Overview / thesis

A modern AI training baseboard ships with somewhere between 15,000 and 25,000 multilayer ceramic capacitors on it. A flagship smartphone ships with about 1,000. A battery-electric vehicle uses 3,000 to 5,000 — TDK's own per-powertrain data puts it higher, at roughly 10,000 for a full BEV versus 5,000 for an ICE car. The MLCC is the single most populous active electronic component on every modern circuit board, and its supply is gated by three companies in Japan, one chaebol in Korea, and roughly five powder suppliers globally. The cycle bottomed in late 2024, pricing turned in May 2026, and the high-end of the curve is structurally undersupplied through at least 2027. Almost every claim on this sector page backs into one of those four facts.

The four-fact thesis

The whole MLCC investment case compresses into four statements that recur across every source consulted (the FK-style primer, the Dongguan Securities initiation, BofA/JPM/GS sell-side, the Banyan Lane and Collyer Bridge Substack threads):

1. AI server is a brand-new demand vector that did not exist in any prior MLCC cycle. Content per accelerator board has stepped up 2–3x versus the 2024 generation, on a hyperscaler installed base that is still growing ~50% YoY.
2. The cycle has turned, not just bottomed. The trough was Q3/Q4 2024; we are 12–18 months into a typical 18–30-month up-cycle. Pricing returned to the high-end first (Q4 2025) and the commodity tier last (May 2026). Both ends moving at once is what a confirmed inflection looks like.
3. The high-end is structurally undersupplied and cannot be fixed inside the thesis horizon. Capacity additions take 24–30 months from capex commitment, qualification adds 6–12 more, and the dielectric powder comes from 3–4 qualified suppliers globally.
4. There is an under-priced upstream choke point — sub-100nm barium titanate powder — sitting at a fraction of the downstream MLCC market caps.

Why it matters right now (the 'so what')

The opening hook is straightforward. AI hyperscaler capex grew roughly 50% year-over-year into 2026, and the bill of materials per accelerator board added voltage rails as Hopper became Blackwell and Blackwell becomes Rubin. Each rail needs decoupling. Decoupling is what an MLCC does. The result is that AI-server MLCC content per system is up two to three times versus the 2024 generation, on a hyperscaler installed base that is still growing. Murata, Samsung Electro-Mechanics, Taiyo Yuden, TDK, and Yageo all raised prices in May 2026 — Murata 15% to 35% on the high-end (effective April 1, 2026, on AI-server and auto-grade), Yageo 10% to 20% on auto-grade, Taiyo Yuden 6% to 13% on commodity, SEMCO signaling 5–10%. Pricing power returned to both ends of the curve at once, which is what a true cycle inflection looks like.

The structural setup is real, not forecast. The cycle is happening. Five independent sources converged on the same read in the third week of May 2026 — BofA, the sell-side Murata-vs-Yageo desk note, a six-dimension global passives basket framework, Goldman Sachs on Yageo, and Banyan Lane Capital's Substack thesis on Murata. Murata's capacitor backlog surged +89.5% to ¥269.2B, with the computers/servers application up +28.4% YoY, confirming demand is running roughly 2x supply at the high end. Customer inquiries were running at ~2x supply capacity (Bloomberg, Feb 2026), with production lines at 98% utilization. Murata's President Nakajima called capacity "still very much insufficient" and announced the JPY 80B (~$500M) emergency MLCC capex decision made "in haste" alongside the largest buyback in company history (JPY 150B).

Sizing the opportunity — TAM and growth

There is no single agreed TAM figure across the sources; they bracket a range. Reproduce the numbers rather than average them away:

Metric	Figure	Source
MLCC TAM (2025)	$16.3B	JPM industry model
MLCC TAM (2030E)	$25B (+11% CAGR)	JPM industry model
MLCC market (cycle midpoint)	~$15B revenue	FK primer
MLCC market (2030E)	$22–25B	FK primer
Global capacitor market (all types)	$41.23B	Capacitor Dossier (Apr 2026)
Broader passive-component complex (caps + resistors + inductors + transformers + filters)	~$30B annual revenue	FK primer
AI-server mix of MLCC TAM	1.1% (2025) → 4% (2030E)	JPM model
AI-server share of MLCC revenue	~5–10% today → 20%+ by 2027	FK primer
Unit volume (today)	~5.5–6 trillion MLCC/yr	FK primer
Unit volume (2030E)	~8–10 trillion/yr	FK primer
Server capacitor demand CAGR (2025–2030)	30%	Murata IR
AI MLCC demand CAGR	80%+	Banyan Lane / substack scrape
800V/high-voltage EV MLCC growth	17.3% CAGR through 2025 (3x overall cap market)	technical sections / Dongguan
Strategic-tier MLCC growth	12–17% CAGR	Global Passives Basket / Capacitor Dossier
Server MLCC market (2025)	~$1.3B = $600M AI + $700M general	Zephyr (29 May 2026) via Archetype
General-server MLCC CAGR	30–40% (agentic AI → CPU content)	Zephyr / Archetype
Smartphone / mobile MLCC growth	negative through 2026–27	Zephyr / Archetype

The mix-shift detail is the load-bearing part of the TAM story. Strategic MLCCs (high-voltage, polymer electrolytic, ultra-high-CV) are 35–40% of market value but 60–70% of industry profit, growing 12–17% CAGR with 30–80% price premiums and 24–47-week lead times, while commodity MLCCs face -10–20% pricing pressure from Chinese capacity. This bifurcation is why Murata's FY27 operating-profit guide is +34.8% on only +7.1% revenue growth — the profit pool is migrating up the curve faster than units. ASP weighted by mix is roughly flat-to-slightly-up over the cycle because high-end mix shift offsets commodity ASP compression. ASPs span four orders of magnitude: a commodity 0402 X5R goes for fractions of a cent, while a 0805 100V X8R auto-grade or a 008004 10µF AI-server part commands 10–50 cents.

Murata's own demand build for AI server: Medium-Term Direction 2027 put AI servers at 10,000–20,000 MLCC per baseboard; the revised FY25–FY30 estimate is 15,000–25,000 per baseboard, a 3.3x increase in demand versus FY2025, with a separate 2.0x lift from miniaturization and high-capacitance demand. The GB300 platform requires ~30,000 MLCCs per baseboard — about thirty times a mobile phone and three times an automobile — with a single rack consuming as many as 440,000 individual capacitors. The generational BOM step-up is the sharpest sizing input: Morgan Stanley's NVL72 BOM puts MLCC content at $1,530 per GB300 (Blackwell) rising to $4,320 per VR200 (Vera Rubin), +182% generation-on-generation.

The three structural demand drivers

AI server — the new vector. An NVIDIA GB200 NVL72 rack carries 36 GB200 baseboards, each with 25k-class MLCC content; close to 1 million MLCCs per rack versus roughly 10,000 in a standard 2U dual-socket Xeon server. The demand sits in the high-cap, small-form-factor, low-ESR tier where Murata, Samsung E-M, and Taiyo Yuden have structural pricing power. Hyperscalers (Google, Meta, Microsoft, NVIDIA) have locked in 60–70% of available high-CV capacity on multi-year long-term agreements. Dongguan put the AI data center MLCC use case plainly: AI chips run at low voltage with high current and large instantaneous fluctuations, requiring high-capacitance MLCCs for local decoupling and instantaneous power replenishment.

Automotive electrification — the multi-year layer underneath. Electrification roughly doubles MLCC content because EVs add inverters, on-board chargers, battery management, ADAS, and infotainment. The TDK per-powertrain data: ICE 5,000 / HEV 6,000 / PHEV 8,000 / BEV 10,000 MLCCs. The 800V architecture transition is an independent second driver that continues even if AI capex moderates — 23 new 800V production models launched in 2025, 50+ planned through 2027, with OEMs now specifying AEC-Q200 across 100% of capacitors in 800V powertrain circuits. High-voltage MLCC (1kV-rated, 100nF class) grew at 17.3% CAGR through 2025.

Smartphone high-cap mix shift — the slow secular layer. ~1.4B handsets per year × ~1,000 MLCC each ≈ 1.4 trillion units. Flagship content has grown ~50% over the last decade; the high-cap (≥10µF) share of phone MLCC content has roughly doubled in five years as 5G modems, multi-camera ISPs, and on-device AI accelerators added rails.

The aggregate result: MLCC unit demand growth is structurally above GDP for the foreseeable future, plausibly mid-to-high single digits compounded, with the value pool growing faster than units because of mix shift. Adjacent emerging demand — humanoid robotics modules and 01005 MLCCs in smart glasses (150–200 cells per pair per TrendForce) — is real but not yet material to the TAM.

Why the structure holds — concentration and the moat

This is one of the most concentrated component markets in semiconductors, comparable to NAND/DRAM in degree of consolidation but with a more stable competitive ranking. The Tier-1 four (Murata + Samsung E-M + TDK + Taiyo Yuden) hold roughly 77% of global MLCC revenue per the FK primer; Dongguan's 2024 numbers put the top 5 (adding Kyocera) at 83.6%, with Murata alone at 31.8% (FK uses ~34%) and SEMCO at 11.8–22% depending on the cut. In automotive-grade specifically (2022 data), Murata holds 47.0%, SEMCO 14.6%, TDK 14.6%, Taiyo Yuden 10.4%. In AI-server MLCC, Samsung Electro-Mechanics holds ~40% of the high-end segment (Maeil Business Newspaper / Digitimes).

The moat is process IP that compounds over decades, concentrated at two manufacturing crown jewels — electrode printing and co-firing — plus the upstream powder. Only Murata, TDK, Samsung Electro-Mechanics, and Kyocera AVX hold production-scale capability for ultra-high-CV designs above 47µF. The yield math is brutal: 99.5% yield per layer across 1,000 layers compounds to a 0.7% cumulative body yield, so reaching commercially viable >85% yields requires simultaneous mastery of every stage, which took these firms decades. A new entrant with comparable equipment (which they largely cannot even source — Murata's sintering furnaces are proprietary) would need 5–10 years of yield learning at the high end. The AEC-Q200 automotive qualification cycle (12–18 months minimum, decades of field-reliability data) reinforces the structure on top of the manufacturing moat.

The central debate

The bull/bear axis the sources keep circling is whether the headline "price hikes" are real list-price increases, mix shift, or allocation premiums — and, relatedly, how big the AI BOM TAM actually is. Three framings coexist in the public record:

• Framing A — real price hikes: Murata 15–35%, Taiyo Yuden 6–13%, SEMCO 5–10%, with Banyan Lane modeling 4–7 rounds through FY30 and AI MLCC ASP potentially doubling+ from trough (the 2018 cycle ran 4 discrete rounds, ASP roughly doubled).
• Framing B — mix shift, not price: Taiyo Yuden's own sell-side roundtable (May 22–23) said "no MLCC price hikes this fiscal year, but FY3/28 is a possibility"; ASP lift comes from migration to smaller, higher-cap parts (47µF moving from 2012/1608 to 1005, 100µF from 2012 to 1608 in 2H 2026).
• Framing C — allocation premiums: the claim that "Murata hasn't really raised prices, customers pay premiums to secure capacity." Banyan Lane's public version: the shortage "allocates pain to tier-2/tier-3 customers while tier-1 customers pay through the nose to maintain supply." Murata's multi-year LTAs give hyperscalers full allocation while mid-tier OEMs face shortages — echoing 2018, when line-stops hit Tier-2/3 OEMs (GoPro under-produced cameras, Sony PS4 was constrained). TrendForce notes AI customers use 2–3-year planning models, making the price trend more durable.

The second debate, surfaced in the Collyer Bridge chat, is how to size the TAM correctly. Subscriber "Zrave" (reposted by Collyer as his anchor thesis): "MLCC TAM is 99.99% outside that BOM... PSU passives scale with wattage... Lifetime Rubin MLCC BOM won't cover a single quarter of Yageo revs." The point: bulls who fixate on per-rack MLCC content ($4.3K for VR200) miss that the real TAM is wattage-scaled passive demand across the whole datacenter, not any single GPU rack. The datacenter-level BOM for power and cooling looks very different from the rack-level BOM.

What can break it

The bull case does not break on any single risk in isolation; it breaks if two or more arrive within 12 months. Ranked:

1. AI capex digestion in 2027 — the largest risk. The thesis embeds AI-server BOM expansion continuing through 2027. Any of AMZN/MSFT/META/GOOG/ORCL cutting 2026/2027 capex by 15%+ would cascade through the complex within days, as smartphone misses did in the 2017–2018 cycle.
2. BEV demand softening — flattens the multi-year auto layer underneath the AI cycle; shortens the tailwind by 6–12 months rather than ending the cycle.
3. China commodity flood reaching mid-cap — Chinese new entrants (Fenghua, Sunlord, Sanhuan, Guoci) currently pressure only the commodity 0603/0402 tier; if they credibly close the gap on 0201 high-cap, Tier-2 Taiwanese pricing power compresses. A 5–10-year question, not a 2026 one.

Secondary: JPY/KRW volatility (a swing through ¥140/USD compresses Murata/Taiyo Yuden reported earnings by mid-single-digit percentages), and a Chinese rare-earth export restriction on dysprosium/yttrium/holmium dopants that could complicate non-Chinese production within 6–12 months.

The non-consensus angle

The named winners are priced. The angle with alpha potential is the upstream choke point the market has not priced: sub-100nm barium titanate powder, which cannot be bypassed for high-cap small-form-factor MLCC. Powder market share per Dongguan (Sakai 28%, Ferro 20%, Nippon Chemical 14%, Guoci 10%, Fuji Titanium 9%, with the top 5 at 81%) corrects the "40–50% merchant share" Sakai self-reports. The technology gap is concrete: Japanese hydrothermal BaTiO3 reaches 80–100nm particle size enabling 1,200 layers at 0.5–0.6µm (Murata 1,600 layers); Chinese producers reach only 120–150nm and 800 layers at 1–2µm. Company-level treatment lives on the ticker pages: see 4078.T (Sakai Chemical), 4092.T (Nippon Chemical Industrial), and 6173.TWO (Prosperity Dielectrics) for the powder layer, and 6981.T, 009150.KS, 6976.T, 6762.T, 2327.TW for the manufacturer tier. The adjacent non-MLCC capacitor categories — silicon capacitors and supercapacitors/LIC — are covered separately; see 6531.TW (AP Memory) and 7220.T (Musashi) on those threads.

China's domestic-substitution overlay

Distinct from the investment thesis but recurring in the Chinese sell-side: in 2025 China imported 2.56 trillion MLCCs worth US$6.179B (avg US$2.41/thousand units imported vs US$2.11/thousand exported — the price gap reflects imported high-end value). If 50% of import volume were domesticated, the substitution opportunity reaches 1.28 trillion units. This is the policy tailwind behind state-funded Fenghua and Sanhuan, and the structural source of long-run pressure on Tier-2 Taiwanese commodity pricing.

How it works

An MLCC is the single most populous active-circuit component on every modern board, and the reason it exists comes down to one physical problem: a silicon chip needs power delivered as a clean, instantaneous voltage, and the power supply is too far away and too slow to react. A modern GPU like an NVIDIA GB200 module swings hundreds of amps in nanoseconds across thousands of pins — context switching, memory access, and tensor-core activation create transient current swings approaching 100 amperes per microsecond (100A/µs), and those transients cause voltage droops that corrupt computation if not compensated within nanoseconds. The package and board cannot react that fast on their own. The fix is to park enough stored charge as close to the silicon as physically possible, ready to discharge into the chip the instant current spikes. That charge lives in capacitors. The useful analogy: a chip is a thirsty city, the power supply is a distant reservoir with long pipes that lose pressure on a demand spike, and capacitors are the water towers placed inside the city itself, full and ready to dump water the moment a hydrant opens. Modern AI accelerators are not cities, they are wildfires — they need water towers stacked on top of each other. MLCCs replaced the older tantalum, aluminum electrolytic, and film capacitors in nearly every digital application because they could be made small enough to tile a board, cheap enough to use by the thousand, and fast enough to source charge in picoseconds.

The physics: C = εA/d

An MLCC is mechanically simple. It is a sandwich of alternating ceramic dielectric layers and metal electrode layers, fired into a single monolithic block smaller than a grain of rice, with metal terminations at the two ends. The governing equation is the standard parallel-plate capacitor relation, capacitance C = εA/d, where ε is the dielectric constant (permittivity) of the ceramic, A is the effective electrode area, and d is the gap between electrodes (the dielectric thickness). To pack more capacitance into a smaller body you do three things simultaneously, and all three are hard:

• Raise ε by using a high-permittivity ceramic. Bounded by chemistry.
• Raise A by stacking more layers in parallel. Bounded by yield — every additional layer is another chance for a short or a crack.
• Lower d by making each dielectric layer thinner. Bounded by powder physics — you cannot make a dielectric layer thinner than the grain size of the ceramic powder you are using.

Modern high-cap parts stack 500 to 1,200 dielectric layers (Murata reaches 1,600) at roughly 0.5–0.8µm thickness in a body smaller than a grain of rice. Murata data puts the capacitance-to-volume ratio at 1µF/mm³ in 1996 rising to 40µF/mm³ by 2020. Maximum capacitance per body has roughly doubled every three years; minimum body size has halved roughly every five — a cleaner shrink cadence than logic's Moore's Law because each generation needs ceramic, equipment, and process retooling rather than a new lithography platform.

Barium titanate: the ceramic moat

The dominant dielectric is barium titanate (BaTiO3), a perovskite crystal with a dielectric constant around 3,000. For scale: air is 1, silicon dioxide about 3.9, alumina about 9 — so BaTiO3 is roughly 1,000× the air gap in a textbook capacitor. Nothing else in the materials catalog combines BaTiO3's permittivity with thermal stability and manufacturability at sub-100nm grain size. Doping the lattice with rare earths (yttrium, dysprosium, holmium) flattens the temperature-capacitance curve and improves voltage stability — these doping recipes are exactly what create the X7R, X5R, X8R, and C0G temperature classifications engineers select by application.

Barium titanate is the moat because the powder is not a commodity chemical: particle size, shape uniformity, and chemical purity directly determine the performance ceiling of the finished part. Particle size matters because you cannot make a dielectric layer thinner than the grain size of the powder. Japanese producers (Sakai Chemical, Nippon Chemical) achieve average BaTiO3 particle sizes of 80–100nm via hydrothermal synthesis; Chinese producers (Guoci Materials, Fenghua) currently reach only 120–150nm. That 40–50nm gap has concrete consequences: Japanese makers stack up to 1,200 layers at 0.5–0.6µm dielectric thickness (Murata 1,600 layers), while Chinese makers average 800 layers at 1–2µm. In the same body size, finer powder means more layers means more capacitance. The result: Murata can produce a 220µF MLCC in a 1206 case (3.2mm × 1.6mm) while the best Chinese maker maxes at about 100µF in the same package.

