Semiconductors (cross-cutting / geopolitics)

Overview / thesis

The semiconductor supply chain is fracturing along geopolitical lines, and that fracture is the trade. US-China decoupling, friend-shoring, and the AI capex supercycle have collided at the same moment — turning what used to be a globally optimized, cost-minimizing chip supply chain into a contested map of chokepoints, each one a potential investment. The cross-cutting thesis across every source in this sector is the same: trace the demand backwards until you hit a node that cannot be bypassed and is not yet priced, and you find the asymmetric bet. Geopolitics decides which nodes are durable; AI decides how fast they grow.

The structural opportunity

Three forces are compounding on top of each other, and the sources treat them as one story rather than three:

1. US-China tech decoupling. Export controls on advanced chips (the January 2025 AI diffusion rules restricting advanced computing ICs, AI model weights, and license exceptions) are pushing the supply chain to bifurcate. China retaliates on the mineral side — germanium and gallium restrictions (2023), antimony (2024) — which directly targets semiconductor materials and the defense supply chain. The net effect is not a clean cut but a permanent friction tax that makes Western-sourced, allied-sourced, or hard-to-replicate supply structurally more valuable.

2. Friend-shoring / reshoring. The CHIPS Act is pumping $52B into domestic US fabs. But the recurring, hard-edged conclusion across the sources (Barclays manufacturing-footprint analysis) is that reshoring only works for leading-edge (<7nm) — mature nodes (28nm+) will stay concentrated in Asia because the economics don't support onshoring commodity production. This is the central debate: the CHIPS Act is subsidizing what is economically marginal, and the open question is whether it sticks when subsidies expire. Intel 18A is named repeatedly as the acid test — if 18A and Intel Foundry Services fail to win external customers, the domestic-leading-edge reshoring thesis collapses.

3. The AI capex supercycle. Underneath the geopolitics is the demand engine. Hyperscaler AI buildout is driving demand backwards through transceivers, lasers, epiwafers, substrates, metrology, test, and power. The Intel 18A cycle is explicitly noted as happening simultaneously with the AI infrastructure buildout — supply chain names supplying both advanced packaging AND hyperscaler AI chips get two demand curves instead of one ("dual tailwind").

The 'so what' — where the money is

The sources converge on a clear hierarchy of where the opportunity actually sits, and it is deliberately not the obvious names:

• Go upstream until you hit a bottleneck the market hasn't priced. Serenity's (@aleabitoreddit) core construction — and the dominant method across the screenshot archives — is to trace hyperscaler capex backwards through the supply chain until landing on a $100M–$700M-MC upstream node with exclusive exposure. The mental test: can this node be bypassed? If not, it's a bottleneck, not a supplier. "The more well known the name ($LITE, $COHR) the more crowded/priced-in. Go upstream for hidden gems." Example chain: hyperscaler transceiver demand → AAOI → MBE/MOCVD reactors (VECO, Aixtron) → epiwafer foundries (IQE ~$175M, Landmark ~$3.6B) → substrate (AXTI).

• Friend-shoring's sweet spot is Japan. Japan is the most-developed friend-shoring story in the vault — not competing with TSMC on logic, but dominating niche mid-stream component markets that are hard to replicate and increasingly strategically valuable. The pitch is "the friend-shoring thesis in stock form": allied with the US, technologically advanced, but not directly in the China export-control crosshairs the way Taiwan is. The component layers — optics, test/measurement, connectors, metrology, fiber, substrates — are exactly what every GPU cluster needs. See japan-semi and jp-semi-supplychain-comparison for the 6-company map.

• Semicap and subsystems are the broad reshoring beneficiary. From the buy list: AMAT, LRCX, CAMT, ONTO, TOELY positioned for the "strongest and most meaningful year in WFE history"; subsystems MKSI, ICHR, UCTT as the levered play on fab buildout.

• Mature-node / lagging-edge. A persistent secondary thesis (Fabricated Knowledge, "The Long Tail of Smaller Fabs"): the fastest capex growth is not TSMC's ~40% YoY, but the long tail of IDMs and smaller fabs — "lagging edge semiconductor is finally getting the investment it so desperately needs."

Sizing the opportunity

The vault contains a historical-analogy frame for the AI capex wave rather than a clean top-down TAM build. The recurring numbers (from the AI capex notes):

• 1999 fiber spend ~$120–140B → equivalent of ~$213B in 2024 dollars by CPI.
• 2024 AI capex spend ~$100B against 58M people in US.
• Telecom-era benchmark: ~$600/person during 1992–2000, ~$200 cumulative for the whole US → equivalent of ~$1.5T of AI spending given ~300M US population today. The implication is that AI capex still has substantial runway versus the dot-com fiber buildout on a per-capita-normalized basis.
• NVDA performance/TCO trajectory: A100 → H100 → B100 (Blackwell) → "X100", ~3x performance per generation at ~50%+ price increase = ~2.5x TCO improvement. "H100 in 2030 is gonna be outdated" — the relentless obsolescence cycle is itself a demand driver for the supply chain.

For specific sub-markets, the friend-shoring component names carry their own pockets: Nitto Boseki holds ~90%+ share of specialty glass for GPU substrates and a 100% T-Glass monopoly; Rigaku is #1 globally in XRD (~29% share); JEM is a top-5 probe-card player (~8–10% share) levered to HBM. These are micro-TAMs, not a sector top-down. The Japan ecosystem trades at 10–20x earnings on the TSE versus 25–40x for Western semi peers — the valuation gap is itself the opportunity (or a value trap from low growth/liquidity; the sources flag both readings).

The growth rates that recur: Santec ~20% forward revenue growth (39% 3yr CAGR), JEM +17.9% forward (HBM probe-card driven), Furukawa optical fiber pricing in transformative margin expansion (+380% 1yr stock move on +5.6% forward revenue), CPO requiring 4x the SiPho wafer content of pluggables. The probe-card cycle shows AI-chip testing lead times at 6 months to 16–24 months with 2-year visibility on some orders — a direct read on how tight the bottleneck has become.

The demand drivers (recurring across sources)

• AI datacenter buildout — GPU clusters need optical connectors, fiber, transceivers, lasers, test equipment, metrology, and power electronics. Every node in the AI BOM is a demand vector. CPO (co-packaged optics) is arriving earlier than expected and is a named near-term catalyst.
• HBM ramp — HBM3e/HBM4 mandates wafer-level burn-in with no substitute; drives probe-card, burn-in test, and metrology demand. "All revenue related to HBM continues to shine."
• Advanced packaging — EMIB, Foveros, chiplets, glass substrates. New package architectures require 100% burn-in. The glass-substrate shift at 1.4nm / 1.6T speeds is a flagged mid-2026 inflection.
• 5G/6G rollout — test & measurement demand (Anritsu's core), defense electronics overlap.
• Reshoring capex — WFE buildout from CHIPS Act, EU Chips Act, and friend-shoring fab expansion.
• "Made in America" supply premium — geopolitically resilient supply commands a pricing premium (AAOI transceivers vs FN/Innolight/Eptolink; analogs INTC-to-TSM, MU-to-Hynix). Matters more as export controls tighten.

The key debates

1. Can reshoring work at scale? Yes for leading-edge, no for mature nodes (Barclays). Will CHIPS Act subsidies for economically marginal capacity survive their own expiry?

2. Is Intel 18A the linchpin or the failure point? It is the CHIPS Act's biggest bet — domestic leading-edge foundry. If 18A process technology doesn't work and IFS can't win external customers, the reshoring thesis and a whole EMIB/advanced-packaging supply chain (see intel-supply-chain) deflate. Pink's own INTC trade log is a cautionary note on confusing hype with the long execution game.

3. Is the friend-shoring premium permanent or transient? If US-China tensions de-escalate, does the strategic-value argument for Japan's mid-stream collapse? Each Japanese name has different China revenue exposure — decoupling can hurt (lost market) or help (strategic premium).

4. Do export controls actually work? The split view (Gregory Allen): controls slow China but don't stop it — Huawei's Kirin 9000S proved SMIC can produce at 7nm, DeepSeek showed China can do more with less. Controls are buying time, not solving the problem, and are simultaneously driving Chinese self-sufficiency (Huawei Ascend, SMIC N+2). The unresolved question is how much demand is truly inaccessible vs merely displaced, and who gains the displaced share (Taiwan, Korea, US).

5. How far does escalation go? What triggers serious Chinese retaliation beyond minerals — rare-earth magnet export restrictions (critical for defense, EVs, wind)? And what does a Section 232 semiconductor tariff regime do to the supply chain (Oliver Wyman projects significant Asia supply-chain diversion)?

6. Is the move fundamental or a "paper game"? A recurring discipline: when a move is driven by exchange rules, margin hikes, forced liquidation, or index rebalance rather than fundamentals, fade the narrative. And when news implies a timeline that violates engineering reality (silicon validation cadence of 30–38 months), the market is wrong.

7. EPS-driven vs multiple-driven. The Caughran (FundamentEdge) overlay that frames how to read the whole sector: sustained 10%+ revenue growth is chronically underpriced; valuation-shorting into positive revisions is a "fatal, rookie mistake"; decompose every move into EPS roll-forward vs revision vs P/E expansion. Most of the 2021–22 semi winners were multiple expansion, not earnings — the current concern is how much of the 2025–26 supply-chain run is the same.

The structural conclusion

The single sentence the sources keep returning to: the value has migrated from the famous names to the upstream chokepoints, and geopolitics is the thing that makes a chokepoint durable. Decoupling and friend-shoring convert hard-to-replicate, allied-sourced, or sole-source supply into a strategic-premium asset; the AI supercycle supplies the volume; and the disciplined trade is to find the sub-$1B-MC node that everyone downstream depends on before the market maps the dependency. Company-specific theses live on the ticker pages — japan-semi for the friend-shoring component ecosystem, intel-supply-chain for the 18A/EMIB cluster, supply-chain-security for the geopolitical map, and china-semi for the indigenization counter-story.

How it works

The semiconductor industry is not one business — it is a stack of physically distinct businesses sitting on top of each other, each with its own physics, its own cost structure, and its own moat. Understanding "how it works" means understanding where value concentrates along that stack, why certain nodes are bypassable and others are chokepoints, and why the economics differ by an order of magnitude between, say, a fabless designer and a glass-cloth maker. The cross-cutting/geopolitics lens cares about the second question above all: a bottleneck is anything in the chain that cannot be bypassed, and that is where pricing power, strategic leverage, and export-control sensitivity all collect.

