CCLs - The Invisible Bottleneck | Long-term Investing : Thematic

M8 → M9 → M10 — the spec migration in AI server laminates, and the names positioned to monetise it

LC Conviction · HIGH

A sheet of resin-impregnated glass cloth, clad in copper foil — the most boring-sounding material in the AI stack, and quietly one of its most supply-constrained.

What is a CCL? — The 30-second primer

A Copper Clad Laminate (CCL) is the raw sheet material from which every printed circuit board (PCB) is fabricated. Mechanically, it is a sandwich: a layer of woven glass-fibre cloth (often called “glass cloth” or “fibreglass”) impregnated with a thermoset resin, with one or two ultra-thin copper foils bonded to its surfaces under heat and pressure. The PCB fabricator then etches the copper into circuit traces, drills holes for vias, and stacks multiple CCL sheets together to make a multi-layer board. A modern AI server compute board can have 26 to 36 such layers stacked.

The CCL does two things that matter: (i) it provides the mechanical and dimensional stability on which all the high-precision circuit features sit, and (ii) its dielectric properties — the famous “Dk” and “Df” numbers — directly determine how cleanly a high-speed signal can travel from one chip to another. As signal frequencies climb from 56 Gbps to 112 Gbps to 224 Gbps PAM4 (i.e. the move from 400G to 800G to 1.6T networking), the resin and glass cloth must absorb less and less of the signal — driving the industry up the spec ladder from M4 (standard FR-4) through M6, M7, M8 and now M9 / M10 grades, each demanding fundamentally different raw materials and tighter manufacturing tolerances.

The Top-Down Thesis

CCL is becoming the AI server’s silicon-equivalent gatekeeper — a small, technical, oligopolistic market with multi-year qualification cycles, exploding content per rack, and a raw-material bottleneck (Q-glass / T-glass) the industry cannot solve on demand.

• 01. The spec ladder is accelerating, and each rung is a different industry.

  AI servers have moved from M4/M6 (Hopper) → M8 (GB200) → M8U/M8.5 (GB300) → M9 (Rubin switch tray, mid-plane, CPX) — with Nvidia having initiated M10 testing in March 2026 for Rubin Ultra and Feynman platforms. Each generational step requires fundamentally new dielectric chemistry (PPO/PTFE blends), new glass cloth (Q-glass or quartz cloth), new copper foil (HVLP-3, HVLP-4) and new press / lamination processes — and crucially, the qualified supplier list shrinks at every step. M9 currently has fewer than five qualified producers globally.

• 02 Content per rack is materially expanding, not just substituting.

  AI server PCB demand jumped from ~15% of total PCB market in 2025 to over 25% in 2026, with PCB value per AI server unit up >30% generationally. Layer count has leapt from 16–20 layers (general server) to 28–36 layers (AI server). Rubin Ultra’s confirmed switch from copper cables to orthogonal backplanes will create

  entirely new CCL demand pockets— three 26-layer M9 boards laminated into 78-layer assemblies per backplane.

• 03 Qualification cycles are an absolute moat.

  Customer qualification at Nvidia / AWS / Google / Meta runs 12–18 months and routinely fails — EMC, the world’s #1 specialty CCL maker, failed Nvidia’s GB300 compute tray certification in late 2025, handing Doosan an effectively exclusive position on that platform. This is a market where capacity is meaningless without qualified status, and qualified status is irrevocably scarce. The result: when AI demand surges, marginal supply cannot simply scale in — incumbents capture the entire delta.

• 04 The raw-material chokepoint is upstream.

  The advanced glass cloth (Q-glass, T-glass, low-Dk3 variants) needed for M8/M9/M10 grades is produced almost entirely by Japan’s Nittobo, with smaller capacity at Asahi Kasei, Shin-Etsu and AGC. Apple, Nvidia, AMD, Microsoft, Amazon and Qualcomm executives have all visited Nittobo in the past year; new capacity does not arrive until 2H 2027. Even tier-1 CCL makers are now bidding against each other for glass cloth allocation. This converts to structural pricing power for whoever already has the allocation.

• 05 Equipment lead times are now 2 years.