Four preparation methods exist for barium titanate, each with a distinct tradeoff:

Method	Cost	Application scope	Component control	Shape control	Particle size	Calcination	Grinding
Solid-state (ball-mill BaCO3 + TiO2, calcine at 1,100–1,200°C)	Low	Commercial	Poor	Poor	>1.0µm	Required	Required
Co-precipitation (dissolve Ba/Ti salts, precipitate at controlled pH)	Moderate	Commercial/experimental	Good	Moderate	<1.0µm	Required	Required
Sol-gel	High	Experimental	Excellent	Moderate	<1.0µm	Required	Required
Hydrothermal (crystallize from aqueous solution at 150–250°C under pressure)	High	Commercial/experimental	Excellent	Good	<1.0µm (sub-100nm spherical)	Not required	Not required

The hydrothermal method is the high-end route: BaTiO3 crystallizes directly from aqueous solution, producing sub-100nm spherical particles with excellent uniformity and no calcination or grinding (which preserves particle shape). This is Sakai Chemical's core IP and the only viable route to the thinnest dielectric layers needed for high-cap AI-server parts; it commands a significant price premium. Guoci Materials uses a hydrothermal variant too. The next-generation dielectric problem — finding a chemistry that holds ε above 2,000 at sub-50nm grain size — is the long-term R&D frontier, because BaTiO3's own dielectric constant drops as grain size shrinks below 100nm (the tetragonal phase transitions toward cubic). The very physics that makes BaTiO3 useful starts to work against the engineer at the smallest scales. See 4078 (Sakai Chemical) and 4092 (Nippon Chemical) for the upstream pure-play cases.

Powder market share (Dongguan Securities): Sakai Chemical 28%, Ferro (US) 20%, Nippon Chemical 14%, Guoci Materials (China) 10%, Fuji Titanium 9%, Kyoritsu 8%, Toho 6%, other 5%. Top five hold 81% — a highly concentrated layer. (Note: this corrects Sakai's own "40–50% merchant share" self-report, which is share of merchant supply, not total market.) Murata produces its own BaTiO3 internally and is structurally self-sufficient; Murata also formed the MF Material JV with Ishihara Sangyo and Fuji Titanium in September 2023 to internalize barium titanate supply, structurally diluting external powder dependence over time. Samsung Electro-Mechanics, Taiyo Yuden, TDK, Yageo, Walsin, and Holy Stone buy a meaningful fraction of high-end powder externally.

By production route — a different cut from the company-share split above: the BaTiO3 market divides roughly solid-state 61.8% / oxalate 21.4% / hydrothermal 13.5% (Archetype Capital, 2026-06-01). The two premium routes — hydrothermal (Sakai Chemical's core IP) and oxalate (Nippon Chemical's) — are only ~35% of supply, and AI-server demand (high-capacitance, high-reliability MLCC) skews toward exactly these two, so the premium powder layer tightens faster than the headline market grows. Archetype frames this as a "capacity crunch, not a bottleneck" and prefers Nippon Chemical (4092) over Sakai (4078) — the same ordering this page's peer-swarm reached independently. See 4078/4078, 4092/4092.

The electrode side: BME nickel vs PME palladium

The electrode side carries its own history and is the second moat axis. Until the late 1990s, MLCC inner electrodes were palladium (PME, precious metal electrode) — conductive, oxidation-resistant, and expensive, capping capacitance density and tying margins to the palladium spot price. The shift to nickel inner electrodes (BME, base metal electrode) was the single largest cost reduction in MLCC history and the moat event that defined modern industry structure. The catch: nickel oxidizes during co-firing unless fired in a reducing atmosphere with precisely controlled oxygen partial pressure. Murata, Samsung Electro-Mechanics, and Taiyo Yuden mastered BME nickel co-firing on Class II X7R/X8R chemistry first; second-tier players (Walsin, Holy Stone, Yageo at the low end) followed years later and still trail on high-cap, high-voltage, and small-form-factor BME parts. Chinese new entrants are catching up on BME at large case sizes but cannot reliably co-fire small-form-factor high-cap parts at scale. PME survives only for high-reliability military, aerospace, and medical applications where palladium's process forgiveness is worth the cost — a rounding error of total volume. BME is roughly 95% of the market.

Key terminology

• BME / PME: base metal electrode (nickel) vs precious metal electrode (palladium). BME ≈ 95% of the market.
• Class I dielectric (C0G/NP0): temperature-stable, low-permittivity, low-capacitance, high-precision, very low loss, near-zero temperature drift. RF, timing, filter applications.
• Class II dielectric (X5R, X7R, X8R): high-permittivity, capacitance varies with temperature and voltage. The bulk of the market by volume and revenue. X7R holds capacitance within ±15% from -55°C to +125°C and dominates AI-server decoupling. X8R extends the upper bound to +150°C and is the auto-grade workhorse.
• Case size: imperial-inch dimensions in hundredths. 0402 = 0.04"×0.02" = 1.0mm×0.5mm. 0201 = 0.5mm×0.25mm. 008004 = 0.008"×0.004" = 0.25mm×0.125mm. 006003 = 0.16mm×0.08mm (Murata's newest, ~75% smaller in volume than 008004). The trend is monotonically smaller; each shrink is a 2–3 year, multi-hundred-billion-yen retool (~¥150–200B capex per generation per major maker).
• High-cap MLCC: typically ≥10µF in case sizes 0402 or smaller — where AI-server and high-end smartphone demand and structural pricing power sit.
• AEC-Q200: automotive electronics qualification standard. Two-to-three-year cycle from first sample to design-win.
• Book-to-bill (BB): order intake ÷ shipments. >1.0 means orders outpace shipments and backlog is building; the industry watches Murata's, Samsung E-M's, and Taiyo Yuden's BB as cycle reads.
• ESR / ESL: equivalent series resistance / equivalent series inductance — both parasitics that limit how fast and cleanly a capacitor delivers charge. ESL below 1 nanohenry is critical for high-frequency transient response; smaller case sizes have lower ESL, which is why 0402 parts sit directly around the GPU package.

Process flow — eight steps grouped into stages, two crown jewels

MLCC manufacturing involves dozens of sequential steps. The three core technologies are materials technology (ceramic powders and formulation), overlay/electrode printing technology, and co-firing technology. The canonical eight-step flow:

1. Ceramic powder synthesis — BaTiO3 via hydrothermal route or solid-state calcination, doped with rare earths for temperature stability. ~50% of dielectric cost. Sakai, Nippon Chemical, PDC are major external suppliers; Murata makes its own.
2. Slurry preparation / tape casting — BaTiO3 powder mixed with organic binders, plasticizers, solvents, and dopant additives into a ceramic slurry, then cast onto a carrier film with a doctor blade into a continuous "green tape" at the target ~0.5µm (0.5–2µm for high-cap) thickness, dried. Thickness uniformity must hold to within 2–5% across the tape width and length. Any defect — a surviving particle agglomerate — propagates through every subsequent layer and becomes a short-circuit site. This is where the particle-size ceiling bites: 150nm powder cannot reliably cast a 500nm tape because a single oversized particle spans the entire layer.
3. Electrode printing (overlay) — nickel paste screen-printed in offset patterns on each green tape at sub-µm line/space precision. For high-cap parts this means printing on 500–1,200 individual sheets. One of the two crown jewels. The side-gap construction method matters here for AI-server parts: maximizing the electrode overlap area within the body (minimizing side margin) raises capacitance per unit volume but demands tighter control to prevent edge shorts.
4. Stacking and lamination — hundreds of printed tapes stacked, precisely aligned, and isostatically pressed into a green block. Registration accuracy between layers determines whether internal electrodes connect properly at the terminations; Murata's equipment achieves sub-5µm registration across 1,000+ layers.
5. Dicing — individual capacitor bodies cut from the laminate block.
6. Debinding and co-firing — slow heating to burn off organic binders (debinding at 300–500°C over 12–24 hours to avoid cracking from gas evolution), then sintering at ~1,200–1,300°C in a precisely controlled reducing/nitrogen atmosphere where nickel does not oxidize and ceramic densifies without delamination. The other crown jewel. Ceramic and nickel have different shrinkage rates during sintering, and controlling that mismatch across 1,000+ layers without cracking or delamination is the multi-decade process-IP problem that separates Tier-1 from everyone else. Any organic residue leaves carbon contamination that creates leakage current. Murata's sintering furnaces are proprietary designs not available on the open equipment market. Yield on 008004-size high-cap parts can be below 50% in trial runs.
7. Termination, plating — copper end-caps applied to connect the alternating internal electrodes, then nickel and tin plating layers for solderability.
8. Test, tape, and reel — automated electrical test (capacitance, dissipation factor, insulation resistance, dielectric withstanding voltage), binning by capacitance and temperature class, packed for SMT pick-and-place. Auto-grade (AEC-Q200) adds thermal-shock cycling, humidity, flex, and accelerated life testing.

The two crown jewels are electrode printing (step 3) and co-firing (step 6) — also the two steps where the gap between Murata/Samsung E-M/Taiyo Yuden and the rest is widest. Co-firing especially is where decades of process IP compound because failure modes are statistical and require enormous yield databases to engineer out.

The yield math that gates entry

The decisive economic fact about co-firing is multiplicative yield loss. A 99.5% yield per layer across 1,000 layers gives a cumulative body yield of only 0.7%. Getting to commercially viable cumulative yields above 85% requires mastery of every stage simultaneously, and that mastery took the Tier-1s decades. This is why only Murata, TDK, Samsung Electro-Mechanics, and Kyocera AVX hold production-scale capability for ultra-high-CV designs above 47µF. A new entrant with comparable equipment — assuming they could even source it, since Murata's furnaces are proprietary — would need 5–10 years of yield learning before reaching commercial viability at the high end. The total yield from raw powder to tested finished part is the most closely guarded secret in the industry, and the margin gap between Tier-1 Japanese and Tier-2 Taiwanese is mostly a yield story.

Cost structure — why upstream powder gains leverage as the mix shifts

The cost structure is the reason barium titanate pricing matters more as the industry shifts to high-cap parts. Ceramic material's share of finished cost rises with capacitance because more dielectric layers require more raw material per unit:

Cost component	Low-capacitance MLCC	High-capacitance MLCC
Ceramic material	20–25%	35–45%
Internal electrode (Ni/Ag/Pd)	5%	5–10%
External electrode (Cu/Ag)	5%	5–10%
Packaging materials	20–30%	1–5%
Labor	10–20%	10–20%
Equipment depreciation + other	20–35%	20–30%

Powder is roughly 30–50% of dielectric cost and roughly 10–15% of finished MLCC cost. As the global mix shifts toward high-cap AI and automotive applications, powder pricing gains increasing leverage over finished-product margins. Dielectric powder pricing follows MLCC pricing with a 2–3 quarter lag.

The metrics that read competitive position

Six numbers tell you most of what you need to know about where a maker plays on the curve: capacitance density (µF/mm³, the single best read); smallest body size in mass production (Murata ships 008004; commodity Chinese makers ship 0603/0402 routinely, struggle at 0201, do not ship 008004 at scale); voltage rating × capacitance, especially in the 100V-plus auto/EV bucket (TDK and Murata lead); yield on high-end SKUs (undisclosed by all, but the source of the Tier-1/Tier-2 margin gap); temperature class (X8R for auto -55 to +150°C, C0G for precision sit atop the price/margin curve); and AEC-Q200 SKU count and active design-in pipeline (the proxy for auto-cycle leverage — a 2–3 year qualification cycle means today's design-ins are 2028 revenue).

Competing capacitor types — MLCC wins almost everywhere, but not the edges

Capacitors break down by dielectric into ceramic (MLCC), aluminum electrolytic, tantalum electrolytic, and polyester film. MLCC's edge is the combination of high compactness, small ESR, high reliability, non-polar operation, high pressure resistance, long service life, and small volume — at moderate capacity and moderate cost. The three classes that compete for sockets and the niches MLCC cannot take:

• Tantalum (KEMET/Yageo, AVX/Kyocera, NEC Tokin): higher capacitance per unit volume than MLCC at low voltage, stable, no DC-bias derating. Used on AI-server VRM rails where high capacitance and low ESR matter and MLCC would need a stack of parts to hit spec. The tantalum-AI-server thesis is real — Yageo's KEMET acquisition (2020) was driven by exactly this design-win pattern on NVIDIA Hopper/Blackwell rails. Downside: conflict-mineral supply risk and inability to scale capacitance down with size the way MLCC does. Tantalum powder itself is a concentrated upstream (HC Starck private, GAM — ~3 suppliers).
• Aluminum electrolytic and polymer aluminum (Nichicon, Rubycon, Panasonic, Chemi-Con, TDK): bulk decoupling on PSUs, motherboard rails, industrial. Larger footprint, shorter life, lower frequency response. Mostly displaced by MLCC at the rail level but retained for bulk hold-up on power delivery and AI-server PSU output. At the facility level, polymer aluminum caps (4–8 per stage, 680–1,500µF each, ESR below 15 milliohms) provide bulk energy storage — the "water tanks."
• Film (Vishay, TDK Epcos, Panasonic, Nichicon): high-voltage, high-reliability, EV inverter DC-link, solar inverter, traction. Will not be displaced by MLCC at the 600V+ end because MLCC voltage-derating and reliability profile do not match film's failure mode. Metallized polypropylene film dominates the EV traction DC-link specifically because of its self-healing property: when a minor dielectric defect causes a localized short, the high local current density vaporizes the thin metallization around the fault, electrically isolating it without catastrophic failure — preventing the cascade failures that would immobilize an EV.

The pattern: MLCC dominates small-signal, decoupling, RF, and low-to-medium-voltage segments; tantalum and polymer aluminum hold the AI-server rail bulk-decoupling niche; film holds the high-voltage power-conversion niche. MLCC's growth in AI server and EV comes from content expansion within its segments, not from displacing alternatives. The adjacent silicon-capacitor category (IC-package decoupling, EMIB/CoWoS-S/I-Cube use case, a 4-supplier oligopoly of Murata/SEMCO/AP Memory/TSMC) and supercapacitors/LIC (rack-level energy buffering) are emerging but not yet at MLCC volume; see 009150 and 6531.

AI-server power delivery — why a board needs tens of thousands of MLCCs

A single NVIDIA GB200 GPU module consumes 500–800W at operating voltages between 0.65V and 1.0V; at 800W and 0.8V, peak current approaches 1,000 amperes, swinging at ~100A/µs. The power delivery network (PDN) solves this through a multi-stage voltage-conversion chain with capacitors at every stage, each tier doing a different job at a different frequency band:

• Facility level: 277V AC → 48V DC at the rack shelf. Polymer aluminum capacitors (4–8 per stage, 680–1,500µF, ESR <15mΩ) for bulk energy storage.
• Intermediate bus: 48V → 12V or 6V. High-CV X6S/X7R MLCCs in 1206/1210 case sizes (100–200 per stage, 100–330µF each) for mid-frequency impedance shaping at 10kHz–1MHz. This is the ultra-high-CV tier where only four suppliers are qualified — Samsung's CL31X227 (220µF, 1206) and CL32X337 (330µF, 1210) are the direct competitors to Murata's equivalents.
• Near the GPU die: 6–8 multiphase voltage regulators deliver the final 0.8V core rail at 50–150A per phase. Ultra-low-ESL MLCC arrays in 0603/0402 footprints (500–1,000 per stage, 10–100µF each) handle high-frequency transient response in the 1–100MHz band. ESL below 1nH is critical, so the smallest case sizes sit directly around the package.
• Closest to the die: embedded capacitors in the package substrate or PCB provide distributed capacitance above 100MHz where even 0402 MLCCs are too far from the silicon. Still emerging (Empower ECAP, IPDiA), not yet at volume.

Total per AI-server board: roughly 10,000–20,000 MLCCs plus 300–600 polymer/hybrid aluminum capacitors across all rails and functional blocks (GPU core, memory, NVLink, storage controllers, auxiliary supplies). Murata states the GB300 platform requires approximately 30,000 MLCCs per baseboard — about thirty times a mobile phone and three times an automobile — with a single rack consuming as many as 440,000 individual capacitors. A GB200 NVL72 rack carries 36 baseboards at 25k-class content each, approaching ~1 million MLCCs per rack versus roughly 10,000 in a standard 2U dual-socket Xeon server. Morgan Stanley pegs MLCC content per rack at ~$130K (Blackwell NVL72, ~441,000 MLCCs at ~$0.30 avg ASP), rising to the next generation — and the generational BOM step is one of the most contested numbers in the sector. (See End markets / Cycle sections for the GB300→VR200 BOM debate, the "TAM is outside the rack BOM" pushback, and the price-hike-vs-allocation-premium-vs-mix-shift framings.)

The 800V EV architecture — an independent second driver

The automotive transition from 400V to 800V battery architectures changes capacitor requirements in ways independent of the AI cycle, which matters for portfolio construction: even if AI capex moderates, the 800V ramp continues. Doubling battery voltage halves the current for a given power (P = V × I), reducing copper weight in the harness and enabling faster DC charging (350kW+ vs 150kW at 400V) — but doubling the voltage increases stress on every capacitor. The DC-link capacitor in the traction inverter must now withstand 1,000V DC continuously (vs 500V at 400V), and energy storage stress scales with the square of voltage (E = ½CV²). Metallized polypropylene film dominates this socket on its self-healing property (above). High-voltage MLCC (1kV-rated, 100nF class, C0G/NP0 dielectric) grew at a 17.3% CAGR through 2025 — triple the overall capacitor market rate — with 23 new 800V production models launched in 2025 and 50+ planned through 2027. OEMs now specify AEC-Q200 across 100% of capacitors in 800V powertrain circuits, recognizing that a single failure could disable the vehicle or trigger thermal runaway.

MLCC content per powertrain type (TDK data): ICE ~5,000, HEV ~6,000, PHEV ~8,000, full BEV ~10,000 — so electrification roughly doubles content. (The primer also cites a 3,000–5,000 per BEV vs ~2,000 per ICE framing; the TDK powertrain-table numbers above are the more granular source.) The incremental ~5,000 MLCCs per BEV vs ICE are disproportionately high-voltage, high-reliability auto-grade parts commanding 30–80% price premiums over consumer-grade equivalents and requiring 12–18 months of AEC-Q200 qualification before a new supplier can ship.