The value chain, layer by layer

The canonical mental model — used by Serenity (@aleabitoreddit) and confirmed across the vault — is to start from end demand (hyperscaler AI capex, 5G/6G rollout, automotive electrification, advanced packaging) and walk upstream until you hit a small-cap node that cannot be designed around. Serenity's worked example for optical interconnect: hyperscaler transceiver demand → transceiver maker (AAOI) → MBE/MOCVD reactors (Veeco, Aixtron) → epiwafer foundries (IQE ~$175M MC, Landmark ~$3.6B) → substrate (AXTI). The further upstream and the more exclusive the exposure, the less the name is priced in. Her rule: "The more well known the name (LITE, COHR) the more crowded/priced-in. Go upstream for hidden gems." The test that separates a bottleneck from an ordinary supplier — can this node be bypassed? If not, it is a bottleneck, and a bottleneck is where the cross-cutting thesis lives.

The full stack, simplified, runs: design tools (EDA) → chip design (fabless/IDM) → wafer substrate → wafer fab (lithography, deposition, etch, implant, metrology) → test → advanced packaging (substrates, bridges, hybrid bonding) → board-level assembly (PCB, connectors, passives) → system integration (servers, optics, power). Geopolitics bites at every layer because the equipment, materials, and software inputs are concentrated in a handful of countries and companies.

Logic vs. memory vs. analog — three different physics, three different economics

Logic is the computation layer — CPUs, GPUs, accelerators. This is where Moore's Law scaling (transistor shrinks) drives value, where leading-edge nodes (sub-7nm) command the highest capital intensity, and where TSMC, Samsung, and Intel compete. Process scaling now leans on architectural tricks the FAQ flags as research priorities: BSPDN (backside power delivery network) is called out as a clear performance advantage, and crucially implementing BSPDN in manufacturing requires more etch and deposition tools — a direct read-through to the WFE (wafer fab equipment) makers (Tokyo Electron, ASML, AMAT, KLAC, LRCX). IMEC (Belgium, a few hours' drive from ASML) is the research house tracking transistor scaling.

Memory is the storage layer — DRAM and NAND — where the unit of demand is bit growth, not chip count, and where the key cross-cutting term to track is "wafer starts per month" and "bit growth for DRAM/NAND." Memory's relevance to AI runs through HBM (high-bandwidth memory), which sits at the intersection of three technologies the FAQ flags as the future of AI/cloud infrastructure: HBM, hybrid bonding, and DRAM scaling. HBM is stacked DRAM, and the packaging method (hybrid bonding vs. micro-bump) determines stack height, thermal performance, and yield.

Analog / power is the boring, durable layer — MOSFETs, IGBTs, RF front-end, automotive sensors — that does not scale with Moore's Law and lives on mature nodes. The FAQ flags the structural problem here: Chinese foundries and IDMs are expanding MOSFET/IGBT capacity rapidly while BYD raises its internal sourcing ratio in NEVs, threatening a glut. The "lagging edge war" over the next five years is the analog/power equivalent of the leading-edge race.

Leading edge vs. lagging edge — the reshoring economics fork

The single most important economic split for the geopolitics thesis comes from Barclays' manufacturing-footprint analysis: reshoring works for leading-edge (<7nm), but mature nodes (28nm+) will stay concentrated in Asia — the economics don't support onshoring commodity production. Leading-edge fabs are strategically scarce and subsidy-justified; mature-node fabs are commodity capacity where cost discipline wins and CHIPS Act dollars are subsidizing what is economically marginal. The open question the vault repeatedly returns to: will reshored mature capacity survive when the subsidies expire?

This fork also drives the "long tail" capex thesis from Doug O'Laughlin / Fabricated Knowledge: the fastest capex growth historically came not from TSMC's ~40% YoY leading-edge spend, but from the long tail of IDMs and smaller fabs — UMC and STM both guided ~100% YoY capex increases, NXP ~70%, with Powerchip and TXN growing faster than 40%. "Lagging edge semiconductor is finally getting the investment it so desperately needs." The cleanest stock expression was ACLS (second-place ion-implant maker behind AMAT, ~80% lagging-edge revenue) — though by Nov 2024 the same analyst reversed, calling ACLS "possibly the best funding short in semi-cap forever" because its Chinese revenue would be replaced and its backlog overstated. The lagging-edge thesis is real but its incumbents are exposed to Chinese domestic substitution.

Wafer fab equipment (WFE) — the economic engine of the fab

Inside the fab, value concentrates in five tool categories, each dominated by one or two players: lithography (ASML — the most geopolitically important equipment maker, uncovered in the vault), deposition and etch (AMAT, LRCX, Tokyo Electron), and metrology/inspection (KLAC, ONTO, Rigaku). The terms to internalize: WFE intensity (equipment spend per unit of capacity), semicap intensity (FB), and semicap yield by player. More process steps — driven by BSPDN, by 3D scaling, by advanced packaging — mean more etch, deposition, and metrology passes per wafer, which is why metrology is a structurally growing share of the tool budget. As the Japan semi ecosystem note puts it: "Advanced node manufacturing requires more metrology steps. As chip complexity increases, X-ray inspection becomes more critical."

Below the tools sit the subsystems — the levered play on the fab buildout. The watchlist names MKSI, ICHR, UCTT as the subsystem layer; the Fabricated Knowledge read is that "all subsystems (except ICHR which sucks) look to be in a much better spot" with HBM-related revenue shining and orders pulling into the back half. The unit economics that make subsystems attractive: UCTT is modeled at 35–40 cents of incremental gross margin per dollar of WFE revenue — high operating leverage on the upswing, which is also why such names run hot (UCTT's 280% YTD was ~65% multiple-driven, not earnings-driven).

Metrology and inspection — why complexity is the moat

Metrology is the quality-control layer, and it is a textbook example of physics creating a moat. Rigaku holds ~29% of the global XRD (X-ray diffraction) market, #1 worldwide, with a moat built on IP/patents, installed base, and switching costs. The thesis extension from Collyer Bridge: Rigaku's true semiconductor exposure is 60%+ of profits, not the reported ~30% of revenue ("Lab to Fab" — a semi play hiding in plain sight). The specific technical catalyst is a hybrid metrology engine, T-SAXS, combining X-ray + optical, targeting mass production, a "$100M new market" with >50% share, co-developed with optical partners (ONTO) and a JEP with major memory customers — applicability debated between 3D DRAM and 3D NAND, with the lean toward 3D DRAM given Onto's involvement. A 2nm or SiPho ramp would see Samsung place major Rigaku kit orders. The general principle: as nodes shrink and stack into 3D, more inspection steps are required, and X-ray inspection (non-destructive, sub-surface) becomes mandatory where optical can't see.

Test and burn-in — the catalyst nobody models

Test is the layer that proves a chip works before it ships, and two structural shifts are driving it. First, new package architectures require 100% burn-in (every unit, not a sample). Second, HBM3e/HBM4 mandates wafer-level burn-in with no substitute — you cannot package an HBM stack and then discover a bad die, so the die must be stressed at wafer level. This is the engine behind the Intel 18A / EMIB burn-in sub-theme (AEM Holdings AWX with Intel-qualified HDMT system-level burn-in; Sunright S71 burn-in services + KESM test). The consumable angle matters: the WaferPak (the interface hardware) is a recurring-revenue consumable, not a one-time tool sale — see burn-in-test-primer.

Probe cards are the test interface that contacts the wafer. Japan Electronic Materials / JEM is a probe-card specialist (~95% of revenue, ~8–10% global share, #4–5 player behind FormFactor and Technoprobe). The economics are highly cyclical (revenue swung 23.6B → 17.5B → back over three years) but capital-light with high ROIC (23%). The current up-cycle is HBM-driven: the screenshot archive notes AI chip testing probe-card lead times at 16–24 months (MPI Corp / 6223 TWSE quotes 6 months on some lines, 2-year visibility on others), with Union Tool (6278) and others adding ~50% capacity at Zhubei H2 2026. Long lead times are the tell of a genuine bottleneck — you cannot will probe-card capacity into existence on demand.

Advanced packaging — where Moore's Law goes when shrinks get hard

As transistor scaling slows, the industry buys performance by stitching multiple dies together. This is the advanced-packaging layer, and it is the fastest-growing part of the cross-cutting thesis because it benefits from both the Intel 18A ramp and the AI buildout simultaneously.

The mechanisms:

• EMIB (Embedded Multi-die Interconnect Bridge) — Intel's method of embedding a small silicon bridge in the substrate to connect chiplets, the alternative to a full silicon interposer. Drives demand for IC substrates, metrology, semicap subsystems, and yield-analytics software.
• Foveros — Intel's 3D die-stacking method (AMKR / Amkor is an OSAT partner).
• Hybrid bonding — copper-to-copper direct bonding with no solder bump, enabling much finer pitch and taller stacks; the key to HBM scaling. BESI is the photonics-table analog to AMAT here (hybrid-bonding tool leader).
• Chiplets on substrate — the substrate (ABF, Ajinomoto Build-up Film) carries the chiplets, which is why ABF substrate is in structural shortage, widening to 42% by 2028 (AT&S, with double-sided ABF capability at its Kulim campus — anchored by AMD, not Intel, despite the Intel framing).

The substrate sub-layer has its own chokepoint chemistry. From the screenshot archive: MEC (4971 TYO) is the sole provider of molecular-bonding "super-roughening" chemistry for FC-BGA substrates — the chemistry that prevents the organic substrate from cracking under GPU thermal loads. Competitors (MKS, JCU) "lack laser focus." The substrate makers (SEMCO, Ibiden, Shinko) depend on it. The catalyst flagged: the shift to glass substrates at 1.4nm / 1.6T speeds makes traditional roughening obsolete mid-2026, with AI-grade substrate demand running +50% against a ~50% supply gap.

The materials chokepoint — glass cloth, the strongest moat in the stack

The most extreme example of physics-as-moat in the vault is Nitto Boseki (3110) in specialty glass cloth for GPU substrates. It holds ~90%+ share in specialty glass for GPU substrates and a 100% monopoly in T-Glass, with pricing power expressed concretely: NE-Glass sells at a 6x premium to commodity E-Glass. The moat is the multi-year qualification process and a massive technology gap in NE-Glass and NER-Glass that Chinese glass-fiber makers cannot yet close. Customers include NVIDIA, AMD, Apple, Google, Amazon; order books were full through 2026. The risk is the mirror image of the moat — single-product dependence and a valuation (37.8x EV/EBITDA, ~66x normalized P/E) priced for monopoly perpetuity.