  Per AMI Asia Metal Industries, CCL production line orders are booked through Q1 2028 — lead times have stretched from 8 months to 24 months. Even if a producer wanted to add capacity tomorrow, it physically cannot until 2028. This is the same equipment bottleneck story that ran through HBM in 2023–24, and it directly underwrites ASPs through that window.

• 06 Rubin CCL TAM = $275m (2026) → $2bn (2027), per GS.

  Goldman Sachs estimates the Rubin-specific CCL addressable market at $275m in 2026 ramping to $2bn in 2027 — a ~7× expansion in one year, and that is just one platform. Add AWS Trainium 2/2.5/3, Google TPU, Meta/Amazon ASICs (all on M8U/M9), and the high-speed CCL TAM is on a 19.8% CAGR through 2032 (QY Research).

A note on valuation: Several names in this report have already begun re-rating as the market starts to price in the M8/M9 spec transition. We view this re-rating as a confirmation — share prices are beginning to reflect what we have identified, while the earnings inflection is still in its early innings. The compounding effect — volume growth from content-per-rack expansion layered on top of ASP uplift from the commodity-to-advanced mix shift — has not yet been captured in consensus estimates. It is the combination of share price momentum validating the thesis and earnings compounding as volume and mix flow through to margins that we expect to propel further gains. The valuation case strengthens as the upgrade cycle matures.

Why This Is Mispriced — The Same Pattern As MLCC

Materials-Science Coverage GapCCL names are covered by Taiwan/Korea local analysts and a thin slice of materials specialists — they do not feature in mainstream US/EU AI infrastructure baskets despite being on the critical path. The same coverage gap that delayed HBM recognition in 2022–23.Buried Inside ConglomeratesThe two most exposed names globally — Panasonic (MEGTRON) and Resonac (HF/HS CCL) — sit inside large diversified industrial groups where the CCL exposure is mathematically immaterial. The pure-play universe is small, illiquid, and largely Asia-listed.Spec Ladder MisreadInvestors who treat CCL as a single market miss the bifurcation — commodity M4/FR-4 is structurally over-supplied and price-deflating, while M8/M9 is structurally short and price-rising. Aggregate CCL data masks both signals, leaving the alpha in the granularity.

The Spec Ladder — CCL Grades Across the Nvidia Roadmap

Sources: Goldman Sachs (Sep 2025 channel checks), Digitimes (Nov 2025, Apr 2026), TrendForce, GlobalTechResearch substack, Tiger Brokers, KGI Securities, Nvidia GTC 2025 / 2026 disclosures. Note: M8U = “M8 Ultra-low loss”; HVLP = “Hyper Very Low Profile” copper foil; PTH = Plated-Through-Hole. “Multi-vendor” status for M10 is deliberately engineered by Nvidia to avoid single-source dependency seen on GB300 compute tray. Specifications subject to revision; per Goldman Sachs, Rubin final CCL spec is expected to be locked around June–July 2026.

Why The Quality Is Going Up — Five Independent Drivers

• 01 Signal speed scaling.

  The roadmap from 400G to 800G to 1.6T networking forces a step-function reduction in dielectric loss (Df). M6 → M8 → M9 each cuts insertion loss by ~25–30%. There is no way to hit 1.6T signalling on FR-4 — the spec literally requires the new materials, and the upgrade is not negotiable.

• 02 Layer-count escalation.

  AI compute boards have moved from 22 layers (GB200) to 26+ (GB300) to potentially 30+ (Rubin Ultra). Each additional layer adds CCL content linearly, and the lamination process itself becomes exponentially more difficult — yield falls as layer count rises, further constraining effective supply.

• 03 Orthogonal backplane architecture.

  Rubin Ultra’s switch from copper cabling to PCB-based orthogonal backplanes is a one-time, irreversible architectural shift that creates a brand-new product category — single 78-layer M9 assemblies — with no historical demand baseline. New demand from new architecture, not substitution.

• 04 HVLP copper foil pairing.

  M8/M9 grades only deliver their performance benefit when paired with HVLP-3 / HVLP-4 ultra-low-roughness copper foil — itself a small oligopoly (Mitsui Mining & Smelting, Furukawa, Iljin Materials, Nan Ya Plastics). This compounds the qualification problem and further restricts effective production.

• 05 Q-glass / T-glass raw-material lock.

