The Board Beneath the Chip: AI's Next Bottleneck
AI CCL CAGR >100%, M9 at 7-10x FR-4 Pricing, Three Upstream Chokepoints, 6 Pure Plays Across 4 Markets, Glass Cloth Sold Out Through 2026, Q-Glass
For Indian businesses, you can check my writings at financepulse.
Disclaimer: Before we begin, I am not a registered analyst. I’m just a student trying to study different businesses. Don’t take my word for it; try to study the business independently. I might have positions in the security discussed.
Everyone knows about the GPU shortages. But the AI infrastructure buildout has a quieter checkpoint forming one layer deeper in the stack, in the laminates that make up the circuit boards connecting all of it together.
Copper clad laminates (CCL), the base material for every PCB in every AI server and every networking switch, are undergoing a generational spec upgrade that is concentrating pricing power in the hands of a small number of qualified suppliers. The market for AI grade CCL is growing at an estimated CAGR north of 100% from 2025 to 2027, and the companies that can produce M8 and M9 grade ultra low loss laminates are entering a period of margin expansion that the market has begun to notice.
As AI server architectures move to 800G and 1.6T switching speeds, the PCBs inside them need laminates with dramatically lower signal loss. Standard FR-4 laminate has a dissipation factor around 0.02. The M9 grade materials going into Nvidia’s Rubin platform and next generation ASIC servers run below 0.002, a full order of magnitude better. Standard FR-4 sells for $15-30 per square meter. M8 and M9 ultra low loss CCL commands $150-200+ per square meter, a 5-7x premium. And the supply of these materials is constrained by upstream bottlenecks in specialty glass cloth and advanced copper foil that cannot be expanded quickly.
The Spec Upgrade Cascade
What makes this setup particularly interesting is that the upgrades are happening across every board in the server rack, not just the GPU substrate. Google’s Ironwood platform (TPU v6p) is the first to massively adopt Axiom CPU boards at a 1:1 TPU-to-CPU board ratio. AWS Trainium 2.5 and 3.0 are adding entirely new PDS switch boards for TPU-CPU communication, with board-to-switch ratios ranging from 1:0.25 in air-cooled configurations to 1:0.55 in liquid cooled setups. These are net new high end CCL demand vectors that did not exist in previous server generations.
The peripheral board content growth alone adds 30-50%+ to the CCL demand per server rack versus prior generations. Layer counts are climbing from 20+ to 30+ layers per board, and each of those layers requires ultra-thin copper foil and low-loss resin systems. When you multiply the higher layer count by the higher price per layer by the greater number of boards per rack, the CCL value per AI server rack is inflecting in a way that makes the current earnings estimates for leading CCL makers look conservative.
CoWoS capacity at TSMC is ramping from roughly 375,000 wafers in 2025 to an estimated 650,000 wafers in 2026, a 73% increase. US CSP capex is projected to grow from $413 billion in 2025 to $575 billion in 2026, up 39% year over year. The CCL capacity to support that buildout is not keeping pace, particularly at the M8 and M9 grade level where qualification cycles take 12-18 months and upstream materials are already fully allocated.
Upstream Bottlenecks That Cannot Be Fixed Quickly
The constraint cascade runs deeper than CCL itself. T-glass, a specialty fine-weave glass fabric made primarily by Nittobo (3110 JP) in Japan, is fully booked through 2026. T-glass is the reinforcement layer in high end CCL that gives the material its low dielectric constant and dimensional stability. NE-glass, the next grade down, is similarly tight. Without these glass cloths, CCL manufacturers simply cannot produce M8 and M9 grade materials regardless of their resin and copper foil capacity.
Copper foil is the second chokepoint. As CCL grades climb from M6 to M8 to M9, the foil needs to get thinner and smoother. HVLP (hyper very low profile) copper foil comes in grades, each progressively harder to manufacture: HVLP3 for M8, HVLP4 and HVLP5 for M9. HVLP4 carries about a 40% price premium over HVLP3. Mitsui Kinzoku (5706 JP) controls roughly 98% of the global market for ultra-thin copper foil used in semiconductor package substrates and about 60% of the HVLP2+ copper foil going into AI server CCL.