Optics and interconnect — the AI-datacenter bottleneck cluster

Every GPU cluster needs to move data between chips, between racks, and across the datacenter, and at AI scale copper runs out of reach — the signal goes optical. This is the richest small-cap hunting ground in the vault. The key physics and components:

• Optical transceivers convert electrical signals to light. Demand is driven by GPU cluster scale-out; the geopolitics angle is a "Made in America" supply premium — AAOI transceivers command a premium versus FN (Thailand), Innolight (China), Eptolink (China) for geopolitical resilience, the optical analog of INTC-to-TSM.
• CPO (Co-Packaged Optics) moves the optical engine onto the same package as the switch/GPU, eliminating the pluggable transceiver. The decisive economic fact: CPO requires ~4x the SiPho (silicon-photonics) wafer content versus pluggables (SOITEC note), which is why the substrate maker wins disproportionately as CPO ramps. Serenity flags CPO "happening earlier than expected."
• EML / CW lasers are the light sources. NVDA's 2024 EML preallocation drove LITE +955%; the same playbook repeated in March 2026 with NVDA committing ~$2B each into COHR / LITE / MRVL for CW/EML lasers and CPO. The lesson (Serenity's "playbook recognition"): when a big customer does the same capacity-lock move twice in the same chain, the downstream effect repeats.
• Silicon photonics (SiPho) integrates optical functions onto silicon. Soitec holds 95%+ share of Photonics-SOI, the engineered substrate that most of the SiPho fab layer (TSEM, COHR, LITE, AVGO) sits on. Tower Semiconductor and GlobalFoundries position as Western SiPho foundries (friend-shoring plays). PAM4 signaling is flagged in the FAQ as "good and cheap" — the modulation scheme that doubles bits-per-symbol cheaply, central to transceiver economics.

The InP (indium phosphide) substrate supply chain is a clean bottleneck map (Serenity): phosphate mine (DGC) → refining (NCI) → melt (Sumitomo / JX / AXT) → fab (COHR / LITE) → packaging (Innolight / Fabrinet) → end user (NVDA / GOOGL). "West depends on NCI" — a single refining node the whole Western optical chain runs through.

Below the active optics sit the passive components — fiber and connectors — which is where the Japan ecosystem dominates. The technical specifics:

• APC (Angled Physical Contact) connectors are the industry standard for reducing back-reflection in fiber systems; the moat is precision polishing (JEM, Seikoh Giken 6834).
• Multi-core fiber and ultra-low-loss fiber for long-distance datacenter interconnect — Sumitomo Electric (5802), Furukawa (5801), pioneering pre-form-to-cable vertical integration ("Lightera," "Lightera" brand).
• Tunable lasers and optical T&M to validate every transceiver and fiber link — Santec (6777), the optical test specialist.

The cost structure — why margins fan out 10x across the stack

The single clearest data set in the vault on unit economics is the 6-company Japan supply-chain comparison, which shows how radically margins diverge by position in the chain. Reproduced:

Company	Ticker	Position	Gross Margin	Op Margin	Net Margin	ROIC
Santec	6777	Optical components/instruments	58.1%	31.6%	22.1%	~20%
Rigaku	268A/268A	268A	X-ray metrology	~56%	17.7%	12.1%
Anritsu	6754/6754	6754	RF/microwave T&M	50.6%	10.7%	8.2%
JEM	6855/6855	6855	Probe cards	41.5%	22.8%	16.4%
Nitto Boseki	3110	Specialty glass cloth	37.0%	15.1%	11.8%	8.4%
Furukawa	5801	Optical fiber/cable (commodity)	16.8%	3.9%	2.8%	5.7%

The spread tells the whole economic story. Differentiated, IP-dense, hard-to-qualify nodes earn 30%+ operating margins and 20%+ ROIC (Santec, JEM). Commodity volume nodes earn single-digit margins and barely cover cost of capital (Furukawa at 3.9% operating, 5.7% ROIC — "commodity-like returns," the market paying for fiber volume growth, not margins). Capital-light businesses (JEM, ~23% ROIC, net cash) crush capital-heavy ones. The same position-determines-margin logic generalizes across the whole industry: EDA and fabless design earn the highest margins (near-zero marginal cost on IP), leading-edge foundry earns high but capital-intensive returns, OSAT and commodity substrates earn thin margins, and materials monopolies (Nitto Boseki) earn whatever their pricing power allows.

The cyclicality and capital-intensity reality

Two structural features shape every semiconductor sub-business. First, extreme cyclicality — JEM's revenue collapsed 26% peak-to-trough; the automotive/analog names (Renesas, ON, STM, IFX) cycle on inventory turns, with multiples contracting "massively" at the bottom and double-digit revenue/EPS swings on the recovery. The discipline lesson the vault repeats (Caughran's "respect the second derivative"): track the rate of change of growth, not absolute growth — stocks rarely drop hard when fundamentals are accelerating, and bottom calls in semis are made when "no one seems to care."

Second, capital intensity bifurcates the chain. Fabless designers and IP/software (EDA, ARM taking double-digit ASP increases on CSS products) carry almost no fixed-asset burden; leading-edge fabs and the equipment that fills them carry enormous ones. The capex super-cycle thesis — "the strongest and most meaningful year in WFE history" (AMAT, LRCX, CAMT, ONTO, TOELY) — is the mechanism by which AI-datacenter end demand converts into equipment, subsystem, and materials revenue several layers upstream.

The AI-capex demand engine, in historical scale

The cross-cutting demand driver underneath all of this is hyperscaler AI capex, and the sector data note frames its scale against the dot-com fiber build for perspective: 1999 fiber spend was ~$120–140B, equivalent to ~$213B in 2024 dollars by CPI; the 1992-era telecom build ran ~$600/person, ~$200B cumulative for the US, equivalent to ~$1.5T of AI spending scaled to a 300M population. Today's AI capex (~$100B against 58M relevant US workers in the framing) is large but, on this comparison, not yet at the fiber-bubble extreme. The chip-performance treadmill that drives the spend: NVIDIA A100 → H100 → B100 → X100, each generation ~3x performance for ~50%+ more price = ~2.5x improvement in TCO (total cost of ownership) — and "an H100 in 2030 is gonna be outdated." Performance-per-dollar improvement, not raw price, is what keeps the capex cycle compounding, and it is what makes every node in the supply chain a leveraged bet on AI compute scaling.

Why bottlenecks become geopolitical weapons

The economic punchline for the cross-cutting/geopolitics page: a node that cannot be bypassed is simultaneously the highest-margin business and the most potent point of strategic leverage. China's mineral retaliation works precisely because germanium, gallium, and antimony are unbypassable inputs (China controls ~97% gallium, ~82% germanium, 70–90% REE processing) — the same "can't-be-bypassed" property that makes Nitto Boseki's glass cloth a 6x-premium monopoly makes Chinese material control a credible threat. US export controls work the same way in reverse: ASML lithography, advanced EDA, and HBM are chokepoints the US can squeeze. The friend-shoring thesis — Japan's mid-stream components, GlobalFoundries/Tower as Western SiPho foundries — is fundamentally a bet that the market will re-price bottleneck nodes located in allied jurisdictions at a premium, because in a decoupling world the question is no longer just "is this node hard to replicate?" but "is it hard to replicate and on the right side of the line?"

Subsectors

The semiconductor sector as covered in this vault is not one market — it is a stack of distinct sub-areas, each with its own technology, its own set of players, and its own investment angle. They are bound together by two macro forces: the AI infrastructure buildout that drives demand, and the US-China decoupling that fractures supply. What follows enumerates each sub-area found in the sources. Company-specific deep dives live on their ticker pages and are referenced by ticker wikilink rather than re-hosted here.

China semiconductor ecosystem

What it is. China's domestic chip industry and the policy/geopolitical dynamics around it. The vault's dedicated china-semi page is a stub ("To be populated from source files"), and the dedicated export-controls-china-semi page is a framework-only stub marked "Dropbox sync pending" — so most of the substantive China content lives scattered across the supply-chain-security hub, the master-faq, and screenshot archives. The honest state: this is the least-developed subsector page despite being one of the most important themes.

The technology and the self-sufficiency push. China is being forced up the value chain by export controls. Gregory Allen's analysis (in allen-true-impact-export-controls) is the load-bearing finding: controls slow China but don't stop it. Huawei's Kirin 9000S proved SMIC can produce at 7nm; SMIC's N+2 process and Huawei's Ascend AI accelerators are the self-sufficiency evidence. DeepSeek (per allen-ai-race-2025) showed China can do more with less compute. The master-faq flags the "lagging edge war in the next 5 years" — Chinese foundries and IDMs expanding mature-node capacity (MOSFET, IGBT) aggressively, with BYD gaining NEV share and increasing internal sourcing. The note's stark line: "Mosfet, igbt are screwed too. Chinese foundry and IDMs are expanding capacities rapidly."

Who plays. SMIC (leading foundry, target of controls), Huawei (Ascend / Kirin), Chinese WFE makers, and the domestic-substitution equipment names. From the screenshot archive: ACMR (ACM Research) targeting 60% long-term market share in China cleaning equipment, with combined China memory capacity expansion of c120K wpm+ in 2026, most of it DRAM; domestic equipment companies expect significant growth on this WFE wave. Montage Technology's 6809 Jintide line is an Intel JV for China-market server CPUs — a direct China-strategy exposure (founders extracted RMB 1.92B, stake fell from 60% to 4%).

Investment angle. Two opposing reads. The bear read: China is a market that gets "replaced" — see the recurring warning on ACLS (Axcelis), described as "possibly the best funding short in semi-cap forever. Consists mainly of Chinese revenue that will be replaced, almost zero advanced logic revenue, and an overstated backlog." The bull read: domestic substitution is a real capex cycle for Chinese-domiciled equipment names (ACMR). For Western names, the angle is the inverse — China revenue at risk of substitution is a quality discount, not a premium.

Japan semiconductor ecosystem

What it is. Japan's mid-stream supply-chain position — optics, test/measurement, materials, connectors, metrology — explicitly NOT the leading-edge logic fabs. The thesis in one line from japan-semi: "not competing with TSMC on logic, but dominating niche component markets that are hard to replicate and increasingly strategically valuable. The friend-shoring thesis in stock form." Japan is the sweet spot: allied with the US, technologically advanced, but not in the China export-control crosshairs the way Taiwan is. These names trade at 10-20x earnings on the Tokyo Stock Exchange with thin international coverage.