Powertrain	MLCCs	Inductors	Piezoelectric
ICE	5,000	450	100
HEV	6,000	600	140
PHEV	8,000	650	160
BEV	10,000	700	180

The qualification barrier as a moat

AEC-Q200 Revision E (published March 2023, continuously updated) is the de facto automotive passive-component qualification standard, and the test regime is itself a barrier. Mechanical shock at 50G half-sine waves lasting six milliseconds simulates rough terrain; bending-strength testing deforms the substrate beyond 3mm to probe internal micro-crack resistance; thermal cycling extends to +150°C for high-voltage systems. AEC-Q200-certified parts exhibit failure rates one order of magnitude lower (10–100× improvement) than industrial-grade. The practical consequence is a 12–18 month qualification timeline before a new supplier can ship a single production part — covering not just 2,000–5,000 hours of accelerated life testing at elevated voltage and temperature, but PPAP (Production Part Approval Process) documentation, traceability, and statistical-process-control evidence across multiple production lots. This is the moat Chinese makers have not crossed at scale in the high-end automotive tier: Fenghua and Sanhuan are progressing in consumer and industrial segments and have begun small-size high-cap products, but OEM purchasing will not override decades of field-reliability data on price alone. The barrier reinforces structure — automotive MLCC by volume (2022): Murata 47.0%, Samsung Electro-Mechanics 14.6%, TDK 14.6%, Taiyo Yuden 10.4%.

The staircase of entry barriers

Barriers are high at the top and decrease as you walk down the SKU stack — a literal staircase. To enter the commodity 0603 BME nickel market, a Chinese new entrant needs ~$500M–$1B of capex and 3–5 years of process tuning, both achievable. To enter the 0201 small-form-factor mid-cap market, another doubling of capex plus multi-year per-customer qualification plus ceramic-powder supply qualification. To enter the 008004 ultra-miniature high-cap market, everything above plus multi-decade ceramic process IP nobody has replicated outside Murata, Samsung E-M, and Taiyo Yuden. The same staircase applies to AEC-Q200: body electronics is approachable for Tier-2; powertrain 100V-plus is essentially closed (a Murata–TDK duopoly with Samsung E-M trying to break in). Switching costs are real but bounded — for commodity MLCC, OEMs routinely qualify multiple suppliers and split allocation by price; for high-end, AEC-Q200, and AI-server high-cap, qualification cycles run 18–36 months and once a Murata 008004 22µF part is designed into an iPhone power-management IC it stays for the device's lifecycle.

Where the value pool sits across the chain

The chain runs powder → nickel paste / copper terminations / equipment → MLCC manufacturers → distributors → end OEMs. Gross-margin capture by layer: qualified powder ~25–35% (operating margin 25–35% mid-cycle, expanding to 35–45% at peak); equipment ~30–40% on cyclical revenue; MLCC manufacturers ~30–45% at Tier-1 and ~20–30% at Tier-2, where most of the value pool sits; distributors ~5–10% on a large revenue base; EMS assemblers low single digits. The two highest-margin layers are the qualified powder layer and the Tier-1 manufacturer layer; Tier-2 MLCC, equipment, and distribution are structurally lower-margin. Tier-1 operating margin runs 15–25% mid-cycle, 25–30%+ at peaks, compressing to 5–10% at troughs; Tier-2 Taiwanese run 10–15 points below across the cycle; Chinese entrants below that. Within the strategic tier specifically (high-voltage, polymer electrolytic, ultra-high-CV), the bifurcation is stark: strategic MLCCs are ~35–40% of market value but ~60–70% of industry profit, growing 12–17% CAGR with 30–80% price premiums and 24–47 week lead times, while commodity MLCCs face -10–20% pricing pressure from Chinese capacity. Mix shift into the strategic tier is the core of the margin story.

The two crown-jewel process bottlenecks — co-firing IP at 008004-and-below and sub-100nm barium titanate powder — are the structural choke points. Co-fire IP is captured inside the Tier-1 equity (it is why Murata trades at a premium multiple, not a separate investable layer). Sub-100nm BaTiO3 powder cannot be bypassed at the high end of the dielectric curve, is held by ≤4 qualified suppliers controlling >80% of qualified output, and is the one bottleneck with a small-cap pure-play attached. Nickel electrode paste (Shoei Chemical the most-named specialty supplier, ≤5 qualified globally) is a secondary, less-tight bottleneck because Tier-1 makers can substitute among paste suppliers more readily than among powder suppliers.

Subsectors

The passives & components complex is not one market. It is a stack of distinct sub-industries with different competitive structures, different bottlenecks, and different ways into the AI/EV cycle. The mistake that wrecks the analysis is treating "MLCC makers" as homogeneous, or treating the whole capacitor world as MLCC. Below are the sub-areas the sources surface, each with what it is, the technology, who plays, and the investment angle.

MLCC technology and value chain

The multilayer ceramic capacitor is the single most populous active electronic component on every modern circuit board, and its supply is gated by three companies in Japan, one chaebol in Korea, and roughly five powder suppliers globally. Content per device frames the whole demand picture: a flagship smartphone ships ~1,000 MLCCs, a BEV uses 3,000-5,000 (vs ~2,000 for ICE), and a modern AI training baseboard ships 15,000-25,000. Murata states the GB300 platform requires ~30,000 MLCCs per baseboard — about thirty times a phone and three times a car — with a single rack consuming as many as 440,000 individual capacitors. The cycle bottomed Q3/Q4 2024, pricing turned in May 2026, and the high end is structurally undersupplied through at least 2027.

The physics is the parallel-plate equation C = εA/d. To pack more capacitance into a smaller body you raise ε with a high-permittivity ceramic, raise A by stacking more layers in parallel, and lower d by thinning the dielectric. All three are bounded by physics and process control. A modern high-cap MLCC for AI server rails stacks 500-1,200 dielectric layers at 0.5-0.8µm thick, with nickel electrodes printed between each layer, co-fired at ~1,200°C in a nitrogen atmosphere. The dielectric is doped with rare earths (yttrium, dysprosium, holmium) to flatten the temperature-capacitance curve, which creates the X7R, X5R, X8R, and C0G temperature classes engineers select by application. X7R holds ±15% from -55 to +125°C and dominates AI server decoupling; X8R extends to 150°C and is the auto-grade workhorse; C0G/NP0 has near-zero drift but low permittivity, for RF and precision timing. MLCC capacitance-to-volume ratio rose from 1µF/mm³ in 1996 to 40µF/mm³ in 2020 — a Moore-equivalent trajectory where max capacitance per body roughly doubles every three years and minimum body size halves every five.

The two crown jewels of manufacturing are electrode printing (sub-µm nickel paste screen-printing with sub-5µm registration across 1,000+ layers) and co-firing (sintering ceramic and nickel together at controlled oxygen partial pressure without cracking or delamination). The shift from PME (palladium, precious metal electrode) to BME (nickel, base metal electrode) in the late 1990s was the largest cost reduction in MLCC history and the moat event that defined modern industry structure — BME is now ~95% of the market. A 99.5% yield per layer across 1,000 layers gives a cumulative body yield of only 0.7%; getting to commercially viable >85% requires mastering every stage simultaneously, which took the leaders decades. A new entrant with comparable equipment (assuming they could even source it — Murata's sintering furnaces are proprietary, not on the open market) would need 5-10 years of yield learning at the high end.

The value chain has four layers — ceramic powder and precursors, other inputs (nickel paste / copper terminations / equipment), MLCC manufacturers, and distribution/end customers. The two highest-margin layers are the qualified powder layer (~25-35% gross margin mid-cycle, expanding to 35-45% at peak) and the Tier-1 MLCC manufacturer layer (~30-45% gross margin). Tier-2 MLCC, equipment, and distribution are structurally lower-margin. The aggregate MLCC market is roughly $15B revenue at cycle midpoint (JPM industry model: $16.3B in 2025 → $25B by 2030E at +11% CAGR, with AI server mix rising from 1.1% to 4% of total), against ~5.5-6 trillion units shipped per year, growing toward 8-10 trillion by 2030. Geographic split of revenue: Japan ~50%, Korea ~22%, Taiwan ~12%, China ~10%, Rest of World ~6%; demand is ~60% concentrated in China as the assembly hub.

Investment angle: the cycle is the dominant fact — every name's earnings swing 2-3x peak to trough, every stock swings 3-5x, and cycle timing drives more return variance than stock selection. High-end pricing power runs the full 18-30 month up-cycle; commodity runs only 12-18 months before Chinese overcapacity drags it back. The structurally better business is the high-end pure-play (Murata, the auto slice of TDK); the commodity-exposed names (Walsin, Holy Stone) carry lower mid-cycle multiples and higher cycle betas. See 6981, 009150, 6976, 6762 for the manufacturer deep-dives.

Capacitor manufacturing and the BaTiO3 moat

The barium titanate powder that goes into an MLCC is not a commodity chemical — it is a precision-engineered ceramic where particle size, shape uniformity, and chemical purity set the performance ceiling of the finished capacitor. BaTiO3 is a perovskite with a dielectric constant around 3,000 (air is 1, silicon dioxide ~3.9, alumina ~9); nothing else combines that permittivity with thermal stability and manufacturability at sub-100nm grain size. You cannot make a dielectric layer thinner than the grain size of the powder, so powder physics is the binding constraint on the whole roadmap.

Four preparation methods exist, each with a tradeoff. Solid-state (ball-mill barium carbonate and titanium dioxide, calcine at 1,100-1,200°C) is cheapest but coarsest (>1µm), fine for low-cap commodity parts and where Chinese producers are competitive. Co-precipitation reaches sub-1µm with moderate shape control (SEMCO and some Chinese producers use variants). Sol-gel gives excellent compositional control but high cost and limited scale, mostly experimental. The hydrothermal method is the high-end route: BaTiO3 crystallizes directly from aqueous solution at 150-250°C under pressure, producing sub-100nm spherical particles with no calcination or grinding, preserving particle shape. This is Sakai Chemical's core IP and the only viable route to the thinnest dielectric layers for high-cap AI server MLCCs.

The technology gap is concrete: Japanese producers (Sakai, Nippon Chemical) achieve 80-100nm average particle size; Chinese producers (Guoci Materials, Fenghua) reach only 120-150nm. That 40-50nm gap means Japanese makers stack up to 1,200 layers at 0.5-0.6µm dielectric (Murata hits 1,600 layers) while Chinese makers average 800 layers at 1-2µm. The consequence: Murata produces a 220µF MLCC in a 1206 case while the best Chinese maker maxes at ~100µF in the same package. The cost structure amplifies the powder layer's leverage as the mix shifts to high-cap: ceramic materials are 20-25% of manufacturing cost in low-cap parts but 35-45% in high-cap parts.

Powder market share is highly concentrated. The Dongguan Securities figures (the authoritative set, correcting Sakai's own "40-50% merchant share" self-report): Sakai Chemical 28%, Ferro (USA) 20%, Nippon Chemical 14%, Guoci Materials (China) 10%, Fuji Titanium 9%, Kyoritsu 8%, Toho 6%, other 5%. The top five hold 81%. Murata produces its own BaTiO3 internally and is structurally self-sufficient; SEMCO, Taiyo Yuden, TDK, Yageo, Walsin, and Holy Stone buy a meaningful fraction externally, with Sakai the most-named supplier. One structural negative to flag against the clean "Sakai is the upstream alpha" framing: Murata formed the MF Material JV with Ishihara Sangyo and Fuji Titanium in September 2023 to internalize barium titanate supply, diluting external powder dependence over time.

Investment angle: sub-100nm BaTiO3 for next-generation high-cap MLCC is the bottleneck the market has not fully priced — it cannot be bypassed at the high end of the dielectric curve, the qualified supplier list is ≤4 firms controlling >80%, and the cleanest small-cap pure-play is Sakai Chemical at ~$500M-$1B market cap (roughly 3-5% of Murata's cap) with thin analyst coverage. Sakai's powder pricing follows MLCC pricing with a 2-3 quarter lag, so the May 2026 hike round flows into Sakai's earnings through 2H 2026 — the entry window is broader and the cycle-confirmation signal arrives one quarter later than for Murata. The Sakai profile flagged it as WATCH not Buy, naming a European, a Japanese, and "several Chinese competitors" as the active set; Nippon Chemical Industrial (4092.T) is the cleaner starting-condition candidate per that work (+22% YoY revenue vs Sakai's +1.9% guide at similar P/E and EV/EBITDA). Nickel electrode paste (Shoei Chemical 4970.T) is a secondary, looser bottleneck — Tier-1 makers can substitute among ~5 qualified suppliers more readily than for ceramic powder. See 4078, 4092 for the upstream names.

Non-MLCC capacitors

MLCC dominates small-signal, decoupling, RF, and low-to-medium-voltage sockets, but three other capacitor classes hold defensible niches, and their AI/EV growth comes from content expansion within those niches rather than from MLCC displacing them. The Dongguan property comparison frames the tradeoffs: MLCC wins on compactness, frequency response, low ESR, reliability, non-polarity, pressure resistance, service life, and small volume; it sits at moderate capacity and moderate cost.

Tantalum (Yageo/KEMET, Kyocera AVX, NEC Tokin) offers higher capacitance per unit volume than MLCC at low voltage, is stable, and has no DC-bias derating. It holds the AI server VRM rail bulk-decoupling niche where MLCC would need a stack of parts to hit spec. Yageo's KEMET acquisition (2020) was driven by exactly this design-win pattern on NVIDIA Hopper/Blackwell rails. Yageo is #1 globally on tantalum (~46% share). The downside is conflict-mineral supply risk and the inability to scale capacitance down with size as MLCC has. The non-obvious read: Yageo is more accurately a tantalum name with MLCC and chip-resistor optionality, and its AI server exposure is proportionally larger and more durable via tantalum (~46% share, multi-generation NVIDIA qualification cycle) than via MLCC. See 2327.

Aluminum electrolytic and polymer aluminum (Nichicon, Rubycon, Panasonic, Chemi-Con, TDK) handle bulk decoupling and hold-up on PSUs, motherboard rails, and AI server PSU output. In the AI server power delivery network, polymer aluminum capacitors (4-8 per stage, 680-1,500µF each, ESR below 15mΩ) sit at the facility-level 277V AC → 48V DC conversion as the bulk "water tanks." Larger footprint, shorter life, lower frequency response; mostly displaced by MLCC at the rail level but retained for bulk hold-up.

Film (Vishay, TDK Epcos, Panasonic, Nichicon) holds the high-voltage power conversion niche — EV inverter DC-link, solar inverter, traction. It will not be displaced by MLCC at the 600V+ end because MLCC's voltage-derating and reliability profile do not match film's failure mode. Metallized polypropylene film capacitors dominate the traction inverter DC-link because of their self-healing property: a minor dielectric defect vaporizes the thin metallization around the fault and isolates it without catastrophic failure, preventing the cascade failures that would immobilize an EV. The 800V EV architecture is an independent second driver here — doubling battery voltage halves current (lighter copper, 350kW+ DC charging) but forces the DC-link capacitor to withstand 1,000V DC continuously. High-voltage MLCC (1kV-rated, 100nF class, C0G/NP0) grew at 17.3% CAGR through 2025, triple the overall capacitor market, with 23 new 800V production models launched in 2025 and 50+ planned through 2027.

A separate non-MLCC category surfacing in the AI cycle is supercapacitors / lithium-ion capacitors (LIC) for rack-level energy buffering — a fragmented competitive set with Skeleton private and Musashi consolidated. Musashi Seimitsu (7220.T), a Japanese auto-parts company, is expanding into 800V high-voltage server hybrid supercapacitors (HSC): 3.8V HSC vs traditional EDLC 2.8V, a ramp from 200,000 cells in 2026 to 6.5M cells in 2028, with Delta Electronics as downstream customer. The non-MLCC-capacitor primer's call is Musashi PASS at ¥7,960 because LIC is <1-2% of consolidated revenue (watch zone ¥4,500-5,500), with Collyer Bridge having entered then exited a starter within hours on the 7x-vs-consensus-22x PE gap.

Where capacitors sit in the broader passive stack: capacitors are ~65% of the passive component market (2023), with inductors and resistors the remainder. By dielectric type, ceramic capacitors are dominant, and within ceramic the MLCC (multilayer chip) format is >90% — MLCCs are 93% of ceramic capacitor production, vs 4% single-layer and 3% leaded multilayer. The four passive subsegments (tantalum, polymer aluminum, film, MLCC) share AI server and EV as cycle drivers but have different competitive structures and bottleneck profiles. Investment angle: the cleanest non-MLCC AI-server play is tantalum via Yageo/KEMET (captured downstream, with private upstream tantalum-powder supplier optionality at HC Starck / GAM that is low-investability for retail); film is an EV-inverter thesis insulated from AI capex digestion; supercaps are early and revenue-immaterial at the listed names so far.

Silicon capacitors

The newest non-MLCC category, and the one Collyer Bridge turned newly bullish on in late May 2026 ("its pretty wild i did not want to get involved in MLCCs now researching a supposedly superior type of capacitator"). Silicon capacitors are built on a silicon substrate using semiconductor process techniques rather than ceramic sintering, giving them precise, stable capacitance and extreme thinness for IC-package decoupling. The use case is advanced packaging — EMIB, CoWoS-S, and I-Cube — where the capacitor sits in or near the package substrate, closer to the silicon than even a 0402 MLCC. This is the "embedded capacitor" frontier the MLCC technical work flagged as emerging (Empower ECAP, IPDiA) and not yet at AI server production volume, but the AP Memory thesis suggests it is arriving faster than expected and is levered to TPUs as well as GPUs.

The competitive set is a 4-supplier oligopoly: Murata, Samsung Electro-Mechanics, AP Memory (6531.TW), and TSMC (2330.TW). The Collyer Bridge chat read is AP Memory > SEMCO > Murata for pure-play silicon-cap exposure — AP Memory the highest-concentration play, SEMCO flagged but parabolic ("up nearly 100% since Korea opened retail"), and Murata diluted because no segment revenue breakdown is disclosed ("I cant find the revenue breakdown from Murata so far on silicon cap"). Jukan (referenced in bulge-bracket sell-side coverage with decent frequency) had a forthcoming SEMCO/silicon-cap memo as the anticipated next datapoint.

