One-Line Thesis
AI substrates are not simply “the next theme after GPUs.” They are a common bottleneck across the AI system bill of materials. As GPUs, CPUs, DPUs, NICs, switch ASICs and memory modules expand together, FC-BGA, MLB, CCL and low-dielectric material demand expands with them.
What This Hub Answers
| Search Question | Quick Answer | Read |
|---|---|---|
| Why does Korea have so many semiconductor substrate companies? | The U.S. and Europe emphasized design, software, tools and selected materials, while the yield-learning layer of substrate mass production compounded in Japan, Taiwan and Korea. Korea’s edge comes from Samsung / SK Hynix demand, mass-production culture and fast capex execution. | Why Korea Part 1: Semiconductor Substrates |
| What is the AI PCB investment thesis? | Rack-scale AI systems add chips and boards together. Substrates are the common denominator across GPU, CPU, DPU, NIC and switch ASIC expansion. | AI PCB and Substrate Thesis |
| Why do FC-BGA and high-speed PCB matter in chip design? | AI performance depends on moving less data and keeping memory closer to compute. That pushes FC-BGA, high-layer PCB and power-stability parts from “components” into performance bottlenecks. | After NVIDIA: the AI semiconductor bottleneck |
| What did Marvell Q1 FY2027 confirm for AI substrates? | Custom XPU, optical interconnect, scale-up networking and XPU attach are not HBM-only signals. They expand the read-through into FCBGA, high-speed substrates, MLCC, silicon capacitors and test sockets. | Marvell Q1 FY2027 and Korean semiconductors |
| What do Marvell’s trillion-dollar story and Broadcom’s Q2 call mean for Korean substrate names? | The key is connectivity, not HBM alone. Marvell confirms custom XPU and scale-up networking inside NVLink Fusion, while Broadcom must prove AI networking through its $10.7bn Q2 AI semiconductor guide, Tomahawk 6 and CPO commentary. Korea’s first-order read-through is Samsung Electro-Mechanics, Korea Circuit and Isu Petasys. | Marvell’s trillion-dollar story and the Broadcom read-through |
| Which AI PCB names still have ETF under-owned gap? | Daeduck is already in 16 ETFs, but estimated ETF ownership is only 1.95% of market cap, leaving room for AI substrate ETF expansion. Korea Circuit is a quiet weight-expansion candidate at 2.27% and recent active-ETF +0.59%p, while Simmtech is already more reflected at 3.87%. | Korean semiconductor ETF exposure strategy |
| Why is Korea Circuit a two-month relative-alpha candidate? | In the Top 50 semiconductor scan, Korea Circuit combines low ETF exposure, positive one-month foreign+institutional flow and SOCAMM+FC-BGA catalysts, making it the cleanest balanced substrate candidate. | Korean semiconductor Top 50 two-month alpha candidates |
| How do laser-source bottlenecks connect back to the AI substrate thesis? | As AI networking moves into 800G, 1.6T and CPO, the bottleneck is not only substrates and high-speed PCBs but also EML and CW-DFB laser sources. OE Solutions is Korea’s most direct EML / ELSFP proxy, but the current tape is cleaner in AI-RAN / RU names such as Solid, RFHIC and Samji Electronics. | Marvell follow-up: EML and CW-DFB laser sources |
| Is GigaVis an AI substrate stock or an equipment stock? | GigaVis does not manufacture substrates. It supplies inspection and repair equipment for high-end FC-BGA / ABF substrates. Over the latest five sessions, retail sold KRW 19.46B while foreigners plus institutions bought KRW 19.33B, but after the June 4 +11.2% move the better setup is confirmation around KRW 135,000 or a breakout above KRW 144,400. | GigaVis follow-up: AI FC-BGA yield bottleneck |
| If memory ASP peaks out in 2027, do AI substrates peak too? | No. J.P. Morgan’s peak-out applies mainly to commodity-memory price growth. If Goldman’s 2030 token demand holds, server/rack complexity keeps rising, so substrate area, power integrity, FC-BGA and MLCC demand can persist independently of memory-price momentum. | Goldman token demand vs J.P. Morgan memory ASP peak-out |
| Which Korean AI PCB names should investors compare? | The ecosystem is broader than board manufacturers. Compare Samsung Electro-Mechanics, Daeduck, Isu Petasys, Simmtech and Korea Circuit with Doosan Electronic BG, Kolon Industries, Pamicell, TLB and Taesung. | Korea AI PCB Ecosystem: 10 Companies |