The technology and who plays. The jp-semi-supplychain-comparison maps six companies across five niches, every one a friend-shoring component play:

Ticker	Company	Niche	Moat / market share
6754	Anritsu	RF/microwave T&M (5G/6G)	~5-8% global T&M; ~10-15% wireless test
6777	Santec	Optical components & instruments (tunable lasers, WDM, OCT)	Niche leader in tunable lasers
6855	Japan Electronic Materials (JEM)	Semiconductor probe cards	~8-10% global probe cards (#4-5)
5801	Furukawa Electric	Optical fiber/cable (diversified)	Top-5 global optical fiber
268A	Rigaku	X-ray analytical & semi metrology	~29% global XRD (#1)
3110	Nitto Boseki	Specialty glass fiber for GPU substrates	~90%+ specialty glass; 100% T-Glass monopoly

Additional Japan names flagged but not deep-dived: 5802 Sumitomo Electric (multi-core / ultra-low-loss fiber for datacenter interconnect; optics is 4.8% of revenue at 8.9% operating margin, targeting record ¥37B operating profit), 6834 Seikoh Giken (APC connectors, FY25 guidance raised to ¥30B revenue / ¥7.0B operating income), and gaps the vault itself names: Hamamatsu Photonics, Disco Corporation (dicing/grinding leader), Advantest (Japan's largest semi test).

The valuation data. From the 6-company comp (data late Feb / early Mar 2026, JPY): pure-play median NTM EV/EBITDA ~19.4x, P/E ~33x; non-pure-play median ~23.9x EV/EBITDA, P/E ~44.2x. Standouts: JEM cheapest at ~12.5x NTM EV/EBITDA with the highest ROIC (23%) and strongest forward revenue growth (+17.9%) on HBM probe-card demand; Nitto Boseki most expensive at ~31x on monopoly economics (NE-Glass at 6x premium to commodity E-Glass; customers NVIDIA, AMD, Apple, Google, Amazon) but >60% of revenue is from unrelated/unprofitable businesses and consensus PT sits ~40% below spot; Santec best quality (58% gross / 32% operating margin, ~20% growth, 25% ROE); Furukawa the highest-conviction-but-lowest-quality bet (+380% one-year move, 3.9% operating margin, market paying for optical-fiber volume not current margins); Rigaku the most balanced (#1 in XRD, ~16% operating-profit growth, 17.8x NTM EV/EBITDA). The Rigaku "Lab to Fab" thesis (268a-dong-thesis-analysis) argues true semiconductor exposure is 60%+ of profits versus the reported ~30% of revenue — a semi play hiding in plain sight. The Collyer Bridge Discord thesis extension adds the catalyst: "2nm or SiPho ramp would see Samsung place major Rigaku kit orders," plus Onto (ONTO) co-development on a hybrid X-ray/optical (T-SAXS) metrology engine, a claimed $100M new market.

Investment angle. Friend-shoring premium: as decoupling restricts China's access and the West tries to break China's supply-chain stranglehold, Japan's mid-stream position becomes more valuable. The open tensions the vault flags: is the 10-20x TSE multiple a discount to capture or a value trap from low growth/liquidity; how much does each name's China revenue exposure hurt (lost market) vs. help (strategic premium); small-cap liquidity/position-sizing risk; JPY currency risk (weak yen has been a tailwind); conglomerate discount (Sumitomo's optics is 4.8% of revenue); and friend-shoring durability — does the premium collapse if US-China tensions de-escalate. Disruptive threats from outside the peer set: Keysight (~3x Anritsu's T&M revenue), FormFactor + Technoprobe (dominate probe cards, bigger R&D than JEM), Bruker (Rigaku's main XRD competitor), and Chinese glass-fiber makers (a long-term but multi-year-qualification threat to Nitto Boseki).

Export controls & geopolitics

What it is. The US restriction regime on semiconductor technology exports to China, plus China's retaliation. The dedicated export-controls-china-semi page is a framework stub only — it lays out the intended structure (BIS restrictions, entity list / end-use controls, deemed-export rules, Chip 4 multilateral coordination, ASML/KLA/LRCX equipment-vendor responses, EDA software restrictions) but is "Awaiting source material from Dropbox." The substantive content sits in supply-chain-security.

The mechanisms. The January 2025 ai-diffusion-rules-2025 are described as "the most aggressive expansion yet" — Commerce Department controls on advanced computing ICs, AI model weights, and license exceptions for allies, creating a tiered system of compute access. Key thresholds (from the stub framework): advanced logic nodes sub-7nm, sub-14nm for memory, GPU/AI accelerator restrictions, lithography/deposition/inspection equipment, and design tools/IP licensing. Enforcement runs through the entity list, end-use controls (supercomputing, AI training, military), and deemed-export rules for foreign nationals.

China's retaliation — mineral weaponization. China has turned mineral export controls into the counter-weapon: germanium and gallium restrictions (2023, directly targeting semiconductor materials), antimony restrictions (2024, targeting the defense supply chain — flame retardants, ammunition). China controls ~97% gallium, ~82% germanium, 70-90% rare-earth processing (see critical-minerals). The open question the vault keeps returning to: how far do US chip restrictions go before triggering serious retaliation beyond minerals — e.g., a rare-earth magnet export ban critical for defense, EVs, and wind turbines.

The "Made in America" supply premium. From the Serenity method: AAOI transceivers earn a geopolitical-resilience premium versus Fabrinet/FN (Thailand), Innolight (China), Eptolink (China). Same pattern as INTC-to-TSM, MU-to-Hynix. This premium "matters more as export controls tighten."

Investment angle. Equipment and components made outside China get a strategic-value premium; Chinese-exposed revenue gets discounted (the ACLS short). The European compound-semi equipment names carry an explicit geopolitical moat: AIXA Aixtron's pitch calls out that China export barriers on GaN substrates make Western-sourced MOCVD/PECVD equipment more valuable; SOITEC is the French national champion in engineered wafers (Photonics-SOI, FD-SOI). The vault flags its own gaps here: no comprehensive export-controls timeline, no EU Chips Act tracking, no India strategy, ASML uncovered (the most geopolitically important equipment maker), no Taiwan-invasion scenario, South Korea (Samsung, SK Hynix) not mapped as a geopolitical actor, no ITAR/defense export-control framework.

Tariffs & trade policy

What it is. US tariff policy, Section 232 national-security investigations, and trade dynamics affecting semiconductor valuations and supply chains. The dedicated tariffs-trade-policy page is also a framework-only stub ("Dropbox sync pending"). It sketches the intended content: Section 232 investigations into domestic semiconductor capacity (equipment, materials, defense-critical chips, datacenter optics/photonics), MFN/de-minimis dynamics, equipment-vendor tariff incidence and pass-through, fab-location decisions, and the CHIPS Act reshoring linkage.

The substantive threads. commerce-232-semiconductor-investigation — a Commerce Department national-security investigation that could produce semiconductor-specific tariffs and re-rate equipment names broadly. apollo-us-consumer-tariff-2025 covers consumer/firm responses, supply-chain adjustment, and pricing pass-through. oliver-wyman-us-tariffs-asia projects significant supply-chain diversion across Asia — manufacturing-location shifts and trade-diversion patterns. warren-buffett-trade-deficit supplies the macro framing (asset erosion from persistent deficits). The stub's pending-content list also names Apollo tariff analysis and Buffett tariff views.

Investment angle. Tariffs reinforce the reshoring/CHIPS-Act thesis — they make domestic fabs and Western equipment more competitive while raising input costs (fab equipment, materials, steel/aluminum for fab construction). Fabless design companies have minimal direct exposure; equipment manufacturers and IDMs bear the incidence. The key unanswered questions the vault poses: do tariffs actually rebuild domestic capacity, who wins/loses under each scenario, and how tariffs interact with export-control policy. Trade diversion is also a SEA beneficiary story — Vietnam, Indonesia, Thailand, Philippines as friend-shoring destinations caught in the crossfire (country-vn, country-id, country-th, country-ph, bcg-sea-datacenters).

Foundry & process nodes

What it is. The leading-edge logic manufacturing race, mature-node economics, advanced packaging, and the reshoring footprint. This is where the CHIPS Act bet and the Intel 18A acid test live.

The manufacturing footprint reality. src-barclays-semi-manufacturing-footprint is the load-bearing finding: reshoring works for leading-edge (<7nm), but mature nodes (28nm+) will stay concentrated in Asia — "the economics don't support onshoring commodity production." The CHIPS Act's $52B into domestic fabs is subsidizing what is economically marginal at the mature end; the open question is whether it sticks when subsidies expire. The "long tail of smaller fabs" thesis (Fabricated Knowledge, Mar 2022) is the trailing-edge corollary: the fastest capex growth was NOT TSMC's ~40% YoY, but the long tail of IDMs and smaller fabs — UMC and STM guiding 100% YoY capex increases, NXP at 70%, Powerchip and TXN above 40%. "Lagging edge semiconductor is finally getting the investment it so desperately needs." ACLS was the equipment expression of this (second-place ion implant to AMAT; 80% of business lagging edge).

Intel 18A — the CHIPS Act acid test. Intel is the CHIPS Act's biggest bet on domestic leading-edge foundry, and the question is whether 18A process technology actually works. The full supply-chain map sits in themes/intel-supply-chain: 18A + EMIB (Embedded Multi-die Interconnect Bridge) advanced packaging ramp as a catalyst spanning burn-in test, IC substrates, metrology, semicap subsystems, manufacturing analytics, OSAT, foundry, and AI server power. Milestones: Panther Lake (first internal 18A product, sampling 2025, production 2026), Clearwater Forest (server, volume 2027), EMIB ramp, and — the critical unlock — Intel Foundry Services external customer wins. The dual-tailwind thesis: the 18A cycle runs simultaneously with the AI buildout, so suppliers get two demand curves. The swarm's top conviction picks across the layers were ONTO (optical metrology, net cash $640M underwriting 46x P/E, SK Hynix $240M agreement), ATS (ABF IC substrates, PEG 0.2-0.3x, ABF shortage widening to 42% by 2028), PDFS (fab analytics, 94% recurring revenue, 76% gross margin), 2308 Delta Taiwan (AI server power, Vera Rubin cycle Q3 2026), and UMC (mature foundry, H2 2026 ASP +10% hike unpriced, Intel 12nm JDA optionality). Burn-in sub-theme: AWX AEM (Intel-qualified HDMT) and S71 Sunright. OSAT: AMKR Amkor (Intel Foveros, Kim family 49.5% stake). Important corrections embedded: ATS's confirmed anchor is AMD not Intel; 8027.TWO's TGV customer is Japanese; 6809.HK is an Intel China-strategy play. The personal trade-log lesson (stocklog-2026-01-23-intc): INTC calls that gained 60-70% (potential 150k+) round-tripped to +10% — "being able to differentiate between hype and fundamentals" with ST/MT/LT perspective; INTC is in a long-term "execution phase" to prove.

