass replacement upside; Taiwanese cluster moat Customer concentration 12-18 months
7 Ibiden 4062.T IC substrate Substrate cycle re-acceleration with AI; cleanest substrate pure-play; cyclically depressed Substrate pricing weak; Japan cost structure 18-36 months
8 Nittobo 3110.T Glass cloth (high-end) Premium mix shift; T-glass growth Diversified parent dilutes 18-36 months
9 Shennan Circuits 002916.SZ AI server PCB + substrate Diversified Chinese AI PCB beneficiary Less explosive than VG 12-24 months
10 Union Tool 6278.T Drill consumables Toll-booth on PCB volume; razor-blade economics Already re-rated 5x; entry too rich Watch for pullback to ¥10-12K

Tier framework

Tier 1 — Core (own through cycle):

  • Fulltech (1815.TWO) — the glass-cloth bottleneck pure-play; ~$1.9B after a ~3x re-rate, so quality remains but the cheap entry is gone.
  • Kitagawa Seiki (6327.T) — equipment bottleneck, no English coverage, 10-year-high orders. Asymmetric.
  • Shengyi Tech (600183.SH) — premium CCL, China-domestic, structurally advantaged.

Tier 2 — Tactical (strong but more cycle-sensitive):

  • Taiyo Holdings (4626.T) — Rubin substrate dry film + FPIM monopoly; substrate-tier exposure.
  • Victory Giant (2476.HK) — best AI mix; tactical buy on pullbacks.
  • Tripod (3044.TW) — NVSwitch tray + Amphenol replacement narrative.
  • Ibiden (4062.T) — substrate inflection re-acceleration.

Tier 3 — Watchlist:

  • Nittobo (3110.T) — wait for clearer T-glass execution.
  • Shennan Circuits (002916.SZ) — wait for AI mix to reach 50%+.
  • Union Tool (6278.T) — wait for ¥10-12K pullback.
  • Fulltech alt: Unitika (3103.T) — flagged by Collyer Bridge but supply-chain feedback "not very good"; trade only.

Avoid:

  • Mainstream Taiwanese PCB (Compeq, GCE) at current valuations — diluted AI exposure, valuations not cheap.
  • Pure commodity CCL (TUC, EMC at top of cycle) — no premium-grade exposure.
  • Western advanced PCB (Sanmina, TTM) — no AI volume; mil-aero cycle weaker.

Pure-play scoring

Company AI server PCB exposure Materials tier exposure Total purity score
Victory Giant 9/10 1/10 High purity, downstream layer
Fulltech 1/10 (indirect) 10/10 Highest purity, alpha layer
Shengyi 4/10 9/10 High purity, premium CCL focus
Ibiden 4/10 (PCB) + 8/10 (substrate) 4/10 Substrate-dominated
Tripod 6/10 2/10 Medium-high, fab tier

PART XX — KEY QUESTIONS TO KEEP RESEARCHING

  • Does Fulltech's revenue share to LDK actually hit ~70% by end-2026? This is the make-or-break of the thesis.
  • Does Nittobo's E-glass mix-shift accelerate, and is a literal capacity exit ever confirmed? Faster mix-shift means a bigger gap; if no capacity withdrawal materializes, the squeeze thesis softens.
  • What's SABIC's Saudi PPO ramp-back timeline? April 2026 outage; if it stays out 6+ months, premium CCL margin spikes further.
  • Does NVIDIA's PCB-replacement for Amphenol overpass happen? Confirms Tripod / Victory Giant volume-per-rack uplift.
  • Vera Rubin layer count specs? Will determine 2H27 demand step-up.
  • Glass-core substrate adoption timeline (Intel)? Affects Ibiden / Unimicron substrate trajectory.
  • STF Research GS PCB/CCL TAM piece (paywalled — fetch separately) — quantitative anchor for revised market size.
  • Chinese-domestic AI accelerator (Ascend) PCB content per package — drives Shennan + Suntak share independently of NVIDIA.

Best sources to follow

  • Collyer Bridge / @illyquid — Asia AI/semis live tracking; has been ahead on Macronix, Rigaku, Victory Giant.
  • Acid Investments — Substack, Asia hardware specifically. Quick-idea posts on listings.
  • STF Research (stfbutnou.substack.com) — focused on AI infra investment angles; paid, paywall-bypassed via subscription.
  • Global Tech Research (globaltechresearch.substack.com) — NVIDIA-specific PCB/CCL design change tracking.
  • Goldman Sachs / Morgan Stanley Asia hardware notes — AI PCB/CCL TAM research, periodic re-rating events.
  • Industry trade press — DigiTimes, Trendforce, EE Times Taiwan.
  • Twitter / X — @QQ_Timmy, @LinQingV (Macro_Lin), @zephyr_z9, @Mike10947310 for real-time supply chain commentary.
  • Macro_Lin's market commentary is the best Mandarin-language source on CCL / PPO / glass cloth squeeze.

2026-07-01 UPDATE — current-price verdicts, the copper-foil second-sourcing delta, and cross-links

Since the 2026-06-30 correction, four dedicated ticker deep-dives were written to the vault, Fulltech, Glotech, Nittobo, and Mitsui, plus an eight-name comparison at _compare/glass-fabric-bottleneck-8name. Five deltas update the picks above.

The whole cohort is now WATCH-grade at current prices. The theme is intact but the easy entry has passed, and every name is capped by a specific thing.