| Is Samsung Electro-Mechanics an AI substrate stock? | Samsung Electro-Mechanics combines FC-BGA and MLCC exposure, making it a premium Korean node in AI server substrates and power-stability components. | Samsung Electro-Mechanics AI Infrastructure Re-Rating |
| What does Mirae Asset’s KRW 1.3M Samsung Electro-Mechanics target mean? | The KRW 1.3M target applies a 37x 2017-style MLCC shortage multiple to 2028F EPS. With the stock already above a normal upcycle multiple, the key evidence is MLCC pricing and FC-BGA margin realization. | Samsung Electro-Mechanics KRW 1.3M Target Analysis |
| Why does Mirae Asset’s KRW 2.8M Samsung Electro-Mechanics target need 2029 proof? | The June 1 target is 2029F EPS of KRW 93,242 multiplied by 30x. It is not multiple expansion; it is a 2027-2029 earnings-duration call on MLCC, FC-BGA and Si-Cap. SEMCO now has to prove stronger duration than memory. | Samsung Electro-Mechanics KRW 2.8M Follow-Up |
| What does SEMCO’s KRW 1.5T silicon-capacitor win change? | A reclassification event — from “MLCC / substrate maker” to “AI-package power-integrity component company.” KRW 1.557T / 2 yrs (~6.9% of FY25 revenue), recognized 2027–28. SEMCO is the only player globally able to bundle FC-BGA + MLCC + Si-Cap. But the May 20 +7.5% / May 21 pre-market KRW 1,109,000 means the price already reflects much of it. | Si-Cap KRW 1.5T — Reclassification Event |
| How do Intel EMIB-T and SEMCO silicon capacitors connect? | EMIB-T adds TSV-enabled power delivery to 2.5D chiplet / HBM packaging, while Si-Cap is a die-near PDN component that reduces voltage noise inside AI packages. The official fact is SEMCO’s KRW 1.5T Si-Cap contract; the Google TPU / EMIB-T link remains an inference. | SEMCO Si-Cap and Intel EMIB-T |
| What does SEMCO’s KRW 100T market cap mean for the AI substrate thesis? | It is a signal that SEMCO is being treated as an AI package power-integrity plus FC-BGA platform, not just a substrate stock. Hyundai Motor is mathematically reachable, but a durable KRW 150T valuation needs a 2028 OP path around KRW 4T and group margin near 20%. | SEMCO at KRW 100T |
| Can SEMCO command a premium over Murata and Ibiden after reaching Hyundai Motor’s weight class? | The premium thesis exists: a long-term Si-Cap contract, MLCC/FC-BGA/Si-Cap bundling, and reclassification as an AI package PDN supplier. But at KRW 138T market cap, SEMCO is already at 94% of Murata’s size; further upside needs repeat orders and a 2027–2028 earnings bridge. | SEMCO at KRW 138T and peer premium |
| Should MLCC and FC-BGA get GPU/HBM-style multiples? | They are real bottlenecks, but not platform monopolies. FC-BGA and MLCC still carry capex, depreciation, generic-cycle and qualification risks. The SEMCO thesis is directionally right, but the current price demands repeat orders and high-margin evidence. | AI Infrastructure Multiple Map |
| How does the passive-component bottleneck connect to AI substrates? | FC-BGA is the physical floor for the chip package, while MLCCs, silicon capacitors and inductors stabilize the power riding on that floor. Larger GPU/HBM packages make substrates and power integrity one system bottleneck. | AI server passive-component bottleneck |
| Why is Pamicell connected to AI CCL? | Pamicell does not make PCBs. The thesis is upstream material exposure through low-dielectric materials tied to Doosan Electronic BG’s high-end CCL cycle. | Pamicell Part 1 |
| Is Pamicell a one-quarter event or a cycle? | KRX reclassification, Doosan Electronic BG margins and CCL undersupply commentary extend the frame from one quarter to a 12-24 month industry cycle. | Pamicell Part 2 |
| Has Pamicell already been reclassified by the market as an AI substrate materials stock? | Fundamentally it is close to an electronic materials stock, but the flow data says not yet. From May 4 to May 18, Pamicell fell -17.9% with only KRW 1.6bn of foreign + institutional net buying, while Daeduck rose +22.7% with KRW 64.8bn. That recognition gap is the key. | Pamicell flow analysis |
| How would overseas retail investors discover Korean AI substrate stocks? | Easier IBKR / foreign-broker access can turn AI substrates from a local supply-chain theme into a searchable global small-mid-cap cluster. | Foreign Investor Korea Stocks Hub |
| How do HBM and substrates connect? | HBM is the memory bottleneck. Substrates are the physical bottleneck for mounting and connecting GPUs, CPUs and networking chips. | HBM / Korea Semiconductor Hub |