Low-Dk resin systems, the third critical input, are also in short supply. PTFE resin used in the highest-grade CCL costs $30-50 per kilogram versus $3-5 for standard FR-4 resin. Resonac (4004 JP) announced a comprehensive 30%+ price hike effective March 2026, reflecting both cost pass-through and the value capture that comes with being one of the few qualified suppliers of AI-grade CCL precursor materials. When all three upstream inputs are constrained simultaneously, the pricing environment for finished CCL becomes extremely favorable for anyone with qualified capacity.
Who Benefits & Key Players
CCL Makers
Elite Material (2383 TT)
Elite Material is the world’s largest high speed CCL manufacturer and the broadest pure play on the AI laminate cycle. As of mid 2024, EMC held roughly 28% global share in high speed CCL, the largest position in the industry. The company makes the laminates that go into Nvidia switch trays, 800G networking infrastructure, and the majority of hyperscaler ASIC platforms. AI GPU servers represented 28-29% of revenue in mid-2024, with the rest spread across switching, networking, and ASIC applications.
EMC is expanding capacity rapidly with monthly output projected to grow from roughly 5.5 million CCL sheets at end of 2025 to 7.6 million by end of 2026, with facilities ramping in Taoyuan (acquired for NT$2.78 billion), Kunshan, Zhongshan Phase 2, and Penang Phase 2. EMC is uniquely the only Asian CCL maker with US-based sales and manufacturing, which provides optionality as tariff dynamics shift supply chain decisions.
EMC is estimated to hold roughly 80% share on AWS Trainium, close to 100% on Meta MTIA/Iris, around 50% on Google TPU, and 50-60% on 800G/1.6T switches. The company began shipping CCL for next-gen AI ASIC server PCBs in December 2025, a month ahead of TUC. EMC also has a medium-term catalyst in ABF substrate CCL, where Resonac currently holds 90%+ share and EMC is pursuing qualification as an alternative supplier.
Doosan Electro-Materials (000150 KS)
Doosan went from a mid tier Korean CCL producer to the most important laminate supplier in the Nvidia ecosystem in under a year. The exclusive Rubin CCL position is the single most valuable design win in the global CCL industry right now. The company passed GB300 certification where EMC did not, and is now projected to generate KRW 1.15 trillion in Nvidia related revenue in 2026.
Doosan’s M8-grade CCL already goes into GB200 and GB300 Bianca boards, and the DS-7409DYQ/DCQ M9Q grade product is qualified for Vera Rubin mid-plane applications, giving the company content across multiple Nvidia board types. If Doosan can leverage the Nvidia relationship to win positions on Google, Amazon, or Microsoft custom silicon platforms, the earnings trajectory extends well beyond Rubin. Doosan’s AI CCL thesis is very Nvidia dependent. Any Rubin delay, spec change, or competitive re-entry by EMC on subsequent Nvidia generations would compress the premium the market is currently assigning to the electronics division.
Taiwan Union Technology (6274 TT)
Taiwan Union Technology is an interesting name. TUC is a networking and ASIC CCL specialist that delivered 18.9% sequential revenue growth and 48.7% operating income growth in Q3 2025. Operating margins expanded to 15.7% from 12.6% the prior quarter, and the trajectory points higher as the product mix shifts toward ultra low loss grades.
TUC’s historical skew toward networking CCL rather than GPU boards has become an advantage as the industry pivots to multi-platform AI ASIC architectures. The company’s TU-953Q M9Q product is qualified for Vera Rubin mid-plane applications, and more importantly, TUC is a qualified supplier for AWS Trainium 2/2.5/3 alongside Panasonic and EMC. TUC entered next-gen ASIC supply at 20%+ share, up from less than 10% on the prior generation, a meaningful share gain that reflects the company’s improving competitive position at the high end.