Investment angle: the non-MLCC-capacitor primer reworked SEMCO fair value to KRW 800K-1.25M after integrating Citi's W609bn FY27E / W905bn FY28E silicon-capacitor revenue path and the W1.6T two-year supply order — still implying 32-55% downside vs the then-current KRW 1.85M. Owning Samsung Electronics (005930.KS) for SEMCO exposure does not work (SEMCO is 1.7% of Samsung Electronics market cap). The risk is valuation friction at every listed entry point (SEMCO parabolic, AP Memory a TWO-listed small-cap, Murata's exposure undisclosed). See 009150, 6531 for the names. The structural watch: AP Memory primer at pequityresearch.substack.com is the reference write-up; whether silicon caps eventually substitute into MLCC sockets (eDTC / HD-MiM on-chip substitution) is a named long-term risk to the Murata MLCC thesis.

Japan / Asia passive component makers

The manufacturer layer is one of the most concentrated in the components world — comparable to memory NAND/DRAM in degree of consolidation but with more stable competitive ranking — and it sorts into three tiers by geography and capability.

Tier 1 — Japan and Korea. Murata Manufacturing (6981.T, ~31.8-34% global share) is Tier-1 across every segment, dominant in high-end (high-cap, 008004, AEC-Q200 100V-plus auto), owns its own powder, carries a fortress balance sheet (~¥600B net cash), and is the price-setter — it trades at premium valuation because it earns it. Samsung Electro-Mechanics (009150.KS, ~11.8-22% global) is the other Tier-1 and the volume leader in AI-server MLCC (~40% of AI-server share per Maeil Business / Digitimes), available at a chaebol discount (~16x forward P/E vs Murata's 28x). TDK (6762.T, ~12% on MLCC) leads 100V-plus auto MLCC and AI passives but MLCC is only ~10% of revenue — owning TDK for MLCC means owning the ATL Apple silicon-anode battery business (55% of revenue, 89% of segment OP) and HDD heads by default. Taiyo Yuden (6976.T, ~9% global) is the pure-play, 71% of revenue MLCC, strongest in smartphone high-cap and increasingly AI server, with the highest cycle beta of the Tier-1s; it shipped 22µF in 1005 (Aug 2025) and 22µF in 0402 (Sept 2025). Kyocera (via AVX) rounds out the four qualified ultra-high-CV suppliers. In automotive-grade MLCC by volume (2022): Murata 47.0%, SEMCO 14.6%, TDK 14.6%, Taiyo Yuden 10.4%.

Tier 2 — Taiwan. Yageo (2327.TW, ~5-6% on MLCC) is the multi-passive consolidator — #1 on tantalum (KEMET 2020), #1 on chip resistor, #3 on MLCC, with a roll-up record (KEMET, Chilisin 2021, Nexensos 2023, Telemecanique 2023, Shibaura 2025) and Pierre Chen's ~20% personal ownership as the strongest founder alignment in the complex. Walsin Technology (2492.TW, ~3.2%) is Taiwan #2 MLCC, a fast-follower not a price-setter (family-controlled Chiao group, 41.7% insider), with China utilization recovery torque but a structural pricing-power discount — its May 2026 case-by-case-only pricing posture is the tell. Holy Stone (3026.TW) is a Tier-2/3 hybrid manufacture-plus-distribution where the split is undisclosed, the single largest information gap in the listed peer set. Prosperity Dielectrics (PDC, 6173.TWO) is sub-scale finished MLCC plus powder, a PSA-group ally of Walsin, where the powder portion is structurally more interesting than the finished-MLCC portion.

Tier 3 — Mainland China. Fenghua Advanced Technology (000636.SZ), Sunlord Electronics (002138.SZ), Sanhuan Group (300408.SZ), Guoci Materials, Huaguo Technology (1.5% share). Government-funded import-substitution capex, aggressive in commodity 0603/0402, struggling at 0201, not shipping 008004 at scale. In 2025 China imported 2.56 trillion MLCCs worth US$6.179 billion; if 50% of import volume were domestically replaced the substitution scale would reach 1.28 trillion units. Dongguan Securities rates Fenghua (000636.SZ) and Sanhuan (300408.SZ) Buy-and-Hold as the domestic-substitution beneficiaries.

The Tier-1 four hold ~77% of global MLCC revenue (Dongguan: top five 83.6%); the barrier staircase rises from commodity 0603 ($500M-$1B capex, 3-5 years) to 0201 mid-cap (another capex doubling plus multi-year customer qualification) to 008004 ultra-miniature high-cap (multi-decade ceramic process IP nobody has replicated outside Murata, SEMCO, Taiyo Yuden). AEC-Q200 qualification adds a 12-18 month barrier per part (AEC-Q200 Rev E, March 2023: 50G mechanical shock, >3mm bend testing, thermal cycling to +150°C; certified parts show 10-100x lower failure rates). Adjacent Japan/Asia names that surface in the same supply chain: ceramic substrates (Maruwa), nickel paste (Shoei 4970.T, Mitsui Mining 5706.T, Sumitomo Metal Mining 5713.T), tape-casting equipment (Hirano Tecseed, NORITAKE), and ruthenium for thick-film resistors (Furuya Metal 7826.T — claims ~70% share of HDD sputtering-target high-purity Ru).

Investment angle: the long-term steady state is probably a Tier-1 oligopoly of Murata, SEMCO, plus one of Taiyo Yuden or TDK as third anchor, with Yageo as the multi-passive consolidator and a consolidating Tier-2/China commodity tail. The Chinese-entrant threat to Tier-1 has been overestimated at every prior cycle and is likely overestimated now — process IP at the high end is not catching up at the pace bears predicted. The right way to play the China-commodity tail is to short the Taiwan Tier-2 names rather than long the hard-to-own Chinese names. Correlation across the peer group is high, so owning seven MLCC names concentrates rather than diversifies the cycle risk; the construction is one to two manufacturer names at conviction sizing plus one upstream name (Sakai/Nippon Chemical) if the bottleneck thesis is taken seriously. Peer-swarm ranked verdict (2026-05-25): Murata (BUY scale-in) > TDK > Yageo > Walsin > Taiyo Yuden > Nippon Chemical > Sakai > SEMCO > PDC > Holy Stone. See the individual ticker pages 6981, 6762, 2327, 2492, 6976, 009150, 6173, 3026, 4078, 4092 for company-level detail.

Value chain

The MLCC value chain runs in six economically distinct stages, but the money sits in two of them. From the ground up: ceramic powder and other input materials, specialty manufacturing equipment, the MLCC manufacturers themselves, distributors, EMS assemblers, and end OEMs. The two highest-margin layers are the qualified barium-titanate powder layer and the Tier-1 MLCC manufacturer layer. Everything else — equipment, distribution, EMS — is structurally lower-margin. The whole chain is gated at the top by three Japanese makers plus one Korean chaebol, and gated at the bottom (upstream) by roughly five powder suppliers globally. That double choke point is the central fact of the industry.

The chain in prose (the four-layer map)

The primer renders the chain as five vertical columns. Reproduced as prose:

• Ceramic powder / precursors: Sakai Chemical (4078.T), Nippon Chemical Industrial (4092.T), Prosperity Dielectrics / PDC (6173.TWO), Fuji Titanium Industry. Murata makes its own.
• Nickel paste: Shoei Chemical (4970.T), Mitsui Mining & Smelting (5706.T), Sumitomo Metal Mining (5713.T); Murata produces in-house.
• Tape-casting / process equipment: Hirano Tecseed, NORITAKE, Mitsubishi Heavy.
• MLCC manufacturers: Murata (6981.T), Samsung Electro-Mechanics (009150.KS), Taiyo Yuden (6976.T), TDK (6762.T), Yageo (2327.TW), Walsin (2492.TW), Holy Stone (3026.TW), PDC (6173.TWO, small), plus Chinese entrants Fenghua (000636.SZ) and Sunlord (002138.SZ).
• Distributors: Mouser, Digi-Key, Arrow, WPG, Avnet.
• End OEMs: Apple, NVIDIA, Foxconn (EMS), Quanta (EMS), Tesla, BYD.

Margin capture across the chain, layer by layer: the powder layer takes roughly 25-35% gross margin on a small revenue base; the equipment layer takes ~30-40% gross margin on cyclical revenue; MLCC manufacturers take ~30-45% gross margin at Tier-1 and ~20-30% at Tier-2, and this is where the bulk of the value pool sits; distributors take ~5-10% gross margin on a large revenue base; EMS assemblers take low-single-digit margins. See 6981, 009150, 6976, 6762, 2327, 2492, 3026, 6173, 4078, 4092 for company-specific positioning.

Layer 1 — Ceramic powder, the upstream bottleneck

Barium titanate (BaTiO3) is the central material. The qualified high-end (sub-100nm, monodisperse, doped for Class II) is supplied by three companies plus a smaller fourth: Sakai Chemical (Japan), Nippon Chemical Industrial (Japan), PDC (Taiwan), with Fuji Titanium a smaller fourth player. Murata produces its own BaTiO3 internally and is structurally self-sufficient. Samsung Electro-Mechanics, Taiyo Yuden, TDK, Yageo, Walsin, and Holy Stone buy a meaningful fraction of their high-end powder externally — Sakai is the most-named supplier in industry press.

Dongguan Securities' market-share build is the most precise and corrects the Sakai self-report (Sakai's profile claimed "40-50% merchant share," but that is share of merchant supply, not total market):

Company	Country	Powder market share
Sakai Chemical (堺化学)	Japan	28%
Ferro	USA	20%
Nippon Chemical (日本化学)	Japan	14%
Guoci Materials (国瓷材料)	China	10%
Fuji Titanium (富士钛业)	Japan	9%
Kyoritsu (共立)	Japan	8%
Toho (东邦)	Japan	6%
Other	—	5%

The top 5 (Sakai, Ferro, Nippon Chemical, Guoci, Fuji Titanium) hold 81% market share. Highly concentrated. Note Murata's vertical-integration move: it formed the MF Material JV with Ishihara Sangyo and Fuji Titanium in September 2023 to internalize barium titanate supply, structurally diluting external powder dependence over time.

Powder economics: powder is roughly 30-50% of dielectric cost and roughly 10-15% of finished MLCC cost. Margins at the qualified powder layer run 25-35% operating margin mid-cycle, expanding to 35-45% at peak-cycle when demand outstrips supply. Gross margins at Sakai run materially higher than at any Tier-2 MLCC maker. Market cap at this layer is small relative to downstream — Sakai sits at roughly $500M-$1B depending on cycle, against Murata's ~$78B. The leverage is real, but float and analyst coverage are thin.

The choke point is not bypassable at the high end. You cannot make a dielectric layer thinner than the grain size of the powder you are using, and sub-100nm BaTiO3 grade is what gates next-generation high-cap MLCC. Pricing-power lag matters for timing: Sakai's powder pricing follows MLCC pricing with a 2-3 quarter lag, so the May 2026 MLCC hike round flows into Sakai's powder pricing through 2H 2026 — Sakai's earnings catch-up is timed roughly one quarter behind Murata's, which widens the entry window on the upstream name.

The Japan-China powder technology gap (the hard bottleneck)

The bottleneck is physical, not commercial. Japanese producers achieve average BaTiO3 particle sizes of 80-100nm via hydrothermal synthesis. Chinese producers (Guoci Materials, Fenghua) currently reach only 120-150nm. That 40-50nm gap has direct downstream consequences: Japanese MLCC makers stack up to 1,200 layers at 0.5-0.6µm dielectric thickness (Murata achieves 1,600 layers), while Chinese makers average 800 layers at 1-2µm thickness. In the same physical body size, finer powder enables more layers, which means more capacitance. This is why Murata can produce a 220µF MLCC in a 1206 case (3.2mm × 1.6mm) while the best Chinese maker maxes out at around 100µF in the same package.

Four preparation methods exist for barium titanate, and the method determines where in the chain a producer can compete:

Method	Cost	Application scope	Component control	Shape control	Particle size	Calcination	Grinding
Solid-state	Low	Commercial	Poor	Poor	>1.0µm	Required	Required
Co-precipitation	Moderate	Commercial/experimental	Good	Moderate	<1.0µm	Required	Required
Sol-gel	High	Experimental	Excellent	Moderate	<1.0µm	Required	Required
Hydrothermal	High	Commercial/experimental	Excellent	Good	<1.0µm	Not required	Not required

The hydrothermal method crystallizes BaTiO3 directly from aqueous solution at 150-250°C under pressure, producing sub-100nm spherical particles with excellent uniformity, no calcination or grinding needed (which preserves particle shape). It is Sakai Chemical's core IP and the only viable route to the thinnest dielectric layers for high-cap AI-server MLCC. Guoci Materials also uses hydrothermal but trails on grain size. The solid-state method (cheapest, coarsest, >1µm) is where Chinese producers are competitive — i.e. low-cap commodity MLCC where layer count is low.

Layer 2 — Other inputs and equipment

Nickel paste (Shoei Chemical 4970.T, Mitsui Mining & Smelting 5706.T, Sumitomo Metal Mining 5713.T) is a smaller market with fewer named pure-plays; Shoei is the most MLCC-specific. Copper termination paste is commoditized. The equipment layer — tape casters (Hirano Tecseed, NORITAKE), stack/lamination presses (Mitsubishi Heavy), co-fire furnaces — is concentrated among Japanese specialty equipment makers but there is no pure-play MLCC name among them. Critically, Murata's sintering/co-fire furnaces are proprietary in-house designs not available on the open equipment market, which is part of why a new entrant cannot simply buy its way to the high end. The aggregate equipment market for MLCC capacity expansion is low single-digit billions of dollars per year at cycle peak, spread across multiple multi-product suppliers. Each body-size shrink generation runs roughly ¥150-200B capex per major maker.

Layer 3 — MLCC manufacturers, the main value pool

This is where most of the value pool sits. Global share is concentrated. Two share builds exist in the sources, on different bases. The primer's 2024-2025 figures:

• Murata (6981.T) — ~34% global, Tier-1 across every segment, ~45% of corporate revenue is MLCC, owns its own powder, the price-setter.
• Samsung Electro-Mechanics (009150.KS) — ~22% global, the other Tier-1, volume leader in AI-server MLCC (~40% AI-server share per Digitimes).
• TDK (6762.T) — ~12% on MLCC, but MLCC is only ~10% of TDK revenue.
• Taiyo Yuden (6976.T) — ~9% global, the pure-play (71% of revenue is MLCC).
• Yageo (2327.TW) — ~6% on MLCC, but #1 globally on tantalum (via KEMET) and chip resistor.
• Walsin (2492.TW) — Taiwan #2 MLCC, fast-follower not price-setter.
• Holy Stone (3026.TW), PDC (6173.TWO), and Chinese entrants Fenghua / Sunlord round out the tail.

Dongguan's 2024 global MLCC share build is more granular and lower for the named leaders:

Company	Share	Tier
Murata (村田)	31.8%	1
Samsung Electro-Mechanics (三星电机)	11.8%	1
Taiyo Yuden, TDK, Kyocera	—	1
Top 5 combined	83.6%	—
Yageo (国巨)	5.0%	2
Walsin Technology (华新科技)	3.2%	2
Fenghua (风华高科)	2.5%	3
Sanhuan Group (三环集团)	1.9%	3
Huaguo Technology (华国科技)	1.5%	3

The Tier-1 four (Murata + Samsung E-M + TDK + Taiyo Yuden) hold roughly 77% of global MLCC revenue per the primer; adding Yageo gets to ~83%. The remaining ~17% is Tier-2 Taiwanese, Chinese entrants, and legacy small players. This is among the most concentrated component markets in semiconductors, comparable to NAND/DRAM consolidation but with more stable competitive ranking.

Operating margin at Tier-1 makers runs 15-25% mid-cycle, expanding to 25-30%+ at peaks (Murata FY3/18) and compressing to 5-10% at troughs (Taiyo Yuden FY3/24-FY3/25). Tier-2 Taiwanese run 10-15 percentage points below Tier-1 across the cycle; Chinese entrants run below that. The gap is real and structural.

The three-tier competitive landscape per Dongguan: Tier 1 (Japan/Korea) — Murata and Samsung E-M, decades of accumulated expertise, specialize in high-value-added (small size, high capacity, high voltage), targeting automotive, high-end smartphone, and AI server. Tier 2 (Taiwan) — Yageo and Walsin, more diverse portfolios concentrated in low-to-mid-range, actively expanding into automotive/industrial. Tier 3 (Mainland China) — Fenghua and Sanhuan, technological level lags Tier 1 and 2, primarily medium-to-large-size low-capacitance products with recent small-size high-cap breakthroughs.

Layer 4 — Distribution and end customers

The distribution layer (Mouser, Digi-Key, Arrow, Avnet, WPG in Asia) is a low-margin pass-through with ~5-10% gross margin on MLCC. Distributors matter to the cycle read because their lead-time data is the single best industry-wide indicator of supply-demand balance. End OEMs split into smartphone (Apple, Samsung, Xiaomi, OPPO, Vivo via Foxconn/Pegatron EMS), AI infrastructure (NVIDIA boards built by Foxconn/Quanta/Wiwynn for hyperscalers), automotive (BYD, Tesla, VW, Toyota, Hyundai-Kia plus Tier-1s Bosch/Continental/Denso), and a long industrial/consumer tail. EMS assemblers take low-single-digit margins.

The crown jewels — where the manufacturing margin actually comes from

The eight-step process flow (powder synthesis → tape casting → electrode printing → stacking/lamination → dicing → co-firing → termination → test/tape/reel) has two crown jewels: electrode printing (step 3) and co-firing (step 6). These are the two steps where the gap between Murata / Samsung E-M / Taiyo Yuden and the rest of the industry is widest. Co-firing in particular is where decades of process IP compound, because the failure modes are statistical and require enormous yield databases to engineer out. Dongguan frames the same three core technologies as materials technology, overlay printing technology, and co-firing technology.

The yield math is the moat in one line: a 99.5% yield per layer across 1,000 layers gives a cumulative body yield of only 0.7%. Getting to commercially viable cumulative yields above 85% requires mastering every stage simultaneously — which took Murata, TDK, Samsung E-M, and Kyocera AVX decades. A new entrant with comparable equipment (assuming they could even source it, which they can't for Murata's proprietary furnaces) would need 5-10 years of yield learning before commercial viability at the high end. Yield on 008004-size high-cap parts can be below 50% in trial runs. This is why only those four hold production-scale capability for ultra-high-CV designs above 47µF, and it is the reason the co-firing bottleneck is captured inside the Tier-1 equity rather than being a separate alpha.