  • Nittobo (3110.T), around ¥4,580: the cleanest monopoly on the bottleneck, but it is not exercising its pricing power. Its NER glass is largely committed to Doosan, and the August 2025 roughly 20% hike reportedly excluded AI-grade fabric, so the premium is leaking to Taiwan and China. The headline PE near 20x is flattered by a roughly ¥34B one-off, confirmed as a Yaesu Tokyo property disposal to Sumitomo Realty, so the clean forward multiple is closer to 47x to 49x. The yfinance forward-PE of 15.5x and price-to-book of 0.96x are broken by the 5-for-1 split; the true P/B is about 4.9x. Verdict WATCH, half-weight, catalyst is the next-earnings pricing posture.
  • Mitsui Kinzoku (5706.T), around ¥41,940: the one monopoly that is pricing (the 12% MicroThin hike stuck), but FY3/27 is guided down about 30% as metals-inventory gains reverse, it trades about 32x forward and above its own sum-of-parts, and roughly half of profit is still a commodity smelter. Verdict WATCH.
  • Fulltech (1815.TWO), around NT$101: cleanest pure-play, best gross margin of Taiwan's five at 35.97% in Q1 2026, but a price-taker at about 31x on peak-cycle EPS, with TPEx plus-or-minus 10% liquidity and an IDEAL BIKE cross-holding governance flag. Verdict NEUTRAL, Medium.
  • Glotech (5475.TWO), around NT$241.5: the Q-glass and M9 Rubin optionality, priced as if the win is in hand at about 30x price-to-book and price-to-sales, and binary on Rubin timing. Verdict Low to Medium.

The copper-foil layer is being second-sourced, in Goldman's own model. Goldman forecasts the qualified second source Co-Tech (8358.TWO) rising from about 5% to about 53% of high-end HVLP3-plus foil share by 2028, with a third plant online in Q4 2027, and NVIDIA is courting Co-Tech directly to break Mitsui's single-point bottleneck. Lotte Energy Materials (KRX) is also sampling. So the durable 90%-plus HVLP4 monopoly framing is contested at exactly the layer that matters, which is why Mitsui is a WATCH rather than a buy.

The quartz window may be shorter than the bulls assume. Nittobo's own improved NEZ-glass, slated for 2027, is expected to gradually displace Q-glass, which caps the runway for the pure quartz names such as Glotech even before the Rubin adoption question.

Kingboard (0148.HK) has run past its value setup. Collyer Bridge's May 2026 sum-of-the-parts thesis (the Kingboard Laminates stake worth more than the whole holdco) has very likely closed after the stock ran about 308% since end-2025 to a market cap near HK$133B, about US$17B. The live sum-of-the-parts must be re-priced before any action, and integration insulates Kingboard from the glass-cloth spike rather than capturing it.

What STF actually holds. STF Research's 28 June 2026 memo names Mitsui (5706) as "our copper foil play" on the HVLP4 shortage, holds Fulltech as the low-Dk demand pick and Glotech as the Q-glass optionality, and holds some Nittobo. Within the glass layer STF prefers the Taiwan weavers because they both expand capacity and raise prices, capturing the Taiwan premium Nittobo leaves on the table.

Net, the tiered framework above stands directionally with two amendments on current prices. Nittobo moves from Tier 3 "wait" to a half-weight WATCH gated on the pricing turn, and the copper-foil layer (Mitsui, with Co-Tech 8358.TWO as the second-source watch) is added to the map, since it gates the same laminate and is where pricing power is actually being exercised.

Sources

  • Collyer Bridge / @illyquid — APAC Wrap 17 April 2026 (Fulltech, Nittobo, SABIC PPO, Macro_Lin) — KB/raw/substack-archive/illyquid/2026-05-02-apac-wrap-17-april-2026-900.md
  • Collyer Bridge / @illyquid — APAC Wrap 21 April 2026 (Victory Giant HK debut) — KB/raw/substack-archive/illyquid/2026-04-26-apac-wrap-21-april-2026.md
  • STF Research — GS Aggressively Revises Up AI PCB/CCL TAM Targets (Jan 7, 2026) — KB/raw/substack-archive/stf-research/2026-01-07-gs-aggressively-revises-up-ai-pcb-ccl-tam.md
  • STF Research — Taiyo Holdings: Quantifying Solder Mask Demand for AI Servers (Jan 31, 2026) — KB/raw/substack-archive/stf-research/2026-01-31-taiyo-holdings-quantifying-solder-mask-demand-ai-servers.md
  • STF Research — Kitagawa Seiki: Hidden Champion Pressing the AI Era (Mar 16, 2026) — KB/raw/substack-archive/stf-research/2026-03-16-kitagawa-seiki-hidden-champion-pressing-ai-era.md
  • Acid Investments — Quick idea: A/H listing of Victory Giant (2476.HK)
  • South China Morning Post — Victory Giant HK debut +57.23%, US$2.6B raise
  • Wikipedia — Victory Giant Technology corporate facts
  • Global Tech Research — GB300 / Vera Rubin PCB/CCL design change post (referenced; full content gated)
  • Vault: KB/wiki/6278/6278.md (Union Tool deep-dive — drill consumables context)
  • Twitter via Collyer Bridge: @QQ_Timmy (Fulltech / Nittobo thesis); @LinQingV (CCL / PPO supply commentary); @zephyr_z9 (E-glass DDR4 analog)
  • Industry technical references on Megtron grade specs and HDI qualification cycles (general knowledge, not a single-source citation)

Pre-delivery checklist: redundancy sweep ✓ (cut three repeated CCL-grade mentions); word justification ✓ (every grade table earns its space — they're how the value chain hierarchy makes sense); guide pass ✓ (Register D — investment writeup; em-dashes used as Register D allows but sparingly in context-setting passages; no marketing words). Length is long but the industry warrants depth — Pink instructed "no length limit; thoroughness beats brevity."

Topics

from the vault · open in obsidian ↗