Foundry players. TSMC (the logic benchmark; ~40% YoY capex reference point), INTC (18A / IFS), UMC (cheapest major foundry, mature), GFS GlobalFoundries (SiPho foundry positioning, sa-globalfoundries-sipho-foundry — the Western silicon-photonics foundry friend-shoring play), TSEM Tower Semiconductor (Israel-based, SiPho), SMIC (China, structural pricing pressure on UMC). The photonics-to-logic analog map (Jason's Chips, Mar 2026) is a useful pricing crib: TSEM↔TSM, LITE↔NVDA, COHR↔AMD, BESI↔AMAT, XFAB↔INTC, AAOI↔MU, POET↔Groq, CIEN↔CRDO, SMTC↔ALAB, SIVE↔LITE.

Investment angle. Semicap is named the biggest beneficiary of reshoring — "strongest and most meaningful year in WFE history" — with AMAT, LRCX, CAMT, ONTO, TOELY as the core and subsystems MKSI, ICHR, UCTT as the levered play. The foundry/node race is also where the bottleneck-hunting method (below) finds its small-cap upstream gems: advanced packaging (HBM hybrid bonding, glass/ABF substrates, EMIB) is the structurally short link.

Semiconductor research methodology

What it is. The how-to-think and where-to-look layers for analyzing this sector — captured as two named-investor method digests plus the basic semi vocabulary the vault is building.

FundamentEdge (Brett Caughran) — how to think. Ex-Maverick/D.E. Shaw/Citadel/Schonfeld PM teaching institutional fundamental analysis. The ten load-bearing moves: revenue-growth primacy (sustained 10%+ growth is chronically underpriced; low P/E as a buy trigger is a value trap); stocks follow numbers/revisions ("never short in front of large positive revisions"); respect the second derivative (rate-of-change of growth beats absolute); business quality is qualitative (high ROIC can be a risk factor without an oligopoly moat); WACC is broken — use true equity cost of capital (a 12x P/E stock has a higher cost of equity than a 30x stock); the comp sheet as a weekly triage engine (decompose YTD move into EPS-roll-forward vs. revision vs. P/E expansion); hidden thematic factors; three-layer-cake DD (desktop → primary → insight formation); the Digest/Contemplate/Ideate/Predict sector-outlook ritual (Core Four: organic revenue, margins, capital intensity, capital deployment); and underwriting the market itself (S&P 25-yr avg P/E 16.6x vs. current ~21-22x).

Serenity (@aleabitoreddit / Aleabit) — where to look. AI/semi supply-chain analyst, ex-RISC-V FDN, 206K followers, built a custom model because "frontier models hallucinate." The ten tactical moves, semi/photonics-specific: bottleneck hunting (trace hyperscaler capex upstream until you hit a sub-$1B-MC name with exclusive exposure — "the more well-known the name ($LITE, $COHR) the more crowded/priced-in. Go upstream for hidden gems"; the test is can-this-node-be-bypassed); second-order earnings read-throughs (after any beat, list the 4 upstream nodes and check which aren't priced); SEC 8-K supply-chain mapping (her SIVE thesis came from Marvell's 8-K disclosing Amazon warrants at $87 → Celestial → Sivers lasers → Amazon); "paper game" unwind recognition (fade moves driven by margin hikes/halts/forced liquidation, not fundamentals); fade news that violates engineering reality (silicon can't be swapped mid-validation — architecture-to-validation is 30-38 months); playbook recognition across cycles (NVDA's 2024 EML preallocation drove LITE +955%; March 2026 $2B each into COHR/LITE/MRVL is the same move); institutional front-run detection; qualification-cycle as leading indicator (AEHR qualification orders = "early AAOI stage"); chokepoint M&A thinking; and "alpha is free on X" (a Reddit experiment found Burry dead last at -11% 60d; only SemiAnalysis, Global Tech Research, Fabricated Knowledge beat benchmarks). Her concrete supply-chain maps: InP substrate (DGC phosphate → NCI refining → Sumitomo/JX/AXT melt → COHR/LITE fab → Innolight/Fabrinet packaging → NVDA/GOOGL); MLCC/PCB-drill (Murata/Taiyo Yuden/TDK/Kyocera/Yageo/Walsin; Sakai Chemical for barium titanate; Topoint/Union Tool 6278 for PCB drill). Her thesis portfolio screenshot showed 11 tickers at YTD +1,116% (HPS.A, SIQE, AXTI, SIVE, AAOI, LITE, NBIS, AEHR, TSEM, SOI), and TSEM was a +759% YTD home run from a 3-week discovery post.

The two layer well. Caughran = how to think (revenue primacy, 2nd derivative, TECC, comp-sheet columns, cap-allocation timing); Serenity = where to look (the upstream bottleneck the market hasn't priced). The vault wired Serenity's moves into the skills: /primer §14b (bottleneck hunting + "follow the capex" 4-layer beneficiary trace), /deep-dive §5 (upstream bottleneck check table), /earnings-recap §6 (second-order read-through 4-node adjacency), /filings (8-K supply-chain disclosures as first-class signal).

The vocabulary build. The master-faq and sector-data-research notes are the working glossary of terms the vault is still researching: wafer starts per month, semicap intensity, semicap yield by player, BSPDN (backside power delivery — a clear performance advantage requiring more etch/deposition tools), the standard WFE peer list (Tokyo Electron, ASML, AMAT, KLAC, LRCX), IMEC (Belgium transistor-scaling research house), MOSFET/IGBT and the IGBT/MOSFET glut, bit growth for DRAM/NAND, mature wafer starts, analog/IC design, coherent optics vs. PAM4 economics, and the HBM / hybrid bonding / DRAM-future packaging relationship (SemiAnalysis packaging series). The AI-capex historical analog: 1999 fiber spend ~$120-140B (≈$213B in 2024 dollars) vs. ~$100B AI capex in 2024.

Doug O'Laughlin / Fabricated Knowledge — the semi-core paid anchor. [@fabknowledge] / [@FoolAllTheTime], one of the canonical paid semiconductor Substacks. His tracked calls run through the screenshot archive: the Jan 2022 "Valuations of the Space" EPS-growth-vs-price screen ("Automotive, RFFE, and solar is the place to play"); the subsystems trade ("I continue to love the Subsystems trade... all revenue related to HBM continues to shine"); the Nov 2024 universe takes (WOLF going-concern at $5B debt / $1.2B market cap / $1B/yr burn — "will go bankrupt by end of 2025 if it doesn't raise cash"; ACLS as "best funding short in semi-cap forever"; ARM doubling CSS ASP; ANET doubling TAM on AI back-end); and the Mar 2025 automotive-semi bottom call (Renesas, ON, STM, IFX cheapest — "in the downturn it never seems like it can be better. Yet I think it will in automotive"). Other tracked sources: Daniel Nystedt (@dnystedt, Asian-tech/semis breaking news, Taiwan/Korea/China primary-language reporting) and Insane Analyst / Irrational Analysis (@insane_analyst, anonymous engineer, semi deep dives).

Semi-IP licensing

What it is. The licensing layer — IP cores, EDA tools, and the royalty/ASP economics that sit above the fab. Thinly covered in the vault but flagged repeatedly.

The players and the angle. ARM is the central name: the Nov 2024 read is "doubling or more of ASP for CSS [Compute Subsystems] products. Taking price on customers," plus an ARM/Apple overhang and a potential Intel-bid angle on QCOM. EDA is named as a changing landscape to watch (sector-data-research: "changing trend in EDA landscape") — the design-tool oligopoly (Synopsys, Cadence, Siemens EDA) that export controls also target via EDA software restrictions to China (per the export-controls stub). Semi-IP licensing connects to the foundry layer because the IP/royalty model is the high-margin, capital-light counterpoint to fab capex — and to export controls because design tools and IP licensing are explicit restriction thresholds. The vault has no dedicated EDA or ARM page yet; this is a coverage gap.

Cross-cutting subsector: silicon photonics & optical interconnect

Worth isolating because it threads through Japan, foundry, and the bottleneck method. CPO (co-packaged optics) is "happening earlier than expected" (Serenity). The technology: CPO requires ~4x the SiPho wafer content versus pluggables (SOITEC, 95%+ Photonics-SOI share, GlobalWafers exiting). Players span TSEM (SiPho foundry / glass-substrate home run), GFS (Western SiPho foundry), the laser/transceiver chain (AAOI, LITE, COHR, Sivers SIVE, Celestial, Ayar, Enplas for Google TPU), Taiwan CPO ecosystem (Shunsin 6451 Foxconn ODM ~$1.4B MC, PCL, Fittech), and the Japan optics names (Santec, JEM, Seikoh Giken, Sumitomo). The investment angle is the purest expression of bottleneck hunting — trace NVDA/hyperscaler CPO demand upstream through transceivers → MBE/MOCVD reactors (VECO, Aixtron) → epiwafer foundries (IQE $175M, Landmark) → InP substrate (AXTI) until you reach the under-priced small cap.

Value chain

The semiconductor value chain is a long, vertical stack that runs from raw substrate inputs through wafer fab, advanced packaging, test, and finally into the systems — datacenters, servers, automotive, telecom — that consume the chips. The economically interesting truth, and the organizing principle of the vault's research, is that the margin pool is wildly uneven across the stack and the bottlenecks rarely sit where the household names sit. The recurring discipline pulled from the sources: trace hyperscaler capex backwards through the chain until you hit a node that cannot be bypassed, and that node — not the GPU vendor everyone watches — is where the pricing power and the under-priced equity usually live.

The stack, front to back

Roughly the chain reads: raw materials and feedstock → engineered substrates / wafers → epiwafer growth → front-end wafer fab (logic + memory) → metrology / inspection / process control → advanced packaging (substrates, interposers, hybrid bonding, EMIB/Foveros) → assembly & test (OSAT, burn-in, probe cards) → board-level (PCB, MLCC, connectors, optics) → system integration (servers, power, optical interconnect) → end markets. Wrapping around the front end is the wafer-fab-equipment (WFE) layer — litho, deposition, etch, ion implant, inspection — and the subsystems and consumables that feed those tools. Two cross-cutting layers sit on top of everything: EDA/design tools, and the geopolitical/export-control overlay that determines who is even allowed to buy from whom.