BOM cost structure — where powder leverage lives

Dongguan's MLCC cost-structure table is the per-unit economics that explains why upstream powder pricing matters more as the mix shifts to high-cap:

Component	Low-capacitance MLCC	High-capacitance MLCC
Ceramic material	20-25%	35-45%
Internal electrode (Ni/Ag/Pd)	5%	5-10%
External electrode (Cu/Ag)	5%	5-10%
Packaging materials	20-30%	1-5%
Labor costs	10-20%	10-20%
Equipment depreciation + other	20-35%	20-30%

Ceramic material rises from 20-25% to 35-45% in high-cap parts because more dielectric layers require more raw material per unit. This is the mechanism by which BaTiO3 powder pricing gains leverage over finished-product margins as the global mix shifts toward high-cap AI and automotive applications. The technical supplement confirms the same figures (20-25% → 35-45%).

End-market BOM builds — content per device

The MLCC is the most populous active component on every modern board, and content per device is the demand driver. Per-device counts:

• AI server: 15,000-25,000 MLCCs per GB200 baseboard (the demand vector that did not exist in 2022). Murata states the GB300 platform requires ~30,000 MLCCs per baseboard — about 30× a mobile phone and 3× an automobile, with a single rack consuming as many as 440,000 individual capacitors. A GB200 NVL72 rack carries 36 baseboards, so a rack approaches ~1 million MLCCs vs ~10,000 in a standard 2U dual-socket Xeon. Total capacitor count per AI-server board is ~10,000-20,000 MLCCs plus 300-600 polymer/hybrid aluminum capacitors. Murata's own guidance: AI servers at 10,000-20,000 MLCC/baseboard (Medium-Term Direction 2027), revised to 15,000-25,000 (FY25-FY30), a 3.3x increase in demand vs FY2025 with miniaturization/high-capacitance demand at 2.0x.
• Smartphone: 1,000-1,200 MLCCs per flagship; mid-tier Android 800-1,000.
• Automotive (TDK powertrain data): ICE 5,000, HEV 6,000, PHEV 8,000, BEV 10,000 MLCCs per vehicle (the primer also cites a 3,000-5,000 BEV vs ~2,000 ICE range). The incremental ~5,000 MLCCs per BEV vs ICE are disproportionately high-voltage, high-reliability auto-grade parts commanding 30-80% price premiums over consumer-grade equivalents.

Powertrain	MLCCs	Inductors	Piezoelectric
ICE	5,000	450	100
HEV	6,000	600	140
PHEV	8,000	650	160
BEV	10,000	700	180

Smart glasses are increasingly using 01005 MLCCs at 150-200 cells per pair (TrendForce). Humanoid robot modules add MLCC demand across control, drive, communication, and power-management systems.

The AI-server BOM in dollar terms — Morgan Stanley generation step-up

The dollar BOM step-up per accelerator generation is the most-cited demand datapoint:

• Blackwell NVL72: ~441,000 MLCCs at ~$0.30 avg ASP ≈ ~$130K/rack.
• MLCC content per rack: $1,530 (GB300 Blackwell) → $4,320 (VR200 Rubin), +182% generation-on-generation per Morgan Stanley NVL72 BOM. (The substack scrape phrases the same figures as GB300 ~$1.5K → VR200 ~$4.3K.)
• Vera Rubin NVL72 potentially 1.5-2M MLCCs per rack (Leo In AI estimate).

The contrarian framing on the BOM build, which is load-bearing for the value-chain thesis, came from the Collyer Bridge subscriber chat (subscriber "Zrave," 21 May, reposted by Collyer as the anchor for his MLCCs thread): "MLCC TAM is 99.99% outside that BOM. The power capacitor shelf is a separate SKU by Liteon & others. PSU passives scale with wattage and obviously the SST itself is outside. Lifetime Rubin MLCC BOM won't cover a single quarter of Yageo revs." And: "The BOM for power & cooling at the datacenter level looks very different than at the rack level." The point: MLCC TAM is wattage-scaled across the whole datacenter, not gated by any single GPU rack BOM. Morgan Stanley's headline VR200 ~$4.3K MLCC content figure (vs GB300 ~$1.5K) is, per the chat, "the wrong way to look at it."

The PDN — why the BOM is structured the way it is

The AI-server capacitor count is a consequence of the power-delivery-network architecture. A single GB200 GPU module consumes 500-800W at 0.65-1.0V; at 800W and 0.8V, peak current approaches 1,000A, with transient swings near 100A/µs that must be compensated within nanoseconds. The PDN places different capacitor types at each conversion stage:

• Facility/rack shelf (277V AC → 48V DC): polymer aluminum capacitors (4-8 per stage, 680-1,500µF each), ESR below 15mΩ — bulk energy storage.
• Intermediate bus (48V → 12V/6V): high-CV X6S/X7R MLCCs in 1206/1210 (100-200 per stage, 100-330µF each), e.g. Samsung's CL31X227 (220µF, 1206) and CL32X337 (330µF, 1210), direct competitors to Murata's equivalents. This is the ultra-high-CV tier where only four suppliers are qualified.
• Near the GPU die (final 0.8V rail, 50-150A/phase): ultra-low-ESL MLCC arrays in 0603/0402 (500-1,000 per stage, 10-100µF each), ESL below 1nH critical, smaller case sizes placed directly around the GPU package.
• Closest to die (>100MHz): embedded capacitors in package substrate/PCB (Empower ECAP, IPDiA), emerging, not yet volume.

The side-gap construction method matters at the printing/stacking stage for AI-server MLCCs: maximizing electrode-area overlap within the body (minimizing side margin) raises effective capacitance per unit volume but demands tighter process control to prevent edge shorts.

The qualification barrier — AEC-Q200 as a moat node in the chain

Auto-grade MLCC sits behind a qualification wall that reinforces the existing supplier structure. AEC-Q200 (Revision E, published March 2023) imposes mechanical shock (50G half-sine, 6ms), bending strength (>3mm substrate deformation), thermal cycling to +150°C, and accelerated life testing (2,000-5,000 hours). Certified parts show failure rates 10-100x lower than industrial-grade. The practical consequence: any new supplier faces a 12-18 month qualification timeline (the primer cites 18-36 months for the full design-in cycle) before shipping a single production part, including PPAP documentation, traceability, and SPC evidence across multiple lots. This is the moat Chinese makers have not crossed at scale in high-end automotive. The auto-grade market structure (2022 volume data): Murata 47.0%, Samsung E-M 14.6%, TDK 14.6%, Taiyo Yuden 10.4% — the 100V-plus powertrain segment is essentially a Murata-TDK duopoly with Samsung E-M trying to break in.

Pricing power by chain position — who sets price and who follows

Pricing power runs high-end-first, commodity-follows, and is asymmetric by tier:

• High-end (high-cap small-form-factor for AI, X8R auto-grade powertrain, ultra-miniature 008004) is structurally undersupplied because expansion needs multi-year retools only Murata / Samsung E-M / Taiyo Yuden can execute. It sees pricing power first and holds longer — roughly the full 18-30 month up-cycle.
• Commodity (0603 X5R, low-cap 0402) has structural overcapacity from Chinese entrants and Tier-2 Taiwanese; pricing follows the high-end with a 1-2 quarter lag and holds perhaps 12-18 months before overcapacity drags it back.

The May 2026 price round shows the full arc: Murata 15-35% on high-end (AI server + auto-grade, effective April 1, 2026), Yageo 10-20% on auto-grade, Taiyo Yuden 6-13% on commodity, Samsung E-M 5-10% under consideration. Walsin's posture — case-by-case on loss-making SKUs only, no broad hike — is the structural tell that Walsin is a price-follower, not a price-setter. Both ends of the curve moving at once is what a true cycle inflection looks like.

There are three competing framings of what the "price hike" actually is, which matters for understanding pricing power in the chain:

• Framing A — real price hikes: the headline 15-35% (Murata), 6-13% (Taiyo Yuden), 5-10% (SEMCO); Banyan Lane models 4-7 rounds through FY30 with AI-MLCC ASP doubling+ from trough.
• Framing B — mix shift, not list-price increase: Taiyo Yuden's sell-side roundtable (May 22-23) said "no MLCC price hikes this fiscal year, but FY3/28 is a possibility"; the ASP driver is migration to smaller, higher-capacity parts (47µF moving from 2012/1608 to 1005, 100µF from 2012 to 1608 in 2H 2026), which lifts ASP at high margin without raising the price of the same SKU.
• Framing C — allocation/priority premiums (the nuance Pink flagged): the closest public version is Banyan Lane — "The shortage is real, but it allocates pain to tier-2/tier-3 customers while tier-1 customers pay through the nose to maintain supply." Murata holds multi-year LTAs with hyperscalers who get full allocation while mid-tier OEMs (Dell, HPE, Supermicro non-priority lines) face shortages. The 2018 parallel: line-stops hit tier-2/tier-3 OEMs (GoPro publicly admitted under-producing cameras; Sony PS4 production was constrained). TrendForce: AI customers mostly use 2-3 year long-term planning, so demand execution is more orderly, making the upward price trend more durable. The distinction: headline 15-35% "hikes" may partly reflect customers voluntarily paying above list to secure allocation priority rather than Murata raising its formal price schedule uniformly. The exact claim "Murata hasn't really raised prices, it's just allocation premiums" was not found verbatim in any public source — its most probable origin is the Collyer Bridge subscriber chat May 17-22 (the "APAC Wrap: 21 May 2026" post cuts off at "Feedback from chat on why this is the wrong way to look at it:" right where the allocation-premium pushback would appear).

Bottleneck-tier analysis — which choke points carry alpha

The primer's bottleneck table ranks the choke points by whether they can be bypassed and whether the market has priced them:

Layer	Smallest pure-play	Approx MC	Concentration	Bypassable?	Market priced-in?
BaTiO3 high-end powder	Sakai Chemical 4078.T	~$500M-$1B	≤4 suppliers >80%	No	Partial
Co-fire process IP	Captured in Tier-1 MLCC equity	n/a	3 firms	No	Yes
Nickel electrode paste	Shoei Chemical 4970.T	<$1B	~5 suppliers	Partial	Unknown
Tape-casting equipment	Hirano Tecseed, NORITAKE	n/a	~3 suppliers	Partial	Unknown
Tantalum powder	HC Starck (private), GAM	n/a	3 suppliers	No	Partial

Top three bottlenecks ranked by alpha potential: (1) BaTiO3 sub-100nm powder via Sakai — small-cap, concentrated, not bypassable, partially priced; (2) tantalum powder for AI-server VRM rails — concentrated, captured downstream by Yageo/KEMET but with private upstream supplier optionality (low investability); (3) nickel electrode paste via Shoei — secondary tightness, lower urgency than powder. The co-firing IP bottleneck is real but captured inside the named winners' equity (it is the reason Murata trades at ~28x forward P/E), so it is not a separate alpha opportunity.

The "follow the capex" trace on Murata's Feb 2026 AI-MLCC capacity-doubling (Q4 FY26 ramp; plus the JPY 80B ~$500M emergency capex on top of the JPY 250B FY2026 plan announced Apr 30) maps the chain again: (1) direct recipient — Murata's own AI-MLCC line (priced); (2) near-competitors — Samsung E-M overflow on dual-source requirements and Taiyo Yuden at 22µF SKUs (partially priced); (3) upstream — Sakai (BaTiO3 to feed the doubled line), Shoei (nickel paste), Hirano Tecseed / NORITAKE (tape casters), co-fire furnace builders (less priced); (4) second-derivative — tantalum powder suppliers as KEMET/Yageo win sockets around Murata MLCC, EV-inverter film-capacitor names. The pattern matches NVDA's 2024 EML preallocation to LITE/COHR/Sumitomo: the named recipient is priced in weeks, the upstream layer takes 6-18 months.

Supply-demand balance through the chain

Capacity utilization rebuilt from low-70s (Q3/Q4 2024 trough) to mid-80s by Q2 2025, to 87-88% in 2H25, approaching the ~90% tight S/D threshold by mid-2026. The Dongguan capacity-utilization reads: Murata 90-95% with zero server-field backlog and capacitor B/B rising 1.04 → 1.01 → 1.12 across FY25Q1-Q3; Samsung E-M plant utilization above 96% (Busan, Tianjin, Philippines) rising to 99% in Q3 2025; Yageo B/B above 1 in all end markets, AI strongest. The substack scrape adds: Murata production lines at 98% utilization, customer inquiries ~2x supply capacity, "still very much insufficient" (President Nakajima); SEMCO at 99-100% utilization with Philippines Plant 3 not shipping until July 2027+.

The supply gap is structural because the chain cannot respond inside a thesis horizon. Capacity additions take 24-30 months from capex commitment, customer qualification adds 6-12 months, and the dielectric powder comes from 3-4 qualified suppliers. Murata's Izumo plant started shipping April 2026 — 25 months from announce; SEMCO's larger Korean expansion is not until mid-2027. Banyan Lane models the annual deficit widening from -20B units (2026) to -160B units (2030E); industry capacity additions of 25-35% through end-2028 run against 80%+ CAGR AI-MLCC demand. The strategic-segment economics underpin the margin pool: strategic MLCCs (high-voltage, polymer electrolytic, ultra-high-CV) are 35-40% of market value but 60-70% of industry profit, growing 12-17% CAGR with 30-80% price premiums and 24-47 week lead times; commodity MLCCs face -10-20% pricing pressure from Chinese capacity. Lead times stretched from 8 weeks (normal) to 24-40 weeks through 2026, with hyperscalers (Google, Meta, Microsoft, NVIDIA) locked into 60-70% of available high-CV capacity. High-CV MLCC (220µF/4V, AI GPU rail) prices rose +29% from 2024 to Q1 2026 per the Capacitor Dossier.

China domestic-substitution dynamics in the chain

China imported 2.56 trillion MLCCs worth US$6.179 billion in 2025; imported unit price ~US$2.41 per thousand vs export US$2.11 per thousand — the gap reflects the higher value of imported high-end product. If 50% of 2025 import volume were domestically replaced, domestic substitution would reach 1.28 trillion units. The Chinese government has explicitly funded Fenghua, Sunlord, and others under passive-component import-substitution policy. Structural impact is concentrated on the commodity tier (0603/0402 low-cap) and Tier-2 Taiwanese pricing power — not yet credibly hitting Tier-1 Japanese or Samsung E-M, who sit in SKUs Chinese entrants cannot yet make at scale. In China's powder layer, Guoci, Fenghua, and Sanhuan have achieved comprehensive coverage of base and formulation powders but still need high-end breakthroughs; Samsung E-M, Yageo, and Fenghua are downstream powder customers. Fenghua and Sanhuan are making progress in consumer/industrial MLCC and have begun small-size high-capacitance products, but have not crossed the AEC-Q200 high-end auto wall.

Geographic distribution of the chain

Revenue by geography: Japan ~50% (Murata + Taiyo Yuden + TDK), Korea ~22% (Samsung E-M), Taiwan ~12% (Yageo + Walsin + Holy Stone + PDC), China ~10% (Fenghua + Sunlord + others), RoW ~6% (legacy Vishay, Kemet pre-Yageo, Kyocera AVX). Demand is heavily concentrated in China at ~60% of consumption (assembly hub for smartphones, AI servers, EVs, consumer electronics), with US/EU end demand routing through China assembly. The geopolitical risk is more about end-market exposure than input restrictions — MLCCs are not on any major export-control list, though a Chinese heavy/medium rare-earth export restriction (BaTiO3 doping uses dysprosium, yttrium, holmium) would complicate non-Chinese production within 6-12 months.

Players

The MLCC industry is one of the more concentrated component markets in semiconductors — comparable to memory NAND/DRAM in degree of consolidation but with more stable competitive ranking. Four Tier-1 makers (Murata, Samsung Electro-Mechanics, TDK, Taiyo Yuden) hold roughly 77% of global MLCC revenue; adding Yageo gets you to roughly 83%. The remaining ~17% is split across Tier-2 Taiwanese, Chinese new entrants, and legacy small players. The mistake is treating MLCC makers as a homogeneous group — each Tier-1 has a different optimal end market, and the right way to read the peer set is by who wins what.

The three-tier structure

Per Dongguan Securities (Mar 2026), the global MLCC industry divides into three tiers:

• Tier 1 — Japan and Korea. Murata and Samsung Electro-Mechanics. Decades of accumulated technology, product, and customer expertise. Specialize in high-value-added products — small size, high capacity, high voltage. Target automotive, high-end smartphone, AI server.
• Tier 2 — Taiwan. Yageo and Walsin. More diverse product portfolios, mainly low-to-mid-range, actively expanding into automotive and industrial. Capacity continuously increasing.
• Tier 3 — Mainland China. Technology still lags Tier 1 and 2. Primarily medium-to-large-size, low-capacitance. Recent breakthroughs in several small-size, high-cap products. Representative names Fenghua Advanced Technology and Sanhuan Group.

Global MLCC market share (2024)

Two share frames appear across the sources. The primer's widely-cited 2024-2025 numbers:

Company	Ticker	Share	Tier
Murata	6981 (6981.T)	~34%	1
Samsung Electro-Mechanics	009150.KS	~22%	1
TDK	6762.T	~12% (MLCC)	1
Taiyo Yuden	6976.T	~9%	1
Yageo	2327.TW	~6% (MLCC; #1 tantalum, #1 chip resistor)	2
Walsin	2492.TW	Taiwan #2	2
Holy Stone	3026.TW	Tier-2/3 hybrid	2/3
PDC	6173.TWO	sub-scale finished MLCC + powder	2

Dongguan's 2024 share table (different methodology, lower absolute numbers but same ranking):

Company	Share	Tier
Murata (村田)	31.8%	1
Samsung Electro-Mechanics (三星电机)	11.8%	1
Taiyo Yuden / TDK / Kyocera	(top-5 combined)	1
Top 5 combined	83.6%	—
Yageo (国巨)	5.0%	2
Walsin (华新科技)	3.2%	2
Fenghua (风华高科)	2.5%	3
Sanhuan Group (三环集团)	1.9%	3
Huaguo Technology (华国科技)	1.5%	3

In automotive-grade MLCC (2022 volume data): Murata 47.0%, Samsung Electro-Mechanics 14.6%, TDK 14.6%, Taiyo Yuden 10.4%. In AI server MLCC, Murata and Samsung Electro-Mechanics dominate, with Samsung holding ~40% of the high-end AI server MLCC market (Maeil Business Newspaper / Digitimes). Only four suppliers are qualified for ultra-high-CV MLCC at production scale: Murata, TDK, Samsung Electro-Mechanics, and Kyocera AVX.