TUC is expanding capacity from 2.3 million sheets per month currently to 2.6 million by 2Q26 via the Thailand ramp, with a path to 3.2 million by 2027. An overlooked growth vector is AI server power board CCL, the laminates going into PSU, DC-DC, and HVDC subsystems within the server rack, where TUC has a strong position. At roughly 22x forward earnings, TUC is the cheapest pure play high end CCL name relative to its growth rate, and the networking and ASIC diversification reduces single platform risk.
ITEQ (6213 TT)
ITEQ is the diversified second source in this supply chain. The company does not dominate any single high end application the way EMC does in ASIC or Doosan does in Nvidia GPU boards, but it has M9-grade certification from a major US AI company, a strategic Thailand manufacturing base that provides tariff advantages, and a customer mix spanning AI servers, non-AI servers, automotive, and 5G infrastructure.
ITEQ’s IT-999GSE3 M9Q grade product is in development for Vera Rubin, though the company’s market share in the top-end segment trails EMC and Doosan. Where ITEQ gets interesting is on the margin recovery. Operating margins currently run 6-8% versus EMC’s 20%, reflecting a product mix still weighted toward mid tier grades. As the Thailand plant stabilizes and AI grade volumes ramp through 2026, there is margin expansion potential. Revenue growth has been lumpier than peers, and Q3 2025 revenue actually declined 13% sequentially, but the company targets a rebound driven by 800G switch substrate and co-packaged optics demand. ITEQ is a name that benefits from the rising tide of AI CCL demand without the binary risk of a single design win, but the lower margins and smaller scale mean it participates less per dollar of industry revenue growth.
Upstream Bottleneck Players
Mitsui Mining & Smelting (5706 JP)
Mitsui Kinzoku is not a CCL maker. It is the company without which no CCL maker can produce AI grade laminates. With 98% of the global ultra-thin copper foil market (MicroThin) for semiconductor package substrates and 60% share in HVLP2+ copper foil (VSP) for AI server CCL, Mitsui sits at the most concentrated chokepoint in the entire AI materials supply chain.
AI servers are moving from 20+ layer PCBs to 30+ layers and beyond. Each layer requires copper foil. The foil needs to be thinner and smoother with each CCL grade upgrade, and the advanced surface treatment processes for HVLP4 and HVLP5 grades cannot simply be bolted onto existing production lines. Mitsui is doubling VSP copper foil capacity from 420 to 840 tonnes per month by September 2026, but industry estimates project the VSP line growing at an 80% top-line CAGR from FY3/25 to FY3/28, implying demand growth will consume that capacity quickly. Copper foil segment operating profit is estimated to compound at 43% over the same period. Without Mitsui’s copper foil, there is no M8 or M9 CCL, no advanced ABF substrates, and no advanced AI packaging.
Resonac Holdings (4004 JP)
Resonac is the world’s largest back end semiconductor materials manufacturer, covering 60-70% of back-end materials including CCL precursors, dry film resist, non-conductive film, and thermal interface materials. The 30%+ price hike effective March 2026 is the clearest pricing signal in this supply chain: Resonac can raise prices aggressively because its customers have no alternative suppliers qualified at the volumes they need.
While EMC and Doosan are primarily CCL plays, Resonac captures value across nearly every material used in advanced packaging and board fabrication. Resonac also holds over 90% of the CCL core material used in ABF substrates, a position that makes it a gatekeeper for the substrate supply chain as much as for the laminate supply chain. The risk is the conglomerate structure: petrochemical and inorganic segments which is a turnoff for me.
Nittobo (3110 JP)
Nittobo does not make CCL, but its allocation decisions on glass cloth capacity are among the most consequential variables in the AI CCL supply chain. The company holds roughly 90% of the global market for low thermal expansion glass cloth used in semiconductor package substrates and is a major supplier of NE-glass and NER-glass for AI server motherboard CCL.