| Why is Haesung DS in the PCB cluster? | Haesung DS is not a PCB or substrate manufacturer per se. It is the eleventh name in the cluster: global lead-frame #2 + DDR memory package substrate + the optional AI heat-spreader second-source angle. The only non-PCB axis in the cluster. | Haesung DS: Lead-Frame Compounder Pivoting to AI Heat-Spreader Second-Source |
| How do the Korean optical / CPO names map? | Of seven listed names only OE Solutions is genuinely close to CPO via its ELSFP external laser source. The new laser-source follow-up extends that logic into EML, CW-DFB and external laser sources, while warning that the 3-4 month optical leaders still need a clean flow turn. | Korea Optical & CPO Value Chain — Seven Companies · Marvell follow-up: EML and CW-DFB laser sources |
| Should substrates and test sockets be grouped together? | No. Both benefit from AI back-end, but substrates are a CAPEX / volume beta (OPM ~12–15%) while test sockets are a consumables / chip-complexity beta (OPM ~35–47%). Margin structures differ by ~3x. Substrates for short-term momentum, test sockets for 1–2 year holding. | Substrate vs Test Socket Comparison |
Start Here
| Step | Question | Read |
|---|---|---|
| 1 | Why does Korea have so many semiconductor substrate companies? | Why Korea Part 1: Semiconductor Substrates |
| 2 | Why are AI substrates a common bottleneck rather than a late-cycle theme? | AI PCB and Substrate Thesis |
| 2+ | Why do data movement, FC-BGA and high-speed PCBs become AI chip-design bottlenecks? | After NVIDIA: the AI semiconductor bottleneck |
| 2++ | What does Marvell’s Q1 FY2027 say about custom ASIC, optical and substrate bottlenecks? | Marvell Q1 FY2027 and Korean semiconductors |
| 2+++ | How do Marvell’s trillion-dollar story and Broadcom’s Q2 call reprice connectivity and substrate bottlenecks in Korea? | Marvell’s trillion-dollar story and the Broadcom read-through |
| 2+++A+ | What does ETF exposure say about Daeduck, Korea Circuit and Simmtech? | Korean semiconductor ETF exposure strategy |
| 2+++A++ | Why does Korea Circuit screen as a two-month alpha candidate versus Samsung and SK Hynix? | Korean semiconductor Top 50 two-month alpha candidates |
| 2+++A | If the connectivity bottleneck moves into laser sources, how should investors read EML, CW-DFB, ELSFP and OE Solutions? | Marvell follow-up: EML and CW-DFB laser sources |
| 2+++B | If GigaVis is not a substrate maker, where does it sit in the AI FC-BGA value chain? | GigaVis follow-up: AI FC-BGA yield bottleneck |
| 2++++ | Once AI usage gets a price, how do FC-BGA and substrates connect to the cost-per-token thesis? | AI token futures and cost per token thesis |
| 2+++++ | Why do MLCC and FC-BGA deserve different multiples from GPU/HBM inside the same AI infrastructure cycle? | AI Infrastructure Multiple Map |
| 3 | How should investors split the ten-company Korean AI PCB ecosystem? | Korea AI PCB Ecosystem: 10 Companies |
| 4 | Which large Korean name anchors FC-BGA and MLCC? | Samsung Electro-Mechanics: Why the Stock Doubled in 90 Days |
| 5 | What valuation frame sits behind Mirae Asset’s KRW 1.3M Samsung Electro-Mechanics target? | Samsung Electro-Mechanics KRW 1.3M Target Analysis |
| 5+ | Why does Mirae Asset’s KRW 2.8M target turn into a 2029 earnings proof problem? | Samsung Electro-Mechanics KRW 2.8M Follow-Up |
| 5++ | SEMCO Si-Cap KRW 1.5T contract — what changes, and is the price already reflecting it? | Si-Cap KRW 1.5T Supply Contract (May 21) |
| 5+++ | Is SEMCO’s KRW 100T market cap a Murata-style rerating or a price that needs earnings to catch up? | SEMCO at KRW 100T |
| 5++++ | Why do components smaller than the GPU become AI server bottlenecks? | AI server passive-component bottleneck |
| 5+++++ | How do EMIB-T and silicon capacitors change the AI package power-delivery bottleneck? | SEMCO Si-Cap and Intel EMIB-T |
| 5++++++ | Can SEMCO command a premium over Murata and Ibiden after reaching Hyundai Motor’s weight class? | SEMCO at KRW 138T and peer premium |
| 6 | Why is Pamicell a Doosan Electronic BG CCL material proxy? | Pamicell Analysis: AI CCL Materials Re-Rating |
| 7 | Has the Pamicell thesis moved into a 12-24 month cycle? | Pamicell Part 2: AI CCL Materials and Industry Cycle |
| 8 | Why are Pamicell’s fundamentals and flow-based market recognition still misaligned? | Pamicell Flow Analysis: Between Doosan and ST Pharm; Daeduck Is Still Far Away |
| 9 | How does Big Tech AI capex reach the Korean supply chain? | Samsung Electronics vs Samsung Electro-Mechanics: Big Tech AI Capex Read-Through |