The named manufacturers — positioning

Murata (6981.T) — ~34% global, the price-setter. 6981 is Tier-1 across every segment and dominant in the high-end (high-cap, 008004, AEC-Q200 100V-plus auto, premium-smartphone power management). ~45% of corporate revenue is MLCC. Owns its own barium-titanate powder, so it is structurally self-sufficient on the upstream bottleneck. Fortress balance sheet (~¥600B net cash). Achieves the industry's highest layer count (1,600 layers) and produces a 220µF MLCC in a 1206 case. The reference price-setter, quality, and balance sheet — trades at premium valuation because it earns it (~28x forward P/E). Wins high-end auto plus AI plus premium smartphone. The cleanest moat in AI MLCC at ~60% high-CV share within the 4-supplier oligopoly. Ranked #1 BUY (scale-in) across both the sector deep-dive and the 10-ticker peer swarm. Subscriber chat (Collyer Bridge) reads community-bullish ("Bullish Murata," "Bought TDK/Murata to make my MLCC play 10% of portfolio"), with Henry siding with Murata on ultra-high-CV pricing power.

Samsung Electro-Mechanics (009150.KS) — ~22% global, the AI-server volume leader. The other Tier-1, volume leader in AI-server MLCC (~40% share per Digitimes), growing in 100V-plus auto. Its CL31X227 (220µF, 1206) and CL32X337 (330µF, 1210) products are the direct head-to-head competition for Murata's ultra-high-CV equivalents. Chaebol-discount valuation (~16x forward P/E vs Murata's 28x) is the cheapest legitimate entry to Tier-1 MLCC exposure — the catch is the governance discount and the chaebol structure that complicates capital returns. The sector deep-dive flags it as AVOID at the 23 May print (9× in 12 months, 4× analyst PT dispersion KRW 250,000-1,050,000, basket 13/30 Hard Pass), but the primer frames it as a core Tier-1 holding on the discount. Note the silicon-capacitor overlay: SEMCO is one of four silicon-cap suppliers (Murata/SEMCO/AP Memory/TSMC), and the non-MLCC primer reworked SEMCO fair value to KRW 800K-1.25M after integrating Citi's W609bn FY27E / W905bn FY28E silicon-cap revenue path and the W1.6T 2-year supply order — still implying 32-55% downside vs the ~KRW 1.85M price. Owning Samsung Electronics (005930.KS) for SEMCO exposure does not work — SEMCO is only ~1.7% of that mcap. Knowledge gap: SEMCO had not been swarm-researched at primer time; the company-specific case needs a /profile run.

TDK (6762.T) — ~12% global on MLCC, but you cannot buy it as an MLCC story. Tier-1 on 100V-plus auto MLCC and AI-server passives, but MLCC is only ~10% of TDK revenue. The corporate story is dominated by ATL Apple silicon-anode batteries (55% of revenue, 89% of segment OP) and HDD heads (a separate duopoly thesis). Highest corporate quality in the swarm (ISS QualityScore 1, ROIC-and-TSR PSU hurdles, majority-outside board). Own TDK for combined Apple-battery / HDD-head / AI-passive / auto-MLCC exposure with MLCC as bonus content — not the other way around. The only Japanese MLCC name still just below its sell-side target (BofA Buy, PO ¥3,500 vs ¥3,370). Ranked #2 in the peer swarm; "honorable mention" Japan-MLCC pivot in the sector deep-dive if Murata's run feels too extended.

Taiyo Yuden (6976.T) — ~9% global, the pure-play with the highest cycle beta. 71% of revenue is MLCC — the only major where MLCC is the dominant economic line, so no diversification dilution. Strongest in smartphone high-cap and increasingly in AI server. First to ship 1005-size 22µF (Aug 2025) and 0402-size 22µF (Sept 2025), and launched a 4532-size MLCC up to 1,000µF. The highest cycle beta in the complex — the right entry at the FY3/25 trough delivered +267% in 12 months, but at the 23 May print it is the worst current entry math. Sector deep-dive: AVOID (BofA Underperform PO ¥3,800 vs spot ¥9,102 = 139% above target; basket Hard Pass on Revision Velocity 1/5 — estimates cut from ¥130 to ¥90 while the stock rallied 77%). Pricing posture is the lag-confirmation signal: +6-13% commodity hikes in May 2026, but its own sell-side roundtable (May 22-23) said "no MLCC price hikes this fiscal year, FY3/28 a possibility" — ASP lift comes from mix shift to smaller higher-cap parts, not list-price hikes. Collyer chat sentiment mixed ("scammed by ppl who don't like Taiyo Yuden"). Primer Tier-3 watchlist, only on a 25-30% pullback.

Yageo (2327.TW) — ~6% global on MLCC, but #1 global tantalum and #1 chip resistor; the consolidator. The cleanest multi-passive roll-up in the complex (KEMET 2020, Chilisin 2021, Nexensos 2023, Telemecanique 2023, Shibaura 2025). The accurate frame is that Yageo is a tantalum name with MLCC and chip-resistor optionality — its AI-server economic exposure is proportionally larger via tantalum (~46% share via KEMET) than via MLCC, because the tantalum thesis is more concentrated and the NVIDIA-platform qualification cycle is multi-generation. Pierre Chen's ~20% personal ownership is the strongest founder alignment in the peer set. Fair price for what it owns; cycle-leveraged but quality-discount-priced. Sector deep-dive: AVOID at 23 May (NT$629, +10% limit-up vs GS PT NT$346 = 82% above target). Yageo is the single most-discussed name in the Collyer Bridge chat (14+ mentions) — community consensus long, Collyer publicly bullish ("Oh shit Yageo almost limit up," "Everybody I talked to yesterday loves Yageo") but privately preferring Kingboard. The Zrave anchor quote that frames the whole chat thesis: "MLCC TAM is 99.99% outside that BOM... Lifetime Rubin MLCC BOM won't cover a single quarter of Yageo revs" — i.e., MLCC TAM scales with whole-datacenter wattage, not single-rack BOM.

Walsin Technology (2492.TW) — Taiwan #2 MLCC, fast-follower not price-setter. Family-controlled (Chiao group, 41.7% insider). Higher cycle beta than Yageo, lower than Taiyo Yuden. The structural admission: case-by-case-only pricing on loss-making SKUs in May 2026 (no broad hike) is the tell that Walsin is a price-follower. The right name for second-leg cyclical torque (China utilization recovery from ~70% toward 85%-plus, real operating leverage) — but wrong entry above TWD 225-250. Lead times: 22 weeks high-cap, 12-14 weeks low-cap.

Holy Stone (3026.TW) — Tier-2/3 hybrid, manufacture-plus-distribution. The mfg-vs-distribution revenue split is non-disclosed and is the single largest information gap in the publicly-listed peer set. Speculative. Sector verdict: AVOID at current prices — parabolic on the smallest float with the weakest underlying business in the listed peer set, trading at peak-cycle multiple without peak-cycle earnings. Ranked last in the peer swarm.

Prosperity Dielectrics / PDC (6173.TWO) — sub-scale finished MLCC + powder. The powder business is structurally more interesting than the finished-MLCC business. PSA-group ally of Walsin. Primer Tier-3 watchlist, only on a 50%-plus pullback and only after the June 11 Q1 2026 print confirms the powder business is operating as the bull thesis requires (53x trailing P/E makes it asymmetric to the downside). Related-party transaction pricing on Walsin powder sales is not verifiable in English-language disclosure — a Taiwan TPEx disclosure gap.

Chinese new entrants — Fenghua Advanced Technology (000636.SZ), Sanhuan Group (300408.SZ), Sunlord Electronics (002138.SZ), Guangdong Fenghua-Aurora, Shenzhen Eyang. Aggressive in commodity 0603/0402, struggling at 0201, not shipping 008004 at scale. The China commodity tail is the secular risk to mid-cap pricing but not to high-end pricing — government-funded capacity is concentrated where Tier-1 makers do not compete. Difficult to own for foreign investors on disclosure and liquidity grounds; the right way to play the China-commodity tail is to short the Taiwan Tier-2 names rather than long the Chinese names. Dongguan rates Fenghua (000636.SZ, RMB 20.10, PE 2026E 47.86x) and Sanhuan (300408.SZ, RMB 54.24, PE 2026E 31.72x) Buy and Hold on domestic-substitution upside. China's substitution arithmetic: 2025 imports were 2.56 trillion MLCCs worth US$6.179B (import unit price ~US$2.41/thousand vs export US$2.11/thousand), and replacing 50% of import volume domestically would equal 1.28 trillion units.

Upstream — barium-titanate powder suppliers

The qualified high-end powder layer (sub-100nm, monodisperse, doped Class II) is the bottleneck the market has not fully priced. Global BaTiO3 ceramic-powder market share (Dongguan):

Company	Country	Share
Sakai Chemical (堺化学)	Japan	28%
Ferro (美国Ferro)	USA	20%
Nippon Chemical (日本化学)	Japan	14%
Guoci Materials (国瓷材料)	China	10%
Fuji Titanium (富士钛业)	Japan	9%
Kyoritsu (共立)	Japan	8%
Toho (东邦)	Japan	6%
Other	—	5%

Top 5 (Sakai, Ferro, Nippon Chemical, Guoci, Fuji Titanium) hold 81% — highly concentrated. The Japanese tech edge is concrete: Japanese producers achieve 80-100nm average particle size via hydrothermal synthesis (Sakai's core IP) vs Chinese 120-150nm, which is why Japanese makers stack 1,200 layers at 0.5-0.6µm (Murata 1,600) vs Chinese 800 layers at 1-2µm.

• Sakai Chemical (4078.T) — the upstream alpha. The cleanest pure-play exposure to the sub-100nm BaTiO3 bottleneck, ~3-5% of Murata's market cap (~$500M-$1B). 28% of the merchant powder market (the "40-50% merchant share" in Sakai's own profile is overstated — see the correction note in the primer). Flagged by three separate sub-agent runs (Taiyo Yuden, TDK, Walsin) as the under-priced theme exposure: less consensus crowding, less retail-flow distortion, same secular thesis, powder pricing lagging MLCC pricing by 2-3 quarters. Catalyst: FY3/26 results pack (late May 2026). Currently WATCH not Buy (~¥3,465). Sector deep-dive flagged +3% on 23 May vs +6-12% across the rest of the complex, 4.62% div yield, no analyst coverage.
• Nippon Chemical Industrial (4092.T) — the cleaner starting-condition candidate. 14% powder share. Per the Sakai sub-agent, cleaner starting conditions than Sakai (+22% YoY revenue vs Sakai's +1.9% guide) at similar P/E and EV/EBITDA. Ranked above Sakai and SEMCO in the peer swarm.
• Guoci Materials (China), Ferro (US), Fuji Titanium (Japan) round out the qualified set. Murata makes its own BaTiO3 internally (MF Material JV with Ishihara Sangyo and Fuji Titanium formed Sept 2023 to internalize supply), structurally diluting external powder dependence over time.

Other upstream — nickel paste, copper, equipment

Nickel paste is a smaller bottleneck than powder (Tier-1 makers can substitute among ~5 qualified suppliers more readily than for ceramic powder). Named suppliers: Shoei Chemical (4970.T) — the most MLCC-specific specialty supplier; Mitsui Mining & Smelting (5706, 5706.T); Sumitomo Metal Mining (5713.T). Copper termination paste is commoditized. The equipment layer — tape casters (Hirano Tecseed, NORITAKE), stack/lamination presses (Mitsubishi Heavy), co-fire furnaces — is concentrated among Japanese specialty makers but with no pure-play. Shoei is the secondary upstream-alpha flag after Sakai.

Non-MLCC capacitor competitors

Three capacitor classes compete for board sockets, each holding a niche MLCC does not displace (detail on the non-MLCC primer):

• Tantalum (Yageo/KEMET, Kyocera AVX, NEC Tokin): holds the AI-server VRM-rail bulk-decoupling niche; Yageo's KEMET acquisition was driven by NVIDIA Hopper/Blackwell design-wins. Upstream tantalum powder (HC Starck private, GAM) is a captured-downstream bottleneck.
• Aluminum electrolytic / polymer aluminum (Nichicon, Rubycon, Panasonic, Chemi-Con, TDK): bulk hold-up on AI-server PSU output and motherboard rails.
• Film (Vishay, TDK Epcos, Panasonic, Nichicon): the EV-inverter DC-link / high-voltage power-conversion niche, not threatened by MLCC at 600V+.
• Silicon capacitors (Murata, SEMCO, AP Memory 6531.TW, TSMC 2330.TW): a 4-supplier oligopoly for IC-package decoupling (EMIB / CoWoS-S / I-Cube) — flagged as a "supposedly superior type of capacitor" by Collyer Bridge, with AP Memory the purest play and SEMCO the volume name.
• Supercapacitors / LIC (Skeleton private, Musashi 7220.T): rack-level energy buffering. Musashi PASS at ¥7,960 because LIC is <1-2% of consolidated revenue.

Adjacent / cross-referenced names (Collyer Bridge chat)

The Collyer Bridge subscriber-chat sweep surfaced names in the broader passives / materials ecosystem that bleed into the MLCC thesis. Collyer's privately-preferred passive name is Kingboard (0148.HK holdco / 1888.HK Laminates subsidiary) — upstream CCL/PCB — over Yageo, and Kingboard Holdings (0148.HK) was the sector deep-dive's #2 pick (the only name in the universe trading below its mean PT). Other chat names: Co-Tech (8358.TW) Taiwan copper foil (Collyer actively accumulating), Furuya Metal (7826, 7826.JP) high-purity ruthenium for thick-film resistors / HDD sputtering targets (~70% share, AI-driven Ru demand thesis), and Musashi Seimitsu (7220.JP) 800V supercapacitors (Collyer entered then exited within hours on valuation discomfort).

Comparative analyses to reference by name

The head-to-head and ranked comparisons that sit behind this peer view:

• MLCC Peer Swarm — 10-ticker ranked verdict (briefings/2026-05-25-mlcc-peer-swarm.md, full report ~/claude/output/compare/ portal). Ranked verdict: Murata (BUY scale-in) > TDK > Yageo > Walsin > Taiyo Yuden > Nippon Chemical > Sakai > SEMCO > PDC > Holy Stone. Seven sell-side sources cross-referenced (BofA, GS, JPM, Capacitor Dossier, Global Passives Basket, Dongguan Securities, sell-side meeting notes).
• MLCC Sector — Top 2 With Crowding in Mind (briefings/2026-05-23-mlcc-sector.md). Picks: #1 Murata (6981.T, High scale-in), #2 Kingboard Holdings (0148.HK, starter + add); honorable mention TDK; AVOID list Taiyo Yuden / SEMCO / Yageo / Walsin / Holy Stone / PDC.
• Murata vs Yageo desk note and a multi-name comp sheet (the 7-name comp authored by chat subscriber "Gman Scotchnova," reposted by Collyer) — both circulated in the Collyer chat as Google Drive PDFs alongside the BofA note.
• Sell-side consensus merge — ~/claude/output/compare/mlcc-sellside-consensus-2026-05-24.md (seven sources).
• Individual canonical ticker pages for company-specific deep detail: 6981, 2327, 6976, 6762, 3026, 2492, 6173 under KB/wiki/{ticker}/.

The crowding overlay — why ranking ≠ buying

The single most important investor-relevant fact about the peer set is that owning seven MLCC names does not diversify the cycle risk — it concentrates it; correlation across the group is high. The May 21 Global MLCC & Passives Basket scored Murata 21/30 Buy (Moat 5/5, Crowding 3/5) and Kingboard 19/30 Watch (Crowding 3/5) — both moderate — while Yageo, Walsin, Holy Stone, Samsung E-M, and Taiyo Yuden all scored Crowding 1-2/5 on limit-up tape and 4-9× moves over 12 months on flat-to-cut earnings revisions. On 23 May 2026 every Taiwan passives name went limit-up at +10%, Taiyo Yuden +12%, Samsung E-M +11%, and every name in the complex except Kingboard traded above its sell-side price target. The right portfolio construction is one to two names at conviction sizing plus one upstream name (Sakai) if the bottleneck thesis is taken seriously — not the whole basket.

Monitor

This is the rolling watch-log for the MLCC / passive-components complex — dated developments, pricing moves, earnings signals, and the standing checklist of leading indicators. It consolidates the dated framing scattered across the sector briefings (2026-05-23 deep-dive, 2026-05-25 peer swarm, 2026-05-28 non-MLCC primer), the Dongguan Securities initiation (2026-03-25), the Substack pricing scrape (2026-05-25), and the Collyer Bridge chat sweep (2026-05-28). Company-specific theses live on the ticker pages — see 6981 (Murata), 009150 (Samsung Electro-Mechanics), 6976 (Taiyo Yuden), 6762 (TDK), 2327 (Yageo), 2492 (Walsin), 3026 (Holy Stone), 6173 (Prosperity Dielectrics), 4078 (Sakai Chemical), 4092 (Nippon Chemical), 0148 (Kingboard Holdings), 1888 (Kingboard Laminates), 6531 (AP Memory), 7220 (Musashi Seimitsu), 8358 (Co-Tech), 000636 (Fenghua), 300408 (Sanhuan).

The standing read in one paragraph

The cycle bottomed Q3/Q4 2024. As of late May 2026 we are 12-18 months into a typical 18-30 month up-cycle, with a plausible peak in late 2026 or early 2027 if AI capex extends through 2027 as consensus assumes — earlier (mid-to-late 2026) if AI capex digestion arrives early. Pricing turned on the high-end first (Q4 2025, when Murata's first AI-MLCC SKUs moved) and on the commodity tier last (the broad May 2026 hike round). Both ends moving is what confirms a turn rather than a bottom. The dominant risk that overrides everything else is AI capex digestion: a 15%+ cut to 2026/2027 capex guidance from any of AMZN, MSFT, META, GOOG, or ORCL would cascade through the complex within days, exactly as the 2017-2018 cycle ended via Apple smartphone-unit misses.