The vault's framing splits the chain into three regimes by margin and moat: (1) the leading-edge fabs and the irreplaceable WFE/EDA oligopolies, where margins are high and the moat is technological; (2) the mid-stream niche components — optics, connectors, metrology, probe cards, specialty glass — where Japan dominates and the moat is qualification difficulty and precision manufacturing; (3) the commoditized layers (mature-node wafer volume, cable, low-end OSAT) where margins are thin and the only edge is scale or geographic/geopolitical positioning. The strategic insight from supply-chain-security: reshoring economics only work for regime (1) — leading edge under ~7nm — because the CHIPS Act can subsidize what's strategically critical. Mature nodes (28nm+) stay concentrated in Asia because the unit economics never support onshoring commodity production. Barclays' manufacturing-footprint work is the anchor for that split.

Where the margin pools sit (Japan mid-stream comparison)

The cleanest per-stage economics in the vault come from the 6-company Japan supply-chain comparison (jp-semi-supplychain-comparison), which maps gross margin, operating margin and ROIC across distinct stack positions. The numbers expose how dramatically profitability varies by where you sit:

Company	Ticker	Stack position	Gross margin	Op margin	ROIC	Moat strength	Market share
Santec	6777	Optical components / T&M (WDM, tunable lasers)	58.1%	31.6%	~20%	Moderate-Strong	Niche leader tunable lasers / OCT
Rigaku	268A	X-ray metrology / process control	~56%	17.7%	9.7%	Strong	~29% global XRD (#1)
JEM	6855	Semiconductor probe cards	41.5%	22.8%	23.0%	Narrow	~8-10% global probe cards (#4-5)
Nitto Boseki	3110	Specialty glass fiber (GPU substrates)	37.0%	15.1%	8.4%	Very Strong	~90%+ specialty glass; 100% T-Glass
Anritsu	6754	RF/microwave T&M	50.6%	10.7%	~9%	Moderate	~5-8% global T&M; ~10-15% wireless test
Furukawa	5801	Optical fiber/cable (commodity-adjacent)	16.8%	3.9%	5.7%	Moderate	Top 5 global optical fiber

Read top to bottom this is a margin ladder: a tunable-laser specialist (Santec) earns a 32% operating margin and ~20% ROIC; the optical-fiber-and-cable conglomerate (Furukawa) at the volume/commodity end of the same optics chain earns 3.9% operating margin and a 5.7% ROIC that barely clears its cost of capital. Same physical product family (light through glass), opposite ends of the value chain, ~8x difference in operating margin. The lesson the vault keeps drawing: the money is in the precision/qualification-gated nodes, not the volume nodes.

Pricing power tracks moat type, not size. Nitto Boseki — the smallest moat-by-revenue-share story but the strongest pricing position — prices its NE-Glass at a 6x premium to commodity E-Glass and runs a near-100% monopoly on T-Glass for GPU substrates, with order books full through 2026 and NVIDIA, AMD, Apple, Google and Amazon as named customers. That is a single-node chokepoint: it sits in regime (2) but behaves like regime (1). The risk is the mirror image — it is single-product-line risk, and ~62% of Nitto Boseki's revenue comes from unrelated low-growth businesses (medical, composites, insulation) that drag the consolidated margin.

Bottleneck-tier analysis (Serenity's upstream walk)

The vault's most developed value-chain methodology is Serenity's "bottleneck hunting" (serenity-method): start at the well-known demand driver and walk upstream until you hit a node that cannot be bypassed and has only one or two qualified suppliers. The crowded, priced-in names sit downstream; the under-priced equity sits at the bottleneck. Her worked example for the optical-transceiver chain:

hyperscaler transceiver demand → AAOI (transceiver) → MBE/MOCVD reactors (VECO, Aixtron) → epiwafer foundries (IQE ~$175M MC, Landmark ~$3.6B) → substrate (AXTI). She keeps walking until she finds a sub-$1B-MC name with exclusive exposure. The test she applies: "can this node be bypassed? If not, it's a bottleneck, not a supplier." And the crowding rule: "The more well known the name ($LITE, $COHR) the more crowded/priced-in. Go upstream for hidden gems."

She maps the InP (indium phosphide) substrate chain for merchant optical explicitly as a prose value chain:

DGC (phosphate mine) → NCI (refining) → Sumitomo / JX / AXT (melt) → COHR / LITE (fab) → Innolight / Fabrinet (packaging) → NVIDIA / GOOGL (ASIC/GPU). Her one-line conclusion: "West depends on NCI." An alternate framing she uses: Vital → AXTI → LITE/AVGO/COHR + MRVL/MTSI → photonics BOM winners. The bottleneck in this chain is upstream refining/melt, not the transceiver fab everyone watches.

She also maps the MLCC / PCB-drill bottleneck chain (relevant because every AI board is a BOM of passives and drilled multilayer PCB):

• MLCC: Murata (6981), Taiyo Yuden (6976), TDK (6762), Kyocera (6971), Yageo (TWSE 2327), Walsin (TWSE 2492), VSH (US)
• Raw material into MLCC: Sakai Chemical (4078) — Barium Titanate (the upstream feedstock node)
• PCB drill: Topoint (8021), Union Tool (6278), Kyocera (repeat)

And the FC-BGA substrate chemistry chain surfaced in the screenshot archive: MEC (4971 TYO, ~$1B MC) is described as the sole provider of the molecular-bonding "super-roughening" chemistry that keeps organic FC-BGA substrates from cracking under GPU thermal loads. Substrate makers SEMCO, Ibiden and Shinko depend on MEC; competitors (MKS, JCU) "lack laser focus." The flagged disruption risk: the shift to glass substrates at 1.4nm / 1.6T speeds makes traditional roughening obsolete mid-2026 — a bottleneck node that could be displaced by a node-change higher in the stack.

The companion CPO/laser chain she traces: NVIDIA's March 2026 commitment of ~$2B each into COHR / LITE / MRVL is read as the same capacity-lock playbook as NVIDIA's 2024 EML preallocation that drove LITE +955%. The bottleneck this cycle is CW/EML lasers and CPO; the upstream small-cap she campaigns on is SIVE (Sivers), the laser supplier into JBL/MRVL/Ayar/Amazon-Celestial whose competitors trade at ~4x the multiple. The structural point for the value-chain map: when a downstream buyer pre-pays to lock capacity, the bottleneck node is being signposted by the buyer's own balance sheet — and chokepoint M&A (e.g., AVGO acquiring a bottleneck laser supplier to deny it to NVDA/MRVL) becomes the strategic endgame.

The photonics-to-logic mapping (pricing-power crib sheet)

A useful cross-stack mental model in the archive (screenshot-sweep-2026-04-05, "Jason's Chips") maps each photonics-chain name to its logic-chain analog — a way to reason about which value-chain position each occupies and therefore what pricing power to expect:

Photonics name	Logic-chain analog
TSEM	TSM (foundry)
LITE	NVDA
COHR	AMD
BESI	AMAT (equipment)
XFAB	INTC
AAOI	MU
POET	Groq
ALMU	Cerebras
AXTI	SNDK
CIEN	CRDO
SMTC	ALAB
SIVE	LITE

The implication drawn: the right-hand (logic) column is mostly already covered/priced; the left-hand (photonics) column — TSEM, COHR, BESI, AAOI, CIEN, SMTC — is where the multi-bagger value-chain positions sit, and Soitec sells the engineered substrate into most of column-left.

Advanced packaging — the EMIB / Foveros / OSAT layer

Advanced packaging is the value-chain stage the vault treats as structurally bottlenecked and demand-stacked, because Intel's 18A/EMIB ramp and the AI-accelerator buildout hit the same packaging supply chain simultaneously (themes/intel-supply-chain, briefings/2026-04-26-intel-supply-chain). The packaging sub-chain maps to distinct economic positions:

• IC substrates (ABF): AT&S ATS — carries the chiplets. The structural call: ABF substrate shortage widens to ~42% by 2028, which is a supply bottleneck that hands pricing power to substrate makers. (Thesis correction logged: AT&S's confirmed anchor is AMD at Kulim, not Intel — the Intel link is pilot-stage double-sided ABF tests.)
• Metrology / inspection / process control: Onto Innovation ONTO, Rigaku 268A. As nodes shrink and packaging stacks (3D DRAM/NAND, HBM), the number of metrology steps rises — more complexity, more inspection, more pricing power for the leader. Rigaku's hybrid T-SAXS engine (X-ray + optical, co-developed with Onto) is flagged as a ~$100M new metrology market.
• Fab analytics software: PDF Solutions PDFS — the highest business quality in the cluster: 94% recurring revenue, 76% gross margin. Software sits above the physical chain and captures yield-optimization value across it.
• Semicap subsystems: Ultra Clean Holdings UCTT — the levered play on fab buildout. Per-unit economics quoted: 35-40 cents of incremental gross margin per dollar of WFE revenue. Subsystems are the volume-leverage node — operationally levered to fab capex but lower-moat than the tool OEMs they feed.
• OSAT / packaging assembly: Amkor AMKR (Intel Foveros packaging, Arizona campus), and the burn-in/test sub-layer below it.

Burn-in and test sit at the back of the chain but become a forced node in two cases the sources call out: new package architectures require 100% burn-in, and HBM3e/HBM4 mandate wafer-level burn-in "with no substitute." That "no substitute" language is the bottleneck signature — AEM Holdings AWX (Intel-qualified HDMT system-level burn-in) and Sunright S71/KESM sit here, with the consumable (WaferPak) the recurring-revenue annuity within the test node.

Probe cards — a chokepoint with multi-year lead times

Probe cards are a distinct test-stage bottleneck with their own supply economics. From the screenshot archive: AI-chip-testing probe-card lead times ran to 6 months with some orders showing 2-year visibility (MPI Corp, 6223 TWSE), and HBM probe-card lead times of 16-24 months were cited via a Taiwan M-H Corp report, with Union Tool / 6278 adding 50% capacity at Zhubei in H2 2026. Lead times that long are the clearest evidence of a value-chain choke point — demand structurally exceeds qualified capacity. JEM 6855 sits here (~8-10% global share, #4-5), competing against the larger FormFactor and Technoprobe (Italy), both with larger R&D budgets — i.e., JEM is the cyclical small-cap inside a node dominated by a few qualified players, with revenue that swung 23.6B → 17.5B → back inside three years.