  M9 requires Q-glass (quartz cloth) or low-Dk3 glass — supplied at production scale almost exclusively by Nittobo, with limited capacity at Asahi Kasei, Shin-Etsu (SQX), and AGC. The CCL supply chain cannot grow faster than the upstream glass cloth supply chain. Nittobo new capacity arrives 2H 2027 at the earliest.

Supply-Side — The Bottleneck Stack

Scoring Framework — Three Pillars, Each Scored 1–5

We deliberately use the same three-pillar framework deployed in the LC MLCC thematic (21 May 2026). The two markets are mechanically different but structurally analogous — both are passive material inputs to AI servers, both have content explosions running into hard supply walls, and both reward the same kind of producer: technically advanced, qualified at marquee customers, and concentrated enough in the relevant segment that earnings translate cleanly into share price. Re-using the framework ensures consistency in how we evaluate the broader “AI passives” basket.

Why These Three And Not Others — Pillars Considered & Excluded

Capacity / scale. Excluded — as in the MLCC piece, capacity without qualification is unmonetisable, and qualification without technology is impossible. Adding capacity as a fourth pillar would triple-count the same underlying competence.

Q-glass / T-glass allocation. Considered but folded into the Technology / IP pillar. Glass cloth allocation is the binding constraint in 2026–2027 and is captured implicitly by a producer’s M9 qualification status — if you are qualified on M9, you have demonstrably secured glass cloth allocation. The two move together.

Geographic risk. Significant in this market — Taiwan-domicile concentration, China customer / U.S. export exposure, and Korea single-supplier risk (Doosan-Nvidia) are real. Handled as a portfolio-construction overlay, not a scoring axis, to keep the ranking focused on commercial exposure rather than macro-geopolitics.

Valuation. Deliberately excluded. The ranking screens for “what to own”; valuation answers “what to pay” — a separate question, handled in name-by-name follow-ups. Mixing them obscures the underlying exposure signal, particularly when names like Doosan have re-rated 50%+ in the past six months.

Composite & Decision Rules

Composite score = simple average of the three pillars, rounded to one decimal — identical methodology to the MLCC piece. Equal weighting preserves the gating-condition logic: a producer scoring 5 on technology but 2 on qualification (e.g. a Chinese newcomer with M9 R&D but no Nvidia design-in) is uninvestable for this thesis, and that imbalance must surface in the underlying scores rather than be hidden by weights.

AI-server torque is reported as a directional tag that combines the qualification depth, purity, and the specific platform exposure (Rubin compute exclusive vs. switch tray vs. ASIC vs. CPX) — this is the position-sizing input. The composite ranks; the torque tag sizes.

Ranking tie-breaks resolved in favour of higher purity, then higher qualification, then higher technology — consistent with the MLCC framework. Where two names are tied on all three pillars, EMC-like leadership names that hold the highest segment share break the tie.

The Top 10 — Global CCL Producers Ranked for the AI Spec-Ladder Thesis

Doosan: per Korea Investment & Securities (May 2026), “it is no exaggeration to say that the entire operating value lies with the Electronic BG” — i.e. parent valuation is functionally a CCL pure-play despite consolidated revenue including Bobcat / Enerbility / Robotics. CCL Rev Share figures are LC estimates from latest available filings, Prismark 2024 CCL market report, Goldman Sachs (Sep 2025), Digitimes channel checks, and TrendForce. Highlighted rows = top-3 conviction names. Composite = simple average of three pillars rounded to one decimal. Isola Group held by Cerberus Capital (private); listed for completeness — investable via secondary market or private credit channels only.

Top-3 Per-Name Conviction Snapshot

Note on ITEQ (6213 TT) — strong #4: ITEQ deserves recognition as a high-quality #4 — M9-grade Nvidia certified, Thailand 800G switch CCL ramp, 1.6T switch and CPO shipments accelerating from November 2025. The reason it ranks below TUC is principally the technology pillar: TUC’s M9Q is a more advanced sub-grade qualified specifically for Rubin mid-plane, and TUC has explicit AWS Trainium 2/2.5/3 design-in alongside EMC and Panasonic. ITEQ is a strong second-tier pure-play and likely sees comparable share-price impact in absolute terms.

AUTHOR

Drishtant Chakraberty, CFA

Vice President — Equity Research

Lighthouse Canton