Dated developments log (most recent first)

2026-06-01 — Archetype Capital "MLCC Capacity Crunch" on Sakai (4078) + Nippon Chemical (4092) (archetype-research.com). Frames the premium-BaTiO3 layer as a capacity crunch: by production route the market is ~solid-state 61.8% / oxalate 21.4% / hydrothermal 13.5%, so the two premium routes (Sakai's hydrothermal, Nippon's oxalate) are only ~35% of supply — exactly where AI-server demand skews. Server MLCC ~$1.3B in 2025 ($600M AI + $700M general; Zephyr, 29 May) at 80%+ AI CAGR; smartphone MLCC negative 2026-27. Prefers Nippon Chemical over Sakai (same ordering as the peer-swarm here). Author flags the work as incomplete/DYOR. See 4078/4078, 4092/4092.

2026-05-29 (Collyer Bridge daily chat thread, fresh): Collyer opened the daily thread with "Here come the upgrades" (🔥 x2) — flagging sell-side broker upgrades arriving on a name he is long. Subscriber Hanbonjovi posted an FT-article screenshot (content not extractable). Tone bullish but context-free without the image. Flagged to re-check in 12-24 hours as the thread matures.

2026-05-28 — silicon-capacitor turn in the Collyer Bridge chat. Collyer, who had previously avoided MLCCs ("Its pretty wild i did not want to get involved in MLCCs now researching a supposedly superior type of capacitator"), turned newly bullish on silicon capacitors via the pequityresearch AP Memory primer ("I'm starting to believe"). Thread direction: AP Memory > SEMCO > Murata for pure-play silicon-cap exposure, with SEMCO parabolic-valuation friction and Murata segment-disclosure opacity flagged as risks. Jukan's forthcoming SEMCO / silicon-cap X memo is the anticipated next datapoint. Same day, the Collyer Bridge post "Small Parts, Big Cycle: Passive Components" published, framing Yageo getting bid on tier-2 spillover — intellectual antecedent is the 21 May Zrave chat quote (below).

2026-05-28 — non-MLCC capacitor primer published (2026-05-28-non-mlcc-capacitor-primer). Two adjacent categories now on the watch list: silicon capacitors (IC-package decoupling; EMIB / CoWoS-S / I-Cube use case; 4-supplier oligopoly Murata / SEMCO / AP Memory / TSMC) and supercapacitors / LIC (rack-level energy buffering; Skeleton private, Musashi consolidated). Key calls: SEMCO fair value reworked to KRW 800K-1.25M after integrating Citi's W609bn FY27E / W905bn FY28E silicon-cap revenue path and the W1.6T 2-year supply order — still 32-55% downside vs the then-current KRW 1.85M. Musashi (7220.T) PASS at ¥7,960 (LIC <1-2% of consolidated revenue; watch zone ¥4,500-5,500). Owning Samsung Electronics for SEMCO exposure does not work (SEMCO is 1.7% of mcap).

2026-05-28 — Musashi Seimitsu (7220) flagged then dropped (Collyer chat). Musashi's 800V hybrid-supercapacitor (HSC) ramp: 200,000 cells in 2026 → 6.5M cells in 2028; 3.8V HSC vs traditional EDLC 2.8V; Delta Electronics as downstream customer; quoted excerpt claimed "PE ratio in 2028 is less than 10." Nomura raised target to ¥7,700 (fair PE 22.5x on 28/3 EPS, up from 12x, switching to sum-of-the-parts). Subscriber OOO flagged the 100% gap between the 7x-PE excerpt and 22x consensus. Collyer entered then exited within hours ("I'm out … Abit of debate in my dms"). Made-in-Japan flagged the HSC ramp "keeps getting delayed."

2026-05-25 — Substack pricing scrape. Three competing pricing framings surfaced — and the one Pink remembered (Murata "hasn't really raised prices, just allocation premiums") was NOT found verbatim in any public post; it most likely lives behind the Collyer Bridge subscriber-chat paywall ("APAC Wrap: 21 May 2026", at the exact "feedback from chat on why this is the wrong way to look at it" cut-off). Murata FY2025 earnings (Apr 30) detail: JPY 80B (~$500M) emergency MLCC capex on top of the JPY 250B FY2026 plan; President Nakajima called capacity "still very much insufficient" (まだまだ不十分), decision made "in haste" (急ぎ); customer inquiries running ~2x supply (Bloomberg, Feb 2026); production lines at 98% utilization; multi-year LTAs give tier-1s full allocation; JPY 150B buyback (largest in company history) announced same day. Banyan Lane models 4-7 hike rounds through FY30 with AI MLCC ASP potentially doubling-plus from trough; next round modeled Sept 2026 - Feb 2027 at 20-35% on AI parts. Supply/demand deficit widens from -20B units (2026) to -160B units (2030E). SEMCO at 99-100% utilization; Philippines Plant 3 doesn't ship until July 2027+. Industry capacity additions 25-35% through end-2028 vs 80%+ CAGR AI-MLCC demand.

2026-05-25 — peer swarm published (2026-05-25-mlcc-peer-swarm). Ranked verdict across 10 tickers: Murata (BUY scale-in) > TDK > Yageo > Walsin > Taiyo Yuden > Nippon Chemical > Sakai > SEMCO > PDC > Holy Stone. Seven sell-side sources cross-referenced (BofA, GS, JPM, Capacitor Dossier, Global Passives Basket, Dongguan Securities, sell-side meeting notes).

2026-05-24 — Taiyo Yuden sell-side roundtable (reported via Collyer "APAC Wrap: 24 May"). Pricing framing B confirmed: "No MLCC price hikes this fiscal year, but FY3/28 is a possibility." ASP lift comes from mix shift, not list-price increases — 47µF migrating from 2012/1608 to 1005, and 100µF from 2012 to 1608 in 2H 2026. Reconciles with the headline "price-hike" narrative: high-end ASP is rising via harder-to-make parts plus allocation premiums, not uniform list-price moves.

2026-05-24 (Sunday) — Co-Tech (8358.TW) accumulation signal (Collyer chat). Collyer: "Co-Tech is out of disposition on Friday, will be adding more for sure." Taiwanese copper-foil maker, relevant to the laminate / capacitor / PCB supply chain. Freshest discrete actionable signal in the 28 May sweep.

2026-05-23 — complex-wide limit-up day and the deep-dive briefing (2026-05-23-mlcc-sector). Every Taiwan passives name went limit-up +10%; Taiyo Yuden +12%; Samsung E-M +11%; Murata closed ¥7,130 (already 27% above the 17-analyst mean PT of ¥5,203 and 15% above BofA's PO ¥6,200). Five independent sources converged the same week (BofA, the Murata-vs-Yageo desk note, the six-dimension Global Passives Basket, Goldman Sachs on Yageo, Banyan Lane on Murata). Crowding asymmetry was the day's real signal: Murata scored 21/30 Buy (Moat 5/5, Crowding 3/5) and Kingboard 19/30 Watch (Crowding 3/5); Yageo, Walsin, Holy Stone, Samsung E-M and Taiyo Yuden all scored Crowding 1-2/5 on limit-up tape with 4-9x moves over 12 months against flat-to-cut earnings revisions. Taiyo Yuden Revision Velocity 1/5 (estimates cut ¥130 → ¥90 while the stock rallied 77%) was called the rubric's worst case. Top-2 verdict: Murata (structural anchor, scale-in below ¥6,500) + Kingboard Holdings 0148.HK (the only name in the universe with mean PT, HK$64, ABOVE spot HK$58.30; 9.88x fwd P/E; 3.81% div yield). Sakai +3% on the day vs +6-12% across the rest of the complex — the non-consensus laggard.

2026-05-21 — the Zrave anchor quote (Collyer chat, 21 May daily thread, 333 replies). The day the passive-components thematic crystallized. Subscriber Zrave: "MLCC TAM is 99.99% outside that BOM. The power capacitor shelf is a separate SKU by Liteon & others. PSU passives scale with wattage and obviously the SST itself is outside. Lifetime Rubin MLCC BOM won't cover a single quarter of Yageo revs … The BOM for power & cooling at the datacenter level looks very different than at the rack level." Collyer reposted it as the anchor of his dedicated MLCCs thread (signalling endorsement). Same day Collyer dropped three sell-side notes (Yageo-vs-Murata, a Gman Scotchnova 7-name comp, BofA), reacted "Oh shit Yageo almost limit up," noted "Everybody I talked to yesterday loves Yageo," but privately prefers Kingboard over Yageo (per Gman's "you like kingboard better correct?"). Jukan dropped Morgan Stanley charts on MLCC and ABF content GB300 vs VR200.

2026-05-21 — Murata buyback / Izumo ramp context. JPY 150B buyback (largest in company history); Izumo plant started shipping April 2026 (25 months from announce). Prosperity Dielectrics doubling capacity; Yageo +15% move in the window.

May 2026 — broad pricing round (cycle-confirmation signal). Murata 15-35% on AI/auto-grade high-end (effective Apr 1, 2026); Yageo 10-20% on auto-grade; Taiyo Yuden 6-13% on commodity (the lag-confirmation); SEMCO 5-10% signaled / under consideration. Walsin took case-by-case-only on loss-making SKUs with no broad hike — the structural tell that Walsin is a price-follower. High-end leads, commodity follows: this is the canonical "cycle confirmed" pattern.

2026-04-30 — Murata FY2025 earnings. Capacitor B/B ratio 1.12 in Q3 (up from 1.04 → 1.01); capacitor backlog +89.5% to ¥269.2B; computers/servers application +28.4% YoY; FY2025 guidance revised upward; capacity utilization 90-95%; server-field product inventory at zero backlog. Emergency JPY 80B capex announced (above the JPY 250B FY2026 plan).

2026-03-25 — Dongguan Securities initiation (dongguan-mlcc-initiation-2026-03-25). Overweight (Maintain), analyst Luo Weibin. Thesis: supply-demand imbalance intensifying, high-end MLCC prices expected to rise. Murata decided 15-35% hikes on AI-server and high-end auto MLCC from April; SEMCO double-digit hikes from April. 2024 global MLCC share: Murata 31.8%, SEMCO 11.8%, top-5 combined 83.6%. BaTiO3 powder share: Sakai 28%, Ferro 20%, Nippon Chemical 14%, Guoci 10%, Fuji Titanium 9% (top-5 = 81%). Japanese 80-100nm vs Chinese 120-150nm particle size; Japanese 1,200 layers at 0.5-0.6µm (Murata 1,600) vs Chinese 800 layers at 1-2µm. China imported 2.56 trillion MLCCs worth US$6.179B in 2025 (import unit price US$2.41/k vs export US$2.11/k) — 50% domestic substitution = 1.28 trillion units. Investment targets: Fenghua (000636.SZ) and Sanhuan (300408.SZ), both Buy and Hold.

February 2026 — Murata capacity-doubling / inquiry signal (Digitimes, Bloomberg). Murata confirmed AI-MLCC capacity doubling with a Q4 FY26 ramp; MLCC order inquiries running ~2x current production capacity; price-hike decision planned for end-March based on actual demand. This broke the Apr-1 hike story.

Q4 2025 — high-end pricing turned first. Murata's first AI-MLCC SKUs began moving on price — the leading edge of the cycle inflection, one to two quarters ahead of the commodity tier.

Q3/Q4 2024 — cycle trough. Utilization at tier-1 makers in the low 70s; the bottom from which the current up-cycle runs.

Standing watch-items / checklist (leading indicators)

The cycle-on confirmation set, in priority order:

1. Murata quarterly book-to-bill — the single most important leading indicator. >1.0 sustained for two quarters = cycle-on; <0.9 sustained for two quarters = cycle-off. Currently 1.12 (FY25 Q3). Followed by Samsung E-M's and Taiyo Yuden's BB.
2. AI MLCC ASP on the high-CV 220µF/4V reference part — already +29% from 2024 through Q1 2026; TrendForce projects +30-40% for full-year 2026. The best read on whether pricing discipline is holding.
3. Murata capacitor backlog + computers/servers application growth — currently +89.5% backlog, +28.4% computers/servers YoY. A sustained sub-+20% reading on either = AI-MLCC demand normalizing.
4. Murata's AI-MLCC share trajectory — Samsung E-M design-win cadence with hyperscalers is the leading indicator. Banyan Lane downside scenario at 50% share (vs working 60%) materially weakens FY28 EPS.
5. Lead times across the high-end stack — 24-40 weeks today vs 8 weeks normal; >1µF ceramic at 18-22 weeks and rising across Murata/Taiyo Yuden/SEMCO/TDK/Yageo (Q1 2026). Return toward 12-16 weeks = supply catching up; further extension = more cycle room.
6. Hyperscaler capex guidance (AMZN, MSFT, META, GOOG, ORCL) — the binary signal on cycle continuation. Any 15%+ guide-down cascades within days.
7. Mouser / Digi-Key distributor lead-time data on auto-grade MLCC — the best industry-wide S/D-balance read; extended through Q1 2026, remains elevated.
8. Kingboard high-spec CCL pricing actions — four announced hikes through April 2026; watch for a fifth, or Resonac / Mitsubishi Gas Chemical extending their 30%+ bonding-sheet / copper-foil resin-sheet hikes, as confirmation of durable upstream margin support.
9. China commodity / mid-cap encroachment — Fenghua, Sunlord, Sanhuan closing the gap on 0201 high-cap (the 5-year question) is the structural tail risk to Tier-2 Taiwanese pricing. Not yet credibly impacting Tier-1.
10. Rare-earth dopant export risk — a Chinese heavy/medium rare-earth restriction (Dy, Y, Ho dopants) would complicate non-Chinese MLCC production within 6-12 months. Tail risk; not materialized this cycle.

Cycle-off mirror set: book-to-bill <0.9 sustained, hyperscaler capex guide-downs, AEC-Q200 lead-time normalization, broad-based price-hold-or-cut from Murata.

Specific earnings / event dates to watch

• Prosperity Dielectrics (6173.TWO) Q1 2026 earnings — 11 June 2026. Asymmetric to the downside given 53x trailing P/E; the print that must confirm the powder business is operating as the bull thesis requires before any PDC entry.
• Murata FY27 Q1 earnings — expected late July 2026. First print showing whether the Apr-1 2026 price hikes flow through to reported financials at the assumed 1-2 quarter lag. Also the read on whether the 23 May complex-wide limit-up was the next leg or the top.
• Samsung E-M design-win announcements on CL31X227MRKNNW# (220µF, 1206) and CL32X337MSVN4S# (330µF, 1210) — direct competition for Murata's ultra-high-CV equivalents.
• Banyan Lane next modeled hike round — Sept 2026 to Feb 2027, magnitude 20-35% on AI parts; the 2018 cycle ran 4 discrete rounds with ASP roughly doubling trough-to-peak.
• SEMCO Philippines Plant 3 — ships July 2027+; Murata Izumo shipping since April 2026; SEMCO's larger Korean expansion not until mid-2027. None of these resolve the high-end shortage inside an 18-24 month thesis horizon.

Open / unresolved threads to chase

• The "allocation-premium not price-hike" claim — most likely behind the Collyer Bridge "APAC Wrap: 21 May 2026" paywall. Requires Substack re-authentication (visit substack.com/settings + collyerbridge.com in Chrome) to re-read the gated "feedback from chat" section. Per the 28 May sweep, the literal Kemet / tantalum / 40% / spillover quote was not in the ~10-day chat scrollback and may predate 18 May.
• Murata's exact AI-server-specific MLCC share — Banyan Lane uses 60%; industry sources range 50-70%; a 50% read weakens the thesis materially.
• Kingboard's high-spec AI-server CCL vs mid-spec mix — disclosure opaque; only 2 English-language analysts.
• Holy Stone manufacture-vs-distribution revenue split — undisclosed; single largest information gap in the listed peer set.
• Jukan's forthcoming SEMCO / silicon-cap X memo — anticipated next datapoint on the silicon-cap turn; set a reminder.
• CTR Holdings — flagged by Collyer Bridge 22 May for follow-up; no vault coverage yet.

Sources to track on an ongoing basis

Murata IR (IR Day annually December; quarterly late Apr / Jul / Oct / Jan — the single most important data source); Samsung Electro-Mechanics IR (segment guidance, AI-server share cross-check); Taiyo Yuden IR (monthly disclosures, highest-beta cycle read); TDK IR; Yageo IR + monthly Taiwan revenue disclosure; Digitimes (broke the Feb 2026 capacity-doubling and Apr-1 hike stories); Mouser / Digi-Key lead-time data; NVIDIA / AMD earnings calls (accelerator content signals); hyperscaler capex guidance. Substack: Collyer Bridge / Illyquid (Asian AI/semis, active chat — see source-illyquid), Banyan Lane Capital (Murata ASP modeling), STF Research (Founding-tier gated), pequityresearch (AP Memory silicon-cap primer).

Sources

The MLCC sector read rests on three source categories: independent Substack/X analysts, sell-side research (Western bulge bracket plus Asian local brokers), and company / industry primary sources. Where an author has a vault page under _sources/, it is linked below. Only one of the MLCC sources currently has a source page — Collyer Bridge / Illyquid; the rest are listed plain and are candidates for source-page creation.