| 10 | Where is the eleventh cluster name with the unique lead-frame-to-AI-thermal angle? | Haesung DS: Lead-Frame Compounder Pivoting to AI Heat-Spreader Second-Source |
| 11 | What’s the next AI data-center bottleneck after substrate — and which Korean names are real vs. thematic? | Korea Optical & CPO Value Chain — Only OE Solutions Is Truly Close to CPO |
| 11+ | What sits inside the CPO bottleneck — EML, CW-DFB and ELSFP external laser sources? | Marvell follow-up: EML and CW-DFB laser sources |
| 12 | Why is it wrong to group substrates and test sockets as one “AI back-end theme”? | Two Betas of AI Back-End: Substrate vs Test Socket |
AI Substrate Value-Chain Map
| Layer | Korean Names | Investment Question |
|---|---|---|
| Package substrates | Samsung Electro-Mechanics, Daeduck Electronics, Korea Circuit | FC-BGA and advanced package substrates for GPU, CPU and ASIC packages; the physical floor that shortens the memory-to-compute path |
| Substrate inspection / repair equipment | GigaVis | AOI, VRS, AOR, FA and software for AI GPU / ASIC FC-BGA and ABF substrates; the yield-control layer that becomes more valuable as substrates get finer |
| MLB and module boards | Isu Petasys, Daeduck Electronics, TLB, Simmtech, Korea Circuit | Server boards, network switch boards, memory modules and SoCAMM; the high-speed roads for data movement inside and across racks |
| Memory package substrate + lead frame | Haesung DS | DDR4/DDR5 R2R substrate + global lead-frame #2; automotive + memory dual axis |
| CCL body | Doosan Electronic BG | High-end copper-clad laminate for AI servers and high-speed networking |
| Low-dielectric materials | Kolon Industries, Pamicell | mPPO, low-loss resin, hardener and upstream material exposure |
| Power-stability components | Samsung Electro-Mechanics and MLCC peers | Rising MLCC content in AI servers and networking equipment; silicon capacitors are not an MLCC replacement but a closer-to-GPU/HBM complement for package power integrity. See the passive-component bottleneck explainer |
| AI package thermal | Haesung DS (option) | Heat slug / heat spreader second-source optionality in the Shinko / Honeywell / Jentech / I-Chiun ~85%-concentrated market |
| AI optical / CPO | OE Solutions (core), Optocore / Daehan Optical / BWE / WooriRo / Lycom / Coset (thematic), Solid / RFHIC / Samji Electronics as AI-RAN/RU flow watches | 800G/1.6T transceivers, CPO ELSFP external laser source, EML / CW-DFB lasers and optical fiber — same system-BOM bottleneck logic as substrate, but OE Solutions still needs customer qualification and repeat orders, while the 3-4 month optical leaders need a cleaner flow turn |
FAQ — AI Substrates and Korean PCB
What are AI substrates?
In this hub, AI substrates refers to the package substrates and PCBs that mount and connect AI accelerators, server CPUs, networking ASICs, DPUs and memory modules. The key terms are FC-BGA, MLB, CCL and low-dielectric materials.
Are FC-BGA and CCL the same thing?
No. FC-BGA is a high-performance package substrate for chips. CCL is copper-clad laminate, the base material for PCBs. High-speed AI systems need better CCL and lower-loss materials to preserve signal integrity.
Does Pamicell make PCBs?
No. Pamicell is not a PCB manufacturer. It appears in this hub because it is an upstream low-dielectric materials supplier linked to Doosan Electronic BG’s high-end CCL cycle.
Does the market already trade Pamicell as an AI substrate materials stock?
Not yet. The 1Q26 low-dielectric materials mix and KRX reclassification support the electronic-materials thesis, but May 4-18 flow and price correlation still look closer to Doosan, ST Pharm and Kolon Industries than to Daeduck. A higher Daeduck correlation and foreign + institutional net buying above roughly 3% of turnover would be the early evidence of reclassification.
Why is Samsung Electro-Mechanics central?
Samsung Electro-Mechanics combines FC-BGA and MLCC exposure. AI servers require both high-performance package substrates and more power-stability components.
This hub is updated as new AI PCB, substrate and CCL coverage publishes. For memory products and KOSPI concentration, see the HBM Hub. For semiconductor equipment and IP, see the Semiconductor Equipment & IP Hub.
Disclaimer: For research and information purposes only. Not investment advice. Names cited are for analytical illustration; readers should perform their own due diligence and consult licensed advisors before any investment decision.