Mature-node and analog — the "lagging edge" sub-chain

A distinct value-chain theme the sources flag (Fabricated Knowledge, via screenshot-insights-archive): the fastest capex growth historically was not TSMC's leading edge (~40% YoY) but the long tail of IDMs and smaller fabs — UMC and STM guiding 100% YoY capex increases, NXP ~70%, Powerchip and TXN above 40%. "Lagging edge semiconductor is finally getting the investment it so desperately needs." ACLS (ion implant, #2 to AMAT, 80% lagging-edge) is named as the pure lagging-edge tool play. The structural counter-risk in this sub-chain: MOSFET/IGBT glut — Chinese foundries and IDMs expanding capacity rapidly while BYD increases internal sourcing in NEVs, compressing pricing for the merchant analog/power suppliers. The mature-node layer is where Chinese domestic substitution bites first (it is the regime-3 commodity end), which is exactly why reshoring economics don't favor it.

Equipment, subsystems and the WFE oligopoly

The front-end equipment layer is the high-moat oligopoly the master FAQ identifies as the peer set: Tokyo Electron, ASML, AMAT, KLAC, LRCX. Below the tool OEMs sit subsystems (MKSI, ICHR, UCTT) — the levered play on fab buildout — and the watchlist framing in supply-chain-security calls semicap "the biggest beneficiary of supply chain reshoring," naming AMAT, LRCX, CAMT, ONTO, TOELY and "the strongest and most meaningful year in WFE history." The technical driver pulled from the FAQ: backside power delivery (BSPDN) is a clear performance advantage, and implementing it in manufacturing requires more etch and deposition tools — a node-transition that pulls demand through the etch/dep tool node specifically. IMEC (Belgium, near ASML) is the upstream R&D node that defines what the tool makers must build next.

The compound-semi equipment niche has its own geopolitical-moat dynamic (supply-chain-security): Aixtron AIXA (MOCVD/PECVD) — China export barriers on GaN substrates make Western-sourced compound-semi equipment more valuable, an explicit case where the export-control overlay creates a value-chain moat rather than just a risk. Soitec SOITEC (engineered wafers — Photonics-SOI, FD-SOI) is the French substrate champion: per screenshot-sweep-2026-04-05, CPO requires ~4x the SiPho wafer content of pluggables, Soitec holds 95%+ Photonics-SOI share, and GlobalWafers is exiting — a tightening at the engineered-substrate node feeding the CPO chain.

Server / power / interconnect — the system-integration tail

At the back of the chain, the AI-server BOM pulls demand into power and interconnect: Delta Electronics Taiwan 2308 and its Thai subsidiary DELTA (AI-server PSUs, ~60% AI PSU share for the parent, Vera Rubin power cycle as the Q3 2026 catalyst), Sumitomo Electric 5802 and Furukawa 5801 (optical fiber/cable for datacenter interconnect — the high-margin Infocomm slice inside a low-margin conglomerate), and the optical-connector specialists JEM/Seikoh Giken 6834 (APC angled-physical-contact polishing, the precision-gated node in fiber termination). The Taiwan CPO ecosystem surfaced in serenity-method extends the system tail: Shunsin (6451, Foxconn CPO ODM for NVDA), Enplas (Google TPU / NVDA / TSMC optical), Fittech (Browave/Foci), and PCL (Broadcom-related) — equipment and ODM nodes the market under-covers.

Supplier/customer relationships and the geopolitical overlay

Two relationship dynamics shape pricing power across the whole chain. First, capacity-lock pre-buys: when a downstream giant (NVIDIA into COHR/LITE/MRVL, Amazon warrants into Marvell-Celestial-SIVE) commits cash or warrants to lock upstream capacity, it both signals where the bottleneck is and transfers value upstream — the 8-K disclosing those agreements is, per Serenity, "a literal map of who's locked into whom." Second, the export-control overlay (export-controls-china-semi, supply-chain-security): US controls choke China's access to leading-edge tools, EDA and HBM, which (a) makes Western/allied mid-stream suppliers strategically more valuable — the friend-shoring premium that underpins the Japan thesis — and (b) drives Chinese domestic substitution at the mature/commodity end (SMIC N+2, Huawei Ascend, ACMR targeting 60% China cleaning-equipment share, ~120K wpm of new China memory capacity in 2026). The "Made in America" supply premium is the chain's pricing expression of geopolitics: AAOI transceivers price at a premium to FN (Thailand), Innolight (China) and Eptolink (China) purely for geopolitical resilience.

The clean summary of where pricing power lives: the irreplaceable nodes are litho/EDA/WFE (technological oligopoly), specialty materials with single qualified suppliers (Nitto Boseki T-Glass, MEC roughening chemistry, NCI InP refining), precision-gated mid-stream components (probe cards, APC connectors, X-ray metrology), and any node a hyperscaler is paying upfront to lock. The thin-margin, low-power nodes are commodity wafer volume, optical cable, low-end OSAT, and the merchant analog/power suppliers exposed to Chinese capacity expansion. The whole map is best read backwards from end-market capex, not forwards from raw silicon — that is the discipline the sources converge on.

Players

The companies that populate the semiconductors / geopolitics map fall into seven clusters: the foundries and IDMs at the center of the reshoring fight, the Japanese friend-shoring ecosystem, the European equipment champions, the US plays, the China-exposed names, the Intel 18A / EMIB supply chain, and the photonics / CPO bottleneck chain. Company-specific deep detail lives on the ticker pages; what follows is positioning plus the comparative view, with the source comp tables reproduced where they carry unique numbers.

Foundries & IDMs — the reshoring fight

INTC (Intel) is the CHIPS Act's biggest single bet: the only Western attempt at a domestic leading-edge foundry. The entire reshoring thesis runs through whether 18A process technology actually works — it's the acid test. New CEO Lip-Bu Tan (appointed Mar 2025), ~$217B market cap as of Mar 2026, trading off a deeply depressed base ($17.67–$54.60 52-week range). The investment angle is binary and long-dated: Intel is in the "execution phase," a multi-year game to prove. The stocklog stocklog-2026-01-23-intc is a hard lesson on the difference between hype and fundamentals — INTC calls that ran up 60-70% (potential gains over $150k) round-tripped to +10% on greed and complacency. Pink's 7-Apr accumulation list flags INTC/AMD on the agentic-CPU angle.

UMC (United Microelectronics) is the cheapest of the major mature-node foundries and the only name in the Intel-supply-chain cluster that isn't obviously overpriced (PEG ~1.0–1.3, FCF yield 5.6%). The Intel 12nm joint development agreement is 2027 optionality; the nearer catalyst is an H2 2026 ASP hike of ~10%, confirmed April 17 but not yet in sell-side consensus. Carries SMIC structural pricing pressure and a Taiwan geopolitical premium. Verdict from the Intel-18A swarm: STAGE IN at $11.00–$11.30, B+ management (1.31% insider ownership). Sits in the mature-node "long tail" thesis below.

TSM (TSMC) is the gravitational center the whole map orbits — it does NOT have a dedicated wiki ticker folder yet, but it anchors comparisons throughout. TSMC's ~40% YoY capex is the benchmark the trailing-edge thesis measures against; its Japan and Arizona expansions are the friend-shoring buildout in physical form. In Jason's Chips photonics→analog mapping, TSM is the logic equivalent of TSEM (the substrate-foundry analog). GFS (GlobalFoundries) is the Western mature-node and silicon-photonics foundry — the friend-shoring SiPho play. The Nov 17 2025 acquisition of Singapore's Advanced Micro Foundry (AMF, 300+ Asia customers, >$75M 2026 revenue contribution, coherent-optics-for-long-haul IP) plus InfiniLink (Cairo) makes GF claim the title of "largest pure-play silicon photonics foundry by revenue." Doug O'Laughlin's Fabricated Knowledge connects GFS via "Micron's Guide and GlobalFoundries."

Mature-node ("trailing edge") foundries are a structural theme of their own, sourced from Fabricated Knowledge Mar 2022 ("The Long Tail of Smaller Fabs"): the fastest capex growth was NOT TSMC's ~40% but the long tail of IDMs and smaller fabs — UMC and STM both guided 100% YoY capex increases, NXP ~70%, Powerchip and TXN faster than 40%. The thesis: lagging-edge semiconductor was finally getting the investment it desperately needed. The Barclays manufacturing-footprint read (src-barclays-semi-manufacturing-footprint) is the bookend: reshoring works for leading-edge (<7nm), but mature nodes (28nm+) stay concentrated in Asia because the economics don't support onshoring commodity production.

Japan friend-shoring ecosystem — the supply-chain sweet spot

Japan is the most interesting friend-shoring story: allied with the US, technologically advanced, but not in the China export-control crosshairs the way Taiwan is. These names dominate niche mid-stream component markets (optics, test, metrology, materials, connectors) that are hard to replicate and trade at 10-20x earnings on the TSE with thin international coverage. The head-to-head analysis lives at JP semi supply chain comparison (6 companies) and Santec vs JEM vs Anritsu (optical T&M players); investment criteria at JP.

Positioning by company:

• 6754 Anritsu — RF/microwave and 5G/6G test & measurement. Most defensive name (net cash ~10% of market cap), slowest growth (~6-8%). Early 6G R&D with Qualcomm. 6754-anritsu-deep-dive, 6754-anritsu-mgmt-dd.
• 6777 Santec Holdings — optical components (WDM) and tunable lasers / optical T&M. Best-in-class margins (58% gross, 32% operating), fastest 3-year growth (39% CAGR), but smallest and priced for perfection (8.0x EV/Rev). 6777-filings.
• 6855 JEM (Japan Electronic Materials) — semiconductor probe cards (~95% of revenue); the APC-connector / precision-polishing story in the optics framing. Cheapest in the group, highest ROIC (23%), HBM-driven demand, but most cyclical and thin coverage. 6855-jem-deep-dive, 6855-mgmt-dd, 6855-profile, 6855-buy-checklist. Probe-card cycle is hot: lead times 16-24 months, Union Tool (6278) adding 50% capacity at Zhubei H2 2026; complementary names MPI Corp (6223 TWSE, 6-month AI-chip probe-card lead times) and Macronix (2337). Collyer Bridge's top list: Metasurface, Rigaku, Macronix, JEM.
• 268A Rigaku Holdings — X-ray metrology (#1 globally in XRD, ~29% share) for semiconductor fabs. The metrology pick in friend-shoring; the Dong/Shim thesis argues true semi exposure is 60%+ of profits (vs reported 30% of revenue) — "Lab to Fab," a semi play hiding in plain sight. Catalyst: 2nm/SiPho ramp = Samsung major Rigaku kit orders; hybrid T-SAXS metrology engine (X-ray + optical), co-developed with Onto (ONTO), $100M new market, likely 3D DRAM. 268a-rigaku-deep-dive, 268a-mgmt-dd, 268a-profile, 268a-buy-checklist.
• 5802 Sumitomo Electric — $33.6B conglomerate; the Infocommunications optics segment is 4.8% of revenue but the highest-margin division (8.9% operating margin, targeting record ¥37B operating profit). Multi-core / ultra-low-loss fiber for datacenter interconnect. Conglomerate-discount question: are you paying for the auto business to get optics? Note: do not abbreviate as "SEI."
• 6834 Seikoh Giken — APC connectors, the fiber-optic industry standard; revenue guidance raised to ¥30B, operating income ¥7.0B (FY25). 6834-buy-checklist.
• Nitto Boseki (3110) — specialty glass fiber for GPU substrates; near-monopoly (~90%+ share, 100% T-Glass), full order books through 2026, customers NVIDIA/AMD/Apple/Google/Amazon. No dedicated deep dive; referenced in the comparison.

The Furukawa Electric (5801) name appears in the comp table as the optical-fiber/cable conglomerate and is the "highest-conviction secular growth bet but lowest-quality earnings" — see 5801T-furukawa-ai-infra and 5801T-furukawa-dc.

The full 6-company comp table (data as of late Feb / early Mar 2026, JPY):

Ticker	Company	Sector / Niche	Market Cap	EV	Key Revenue Segments
6754	Anritsu Corp	Test & Measurement (5G/6G)	384B	261B	T&M equipment (~60%), Product Quality Assurance (~25%), Other (~15%)
6777	Santec Holdings	Optical Components & Instruments	~225B	~225B	Optical components (WDM), optical measuring instruments, sensing/imaging
6855	Japan Electronic Materials (JEM)	Semiconductor Probe Cards	107B	~100B	Semiconductor testing probes (~95%), electron tube parts (~5%)
5801	Furukawa Electric	Optical Fiber/Cable & Diversified	2,130B	2,356B	Electrical & Electronics (~54%), Infrastructure/Optical (~29%), Functional Products (~14%)
268A	Rigaku Holdings	X-ray Analytical & Semiconductor Metrology	425B	~463B	Multipurpose analytical (~45%), semiconductor process control (~26%), parts/services (~29%)
3110	Nitto Boseki (Nittobo)	Specialty Glass Fiber (GPU substrates)	927B	~922B	Electronic materials (~38%), medical (~12%), composite (~12%), other (~38%)

LTM trading comparables (pure-play vs non-pure-play split):

Company	Ticker	Type	EV/Rev	EV/EBITDA	EV/EBIT	P/E	P/FCF	P/B	PEG
Anritsu	6754	Non-Pure	2.3x	14.7x	21.5x	41.3x	22.8x	3.1x	~5.5x
Santec Holdings	6777	Pure	8.0x	23.7x	25.6x	36.3x	52.3x	9.0x	~1.3x
Japan Electronic Materials	6855	Pure	3.4x	12.3x	14.7x	21.8x	39.6x	3.3x	~0.6x
Furukawa Electric	5801	Non-Pure	1.9x	25.5x	46.4x	40.0x	40.5x	5.3x	~2.9x
Rigaku Holdings	268A	Pure	4.9x	21.2x	27.7x	37.3x	119.4x	4.8x	2.4x
Nitto Boseki	3110	Non-Pure	8.5x	37.8x	56.1x	66.2x	154.6x	5.3x	N/M
Pure-Play Median			4.9x	21.2x	25.6x	36.3x	52.3x	4.8x	1.3x
Non-Pure-Play Median			2.3x	25.5x	46.4x	41.3x	40.5x	5.3x	2.9x
Overall Median			4.2x	22.5x	26.7x	38.8x	46.4x	5.1x	2.4x

Operating metrics (LTM):

Company	Ticker	Revenue (LTM)	Gross Margin	Op Margin	Net Margin	FCF Margin	ROIC	Net Debt/EBITDA	Div Yield
Anritsu	6754	113.0B	50.6%	10.7%	8.2%	14.9%	~9%	Net cash	1.33%
Santec Holdings	6777	28.0B	58.1%	31.6%	22.1%	17.9%	~20%	Net cash	0.95%
Japan Electronic Materials	6855	29.8B	41.5%	22.8%	16.4%	9.2%	23.0%	Net cash	1.00%
Furukawa Electric	5801	1,202B	16.8%	3.9%	2.8%	4.4%	5.7%	2.7x	0.43%
Rigaku Holdings	268A	94.2B	~56%	17.7%	12.1%	3.8%	9.7%	1.7x	1.00%
Nitto Boseki	3110	109.0B	37.0%	15.1%	11.8%	5.5%	8.4%	Net cash	0.45%

Moat ranking from the comparison: Nitto Boseki (very strong, ~90%+ specialty glass monopoly, NE-Glass at 6x premium to commodity E-Glass) > Rigaku (strong, ~29% XRD #1) > Santec (moderate-strong, niche tunable-laser/OCT leader) > Anritsu / Furukawa (moderate) > JEM (narrow, ~8-10% probe cards, #4-5 player). Disruptive threats from outside the set: Keysight (~3x Anritsu's T&M revenue), FormFactor + Technoprobe (dominate probe cards, out-spend JEM), Bruker (Rigaku's main XRD rival), Chinese glass-fiber makers (long-term threat to Nitto Boseki, but multi-year qualification gap).

Comparison verdicts: JEM cheapest on forward fundamentals (~12.5x NTM EV/EBITDA vs ~21x group median, 23% ROIC, +17.9% growth) but most cyclical. Nitto Boseki most expensive and highest de-rating risk (~31x NTM EV/EBITDA, ~47x normalized P/E; consensus PT ~15,380 vs ~25,470 price = ~40% gap). Santec best quality/growth combo. Furukawa the most binary bet (+380% 1-year, 44x NTM P/E on 3.9% operating margins). Rigaku the most balanced risk/reward. Anritsu the most defensive (cheapest NTM EV/EBITDA at 13.2x, strongest net cash). The Hyundai Japan/Taiwan field trip (Hyundai-Field-Trip-Japan-Taiwan-Companies) gives the Korean-perspective read on TSMC expansion, HBM localization, and the AI-chip ecosystem.

Other Japan-adjacent / Asia plays: UIS (2404.TW, Taiwan, supercycle thesis — 2404TW-uis-supercycle); Tower Semiconductor TSEM (Israel foundry, SiPho — friend-shoring but not Japanese — TSEM-optical-backbone, TSEM-si-photonics).

European equipment makers

AIXA (Aixtron) — MOCVD/PECVD compound-semi deposition equipment. The investment angle is an explicit geopolitical moat: China export barriers on GaN substrates make Western-sourced compound-semi equipment more valuable. aixa-deep-dive, aixa-profile, aixa-mgmt-dd, aixa-comprehensive-pitch-docx, kerrisdale-aixtron-long-thesis. (Pink held — IBKR snapshot shows AIXA 350 shares, +19.3%.) SOITEC (Soitec) — engineered wafers (Photonics-SOI, FD-SOI), French national champion in advanced substrates, ~95%+ Photonics-SOI share; CPO requires 4x SiPho wafer content vs pluggables, and GlobalWafers is exiting. soitec-deep-dive, soitec-vs-aixa-showdown. ASML is the most geopolitically important equipment maker and is flagged as a coverage gap — not yet researched.

US semiconductor plays

AMD — agentic-CPU thesis (amd-agentic-cpu-thesis), server-market and reshoring-demand beneficiary; on Pink's 7-Apr agentic-CPU accumulation list with INTC. AEHR — power-semi / wafer-level burn-in test equipment; qualification orders from a leading optical-transceiver company flagged by Serenity as "early AAOI stage" — the qualification phase is the entry point, not the volume ramp. AEHR; active buy order @ $44 noted in screenshots. The semicap WFE basket from buy-list-2026 is the biggest beneficiary of reshoring: AMAT, LRCX, CAMT, ONTO, TOELY ("strongest and most meaningful year in WFE history"); subsystems MKSI, ICHR, UCTT as levered fab-buildout plays. The FAQ peer list for transistor-scaling / WFE: Tokyo Electron, ASML, AMAT, KLAC, LRCX (IMEC is the research house; BSPDN cited as a clear performance advantage requiring more etch/deposition).

IP licensors & precision-timing

CEVA — leading IP licensor in wireless comms and edge processing (cellular-modem DSP, Wi-Fi/Bluetooth, edge AI); high-margin royalty model (85-90% gross), ~$107M revenue, 60-75% R&D intensity. Asset-light play on 5G/IoT/edge-AI; watching. IDCC (InterDigital) — asset-light "toll road" patent licensing for wireless (5G/6G), video codec, and AI; recurring royalty streams, ~$9.8B market cap, Anfa PT $700; risk is patent litigation/invalidation. SITM (SiTime) — MEMS timing / silicon oscillators displacing quartz crystals; transceiver and likely GB200 design wins (jitter/power/PPM tradeoffs), flagged for large transceiver-revenue beats. (Pink held — IBKR Jul-2025 snapshot: 30 shares @ $210.06, +1.32%.) SITM-smart-ring covers an adjacent angle. SITM was on the Jan-2022 Fabricated Knowledge "positive EPS growth + negative price return" bargain-hunting list.

China-exposed names

The export-controls / tariffs framing pages (export-controls-china-semi, tariffs-trade-policy) are still framework stubs awaiting Dropbox sync, so most named China players sit in the briefings and screenshot archives rather than dedicated pages. Montage Technology (6809.HK) is the cleanest China-strategy name: its Jintide server CPUs are an Intel JV for the China market, so the exposure is to Intel's China strategy, not 18A/EMIB — with a binary export-control tail (Jintide relies on both TSMC and Intel supply, both restricted) and founders who extracted RMB 1.92B (stake 60% → 4%). China WFE / domestic-substitution names from the screenshot archive: ACMR (60% long-term China cleaning-equipment share target; combined China memory capacity expansion c120K wpm+ in 2026, mostly DRAM); SMIC and Huawei Ascend appear as the self-sufficiency proof points (Kirin 9000S / N+2 at 7nm) per Gregory Allen's export-control analysis. ACLS is flagged by Fabricated Knowledge as "possibly the best funding short in semi-cap forever" — mainly Chinese revenue that will be replaced, near-zero advanced-logic revenue, overstated backlog.

Intel 18A / EMIB supply chain cluster

The full idea map is Intel 18A / EMIB Supply Chain and the briefing is Intel 18A and EMIB Supply Chain