Independent analysts (Substack / X)

• Collyer Bridge — source-illyquid (Substack collyerbridge.com, X @illyquid; ★★ paid, active; same author across both platforms). The single most-cited independent voice in this corpus. Posts driving the sector: "MLCCs ripping" (May 20), "APAC Wrap: 21 May 2026," "APAC Wrap: 24 May 2026," "Discounted exposure to CCL" (May 12), "May Update" (May 8), and "Small Parts, Big Cycle: Passive Components" (May 28). Subscriber chat (chat ID 6169391) is a primary input — paid Founding tier confirmed. Publicly bullish Yageo but privately prefers Kingboard (1888.HK / 0148.HK); newly convinced on silicon capacitors via the AP Memory primer; reluctant to play MLCCs directly.
• STF Research — stfbutnou.substack.com (Pink holds a paid sub since 2026-05-03, but most content is gated behind a Founding Member tier her level does not reach). Relevant gated posts: "AI Supply Chain: Shortage Intensifying Toward 2027" (Apr 7, 2026) and "Maruwa: Conservative Is Your Dip to Buy" (May 24, 2026). One candidate origin for the "Murata hasn't really raised prices — customers pay premiums to secure capacity" claim Pink was chasing.
• Banyan Lane Capital — Substack thesis on Murata. Models 4-7 rounds of MLCC price hikes through FY30, AI MLCC ASP potentially doubling-plus from trough; next round Sept 2026 – Feb 2027 at 20-35% on AI parts; annual supply deficit widening from -20B units (2026) to -160B units (2030E); 50%/60% Murata AI-share scenarios; base-case 22x target ¥7,254, bull case 28x ¥9,235. Part 2 carries the closest public-facing version of the allocation-premium framing ("shortage allocates pain to tier-2/tier-3 customers while tier-1 customers pay through the nose").
• pequityresearch — pequityresearch.substack.com, "AP Memory / Silicon Capacitor" supply-chain primer. The piece that converted Collyer Bridge to the silicon-capacitor thesis.
• Irrational Analysis — Substack post on Musashi Seimitsu (7220), bullish, cited in Collyer's chat.
• Made in Japan (@madeinjapan) — balanced Musashi Seimitsu view; flagged that the HSC ramp "keeps getting delayed."
• Jukan (@jukan / x.com/jukan05) — surfaced Morgan Stanley MLCC content charts (GB300 vs VR200) in the Collyer chat; referenced in bulge-bracket sell-side coverage; forthcoming SEMCO / silicon-cap memo on X.
• Tim Worstall — Seeking Alpha author / rare-metals wholesaler, recommended in the Collyer chat as a Ruthenium / thick-film-resistor demand resource.
• TrendForce — industry data house. Lead-time data (8 weeks → 24-40 weeks); high-CV MLCC ASP +30-40% projection for 2026; smart-glasses 01005 MLCC demand (150-200 cells/pair); AI customers on 2-3 year long-term planning models making the price trend more durable.
• Archetype Capital — archetype-research.com (Substack; small-cap AI enablers, Japan / SEA value chains, undercovered small-cap AI beneficiaries; Pink subscribed). "Sakai Chemical & Nippon Chemical MLCC Capacity Crunch" (2026-06-01) — premium-BaTiO3 capacity-crunch thesis, production-route market-share split (hydrothermal 13.5% / oxalate 21.4% / solid-state 61.8%), prefers Nippon Chemical (4092) over Sakai (4078); author flags it incomplete/DYOR. See 4078/4078, 4092/4092.
• Zephyr (@zephyr_z9 on X, 29 May 2026 — verified at primary source) — MLCC market sizing: total ~$15B; server MLCC $1.3B in 2025 ($600M AI + $700M general); AI-server 80%+ CAGR, general-server 30-40%; smartphone negative 2026-27; high-end lead time >20 weeks; nickel/silver cost push; book-to-bill >1. Zephyr's actual call: the equipment & raw-material suppliers — the upstream (Sakai 4078, Nippon Chemical 4092) — are the biggest beneficiary of the MLCC capex boom and should now OUTPERFORM the MLCC producers themselves (reiterated 31 May & 1 Jun). The data source the Archetype piece builds on.

Collyer chat contributors named (subscriber handles, not publications, but they carry the qualitative thesis): Zrave (the "MLCC TAM is 99.99% outside that BOM … lifetime Rubin MLCC BOM won't cover a single quarter of Yageo revs" anchor quote), Henry and Bob (Yageo-vs-Murata, sided with Murata on ultra-high-CV pricing power), Gman Scotchnova (authored the 7-name comp sheet Collyer reposted), plus Min Htoo, Phoenixheart, P M, Nishant, T-Bone, fubar, Rest.

Sell-side research

Seven sell-side sources were merged into the primer's consensus build (2026-05-24), with the full consensus at ~/claude/output/compare/mlcc-sellside-consensus-2026-05-24.md:

• BofA (Bank of America) — note dated May 19; Murata Buy PO ¥6,200, TDK Buy PO ¥3,500, Taiyo Yuden Underperform PO ¥3,800.
• JPMorgan (JPM) — note dated Apr 3; industry model with MLCC TAM $16.3B (2025) → $25B (2030E) at +11% CAGR, AI server mix rising 1.1% → 4% of total.
• Goldman Sachs (GS) — note dated Apr 16; on Yageo, PT NT$346.
• Morgan Stanley (MS) — NVL72 BOM build: MLCC content per rack $1,530 (GB300 Blackwell) → $4,320 (VR200 Rubin), +182% / +US$4.3K VR200 framing; the framing Collyer's chat (via Zrave) pushed back on.
• Capacitor Dossier (Apr 2026) — global capacitor market $41.23B; strategic segment 35-40% of market value, 60-70% of industry profit; high-CV MLCC (220µF/4V AI GPU rail) +29% from 2024 to Q1 2026.
• Global Passives Basket (May 21) — six-dimension scoring framework (Moat, Crowding, Revision Velocity, etc.); Murata 21/30 Buy, Kingboard 19/30 Watch, SEMCO 13/30 Hard Pass; Taiyo Yuden Revision Velocity 1/5.
• Murata vs Yageo desk note (May 21).
• Dongguan Securities Research Institute (东莞证券研究所) — March 25, 2026 MLCC industry in-depth report, analyst Luo Weibin (罗伟斌, SAC No. S0340521020001), Electronics Industry Overweight (Maintain). The single most data-dense local-broker source: BaTiO3 powder market shares, three-tier competitive landscape, 2024 global MLCC shares, automotive MLCC-by-powertrain (TDK data), cost-structure tables, lead-time tables by manufacturer, China domestic-substitution math (2.56T MLCCs imported in 2025 worth US$6.179B). Covers Fenghua (000636.SZ, Buy and Hold) and Sanhuan (300408.SZ, Buy and Hold). Raw PDF at KB/raw/mlcc-research/dongguan-mlcc-initiation-2026-03-25.pdf.
• Nomura — raised Musashi Seimitsu (7220) target to ¥7,700 on 22.5x fair PE, sum-of-the-parts.
• Citi — SEMCO silicon-capacitor revenue path (W609bn FY27E / W905bn FY28E, W1.6T 2-year supply order).

The Dongguan report itself cites a long downstream-Chinese reference list: China Business Industry Research Institute, Murata China official WeChat, TDK / SEMCO / Yageo / Walsin / Sanhuan / Future Electronics / Micro-Rong official sites, Aibang Semiconductor Network, New Materials Insights, DakCap Prospectus, Qianzhan Industry Research Institute, China Capacitor Network, Maeil Business Newspaper (Korea, source for the SEMCO ~40% AI-server-MLCC share claim), TrendForce, China General Administration of Customs, Hongxing Electronics Prospectus.

Company and primary / market-data sources

• Murata IR (corporate.murata.com) — the single most important data source in the sector. IR Day annually (December); quarterly results late Apr/Jul/Oct/Jan. Source for the GB300 ~30,000 MLCC / 440,000-per-rack figures, 15,000-25,000 per-baseboard FY25-FY30 revision, 30% server-capacitor demand CAGR 2025-2030, capacitor backlog +89.5%, ¥150B buyback, ¥80B emergency capex + ¥250B FY2026 plan, President Nakajima's "still very much insufficient" remark, capacity-doubling announcement (Feb 2026), Izumo plant shipping April 2026.
• Samsung Electro-Mechanics IR — quarterly results + segment guidance; AI-server-share cross-check; CL31X227 (220µF) and CL32X337 (330µF) product specs.
• Taiyo Yuden IR — monthly disclosures and quarterly results; highest-beta cycle read. Source for the sell-side roundtable (May 22-23): "No MLCC price hikes this fiscal year, but FY3/28 is a possibility … mix shift is the ASP driver."
• TDK IR — quarterly results; auto-MLCC-by-powertrain data, AI passives cross-check.
• Yageo IR + monthly Taiwan revenue disclosure — tantalum / consolidation pace.
• Digitimes (paywalled, English) — Asian supply-chain reporting; broke the Feb 2026 Murata capacity-doubling and Apr-1 price-hike stories; source for the SEMCO ~40% AI-server MLCC share figure.
• Bloomberg — Murata customer inquiries running ~2x supply capacity (Feb 2026).
• Mouser, Digi-Key — distributor lead-time data, the best industry-wide leading indicator.
• NVIDIA, AMD earnings calls — accelerator content / roadmap signals.
• Hyperscaler capex guidance — AMZN, MSFT, META, GOOG, ORCL quarterly; the binary cycle-continuation signal.
• Leo In AI — Vera Rubin NVL72 estimate of 1.5-2M MLCCs per rack.

Consolidated vault source files

The sector page consolidates these vault files (all under KB/wiki/):

• mlcc-primer.md — the master industry primer (Register D), with the 2026-05-23 Sakai correction and 2026-05-24 seven-source data-integration block.
• mlcc-technical-sections-2026-05-25.md — standalone technical deep-dives (how an MLCC works, BaTiO3 as moat, the four powder-prep methods, five manufacturing stages, AI server power-delivery problem, 800V EV architecture, AEC-Q200 qualification barrier).
• dongguan-mlcc-initiation-2026-03-25.md — clean English re-extraction of the Dongguan Securities report.
• mlcc-substack-scrape-2026-05-25.md — Substack pricing-dynamics scrape (the three pricing framings; STF Research / Collyer Bridge / Banyan Lane mapping).
• collyer-bridge-chat-2026-05-28.md — Collyer Bridge subscriber-chat sweep (chat 6169391, 21-29 May; eight topic threads).
• briefings/2026-05-23-mlcc-sector.md — "Top 2 With Crowding in Mind" deep-dive briefing (Murata + Kingboard picks).
• briefings/2026-05-25-mlcc-peer-swarm.md — 10-ticker ranked-verdict briefing.
• briefings/2026-05-28-non-mlcc-capacitor-primer.md — adjacent silicon-capacitor and supercapacitor primer.
• _sources/source-illyquid.md — the Collyer Bridge / Illyquid source page.

Supporting outputs outside the vault: the seven peer wiki pages (KB/wiki/{6981, 2327, 6976, 6762, 3026, 2492, 6173}/), the swarm synthesis (~/claude/output/compare/mlcc-peer-swarm-2026-05-20.md), the sell-side consensus (~/claude/output/compare/mlcc-sellside-consensus-2026-05-24.md), the Sakai profile (~/claude/output/profile/4078-t-profile.md), and the adjacent-sector primers (~/claude/output/sectors/{ai-server-pcb, packaging-glass-substrate, cu-wiring-resin}-primer.md). Raw sell-side material is archived at KB/raw/mlcc-research/.

---

Consolidation queue (merged 2026-05-30 — section-scoped rebuild)

Industry-wide content folded in from these source files. They stay live pending Pink's archive confirm.

• [ ] mlcc-primer.md
• [ ] mlcc-substack-scrape-2026-05-25.md
• [ ] mlcc-technical-sections-2026-05-25.md
• [ ] dongguan-mlcc-initiation-2026-03-25.md
• [ ] collyer-bridge-chat-2026-05-28.md
• [ ] briefings/2026-05-23-mlcc-sector.md
• [ ] briefings/2026-05-25-mlcc-peer-swarm.md
• [ ] briefings/2026-05-28-non-mlcc-capacitor-primer.md
• [ ] _sources/source-illyquid.md

---

MLCC sell-side consensus (folded 2026-06-01, from mlcc-sellside-consensus-2026-05-24)

MLCC Sell-Side Consensus — 2026-05-24

Consolidated ratings and estimates from 7 uploaded research reports, cross-referenced against our swarm verdicts (mlcc-peer-swarm-2026-05-20, mlcc-sector-2026-05-23).

Consensus Table

Ticker	Company	Source	Date	Rating	PT	Fwd P/E	Key estimate
6981.T	Murata	BofA	May 19	Buy (B-1-7)	¥6,200	36.7x (FY3/28E)	FY3/27E: Sales ¥2,007B, OP ¥416B, NP ¥329.9B, EPS ¥168.1; FY3/28E: OP ¥511B, EPS ¥204.1; FY3/29E: OP ¥608.7B, EPS ¥242.6
6981.T	Murata	JPM	Apr 3	Overweight	¥3,500	29.8x (FY25E)	Prefer Murata + Taiyo Yuden in short term on tight MLCC S/D
6981.T	Murata	Global Basket	May 21	Buy (21/30)	—	~40x NTM	Moat 5, Crowding 3, Valuation 3, Inflection 4, Paid Now 3, Revisions 3
6762.T	TDK	BofA	May 19	Buy (C-1-7)	¥3,500	—	FY3/27E: Sales ¥2,635.5B, OP ¥301.7B, EPS ¥121.2; FY3/28E: Sales ¥2,821.2B, OP ¥339.8B, EPS ¥138.2; FY3/29E: Sales ¥2,994.6B, OP ¥382.2B, EPS ¥155.4
6762.T	TDK	JPM	Apr 3	Overweight	¥2,700	—	Prefer TDK medium/long term (battery + HDD + sensors)
6976.T	Taiyo Yuden	BofA	May 19	Underperform (C-1-7)	¥3,800	45.3x (FY3/28E)	FY3/27E: Sales ¥384B-387.4B, OP ¥30B-34.9B, EPS ¥143.9-158.7; FY3/28E: OP ¥46.1B, EPS ¥215.0; FY3/29E: OP ¥57.5B, EPS ¥272.0
6976.T	Taiyo Yuden	JPM	Apr 3	Overweight	¥3,600	—	Upgraded from Neutral; prefer alongside Murata on tight S/D
6976.T	Taiyo Yuden	Global Basket	May 21	Hard Pass (14/30)	—	~87x NTM	Revision Velocity 1 (worst) — estimates cut while stock rallied 77%
6976.T	Taiyo Yuden	Sell-side mtg	—	—	—	—	AI server MLCC sales growing 30-35% CAGR FY27-FY30; Murata/SEMCO are real competitors, not Chinese
2327.TW	Yageo	GS	Apr 16	Buy	NT$346	16.8x (FY26E)	FY26E EPS NT$19.12, FY27E NT$25.27, FY28E NT$27.87. 15x PB/ROE valuation. Stock at NT$572 = 65% above GS PT
2327.TW	Yageo	Global Basket	May 21	Pass (17/30)	—	27.2x	Crowding 2, Valuation 2, Revision Velocity 2 — stock +178% in 12mo but revisions only 3-7%
2327.TW	Yageo	Murata vs Yageo	May 21	—	—	~26x	"AI story is more accurately a tantalum story than an MLCC story"; AI rack MLCC = 3-6% of Yageo annual revenue
2492.TW	Walsin	Global Basket	May 21	Pass (16/30)	—	n/a	Crowding 1 (worst — limit-up tape), Valuation 2
009150.KS	SEMCO	JPM	Apr 3	Overweight	KRW 325,000	—	All 4 major Japanese MLCC names OW; prefer JP over KR on undemanding valuations
009150.KS	SEMCO	Global Basket	May 21	Hard Pass (13/30)	—	~82x NTM	Crowding 1, Valuation 1 — 9x in 12 months; analyst PT dispersion 4x (Mirae 250k vs KB 1.05M)
3026.TW	Holy Stone	Global Basket	May 21	Hard Pass (12/30)	—	~65x TTM	Moat 2, Crowding 1, Valuation 1
0148.HK	Kingboard	Global Basket	May 21	Watch (19/30)	—	~11x NTM	CCL + PCB upstream; Inflection 4; Cheung family HoldCo discount
000636.SZ	Fenghua	Dongguan	Mar 25	Buy and Hold	—	47.86x (2026E)	EPS 2025E 0.31, 2026E 0.42
300408.SZ	Sanhuan	Dongguan	Mar 25	Buy and Hold	—	31.72x (2026E)	EPS 2025E 1.40, 2026E 1.71

Where Sell-Side Disagrees with Our Swarm

1. Taiyo Yuden: BofA Underperform vs JPM Overweight. Sell-side is split. Our swarm said WATCH at ¥8,231 (agrees more with BofA's view that current price embeds excessive expectations). BofA PT ¥3,800 is below even our swarm's entry-zone floor of ¥5,000-5,500. JPM just upgraded and has a more optimistic cycle thesis.

2. Yageo: GS Buy at NT$346 vs our swarm WATCH. GS PT is NT$346, stock trading at NT$572 (+65% above). GS uses a 15x PB/ROE upcycle multiple — if you trust that methodology, the stock is overvalued by GS's own framework. Our swarm entry zone of NT$480-520 is above GS's PT but below current price. Alignment: both agree it shouldn't be chased at current levels.

3. SEMCO: JPM Overweight vs our swarm PASS / Global Basket Hard Pass. JPM's PT of KRW 325,000 (set Jan 2026) is 69% below current KRW 1,061,000. JPM prefers Japanese names over Korean. Our swarm PASS at current aligns with JPM's implicit valuation read but not with their OW rating (which was set at much lower prices).

4. Murata: All sources agree Buy/OW, but PT dispersion is wide. BofA ¥6,200, JPM ¥3,500 (very stale, set Dec 2025), Global Basket 21/30 Buy. Our swarm BUY scale-in at current aligns. The JPM PT is obsolete and should be ignored for entry math.

5. Kingboard 0148.HK confirmed as Watch by Global Basket. Matches the SEMCO sub-agent's flag. Not researched in our swarm but warranted as a separate /profile if Pink wants the upstream PCB/CCL expression.

6. Chinese domestic names (Fenghua 000636.SZ, Sanhuan 300408.SZ). New tickers from the Dongguan report, not in our swarm. Both "Buy and Hold" but at high P/E multiples (48x, 32x on 2026E). These are the Chinese domestic substitution plays, not the same quality tier as the Tier-1/2 names.

Key Data Points

Morgan Stanley NVL72 BOM (Blackwell to Rubin)

Component	GB300 (Blackwell)	VR200 (Rubin)	Change
MLCC content per rack	$1,530	$4,320	+182%
Memory	$373,939	$2,001,600	+435%
PCB	$35,100	$116,730	+233%
Total rack	$3,994,551	$7,803,148	+95%

Murata FY2026 Actuals and FY2027 Guidance

• FY2026 Rev ¥1,830,856M (+5.0%), OP ¥281,835M (+0.8%), NP ¥233,920M (+0.0%), EPS ¥127.66
• FY2027 guide: Rev ¥1,960,000M (+7.1%), OP ¥380,000M (+34.8%), NP ¥293,000M (+25.3%), EPS ¥160.96
• Buyback: ¥150B (75M shares, ~4% of float)
• Capacitor backlog surged +89.5% to ¥269.2B; computers/servers +28.4%; total backlog +55.2%
• Segment: Components (MLCC, inductors) rev +12.3%, ROIC 22.4%; Devices & Modules (SAW, RF) rev -5.9%, ROIC -3.5% (destroying value); ¥43.8B SAW goodwill impairment in FY2026