Nittobo is tripling T-glass capacity via a JPY 15 billion Fukushima expansion targeting production by Q4 FY2026, but the glass cloth supply remains the binding constraint on how much high-end CCL the industry can produce through 2026.
Other Beneficiaries
Nan Ya Plastics (1303 TT) is one of the larger CCL producers globally with a roughly 8% market share, and raised CCL prices 8% in late 2025 on copper and fiberglass cost inflation. Nan Ya participates in the CCL upcycle as a broad based producer but does not have the M8/M9 positioning that commands the 5-7x pricing premium.
Taiwan Glass (1802 TT) is doubling glass cloth capacity in 2025 and targeting 30% market share by year end, positioning as a Taiwan based alternative to Japanese glass cloth supply. Taiwan Glass is an upstream beneficiary of the tightening materials market, though its investment case is more nuanced than the primary CCL producers.
Unimicron (3037 TT), while primarily an ABF substrate and PCB company rather than a CCL maker, is the key downstream consumer of the materials discussed in this piece. With roughly 18-26% ABF substrate market share (estimates vary by source) and a leading position in AI server PCBs, Unimicron’s capacity expansion and technology roadmap are directly linked to CCL and glass cloth availability. The ABF substrate market is flipping from roughly 5% oversupply in 2025 to 1% undersupply in 2026 and 9% undersupply in 2027, a dynamic that is both a validation of the CCL shortage thesis and a downstream catalyst for the companies profiled here.
Q-Glass and CoWoP: The Next Inflection
Looking into 2H26 and 2027, the CCL story gets even more interesting with the introduction of quartz glass cloth, known as Q-glass, and the potential acceleration of CoWoP (Chip-on-Wafer-on-PCB) packaging. Q-glass is 99.9% silicon dioxide versus 45-55% in current NE and NER glass, giving it dramatically better dielectric properties and thermal stability. The Nvidia Vera Rubin mid-plane will be the first PCB to use Q-glass-based CCL in mass production, with M9Q hybrid CCL shipments expected to begin in late 1Q26 or early 2Q26.
You can think of Q-glass as a step-function improvement in the glass cloth that reinforces CCL, enabling signal integrity at speeds that current glass cannot support. Q-glass suppliers are concentrated in Japan. Shin-Etsu Chemical (4063 JP) and Asahi Kasei (3407 JP) are the primary producers, with China’s Feilihua (300395 CH) as an alternative.
CoWoP pilot production could start with the Rubin platform, earlier than the market’s expectation of late 2027 or 2028. The SLP (substrate-like PCB) structure for CoWoP uses an mSAP PCB layer on top of two separate 7-layer HDI boards, all built on M9 Q-glass based CCL. If CoWoP pulls in, the demand for M9Q materials accelerates meaningfully, and the pool of qualified suppliers is even narrower than for standard M9. EMC (EM-896K3), TUC (TU-953Q), Doosan (DS-7409DYQ/DCQ), ITEQ (IT-999GSE3), Panasonic (R-5797Q), and Shengyi (Synamic 10GQ) all have M9Q products in development, but qualification status and volume readiness vary widely.
Risks to Thesis
The primary risk is a Nvidia platform delay or spec downgrade on Rubin. If Nvidia de specs certain boards from M9 to M8.5 grade, M9 demand flattens. The mid-plane and CX9 networking card specs look solid after Nvidia confirmed the Vera CPU and CX9 at CES 2026, but the GPU board spec remains fluid. A broader pullback in hyperscaler capex would hit AI CCL volumes directly, though the 2026 capex trajectory of $575 billion looks firmly committed at this point. Chinese capacity expansion from Shengyi (600183 CH) and Kingboard (1888 HK) could pressure pricing in the M6/M7 range over time, but M8 and M9 qualification barriers remain high enough at least until 2027-28.
Thanks for reading. If you want to discuss any idea, you can contact me on Twitter. Next deep dive on co-packaged optics, dropping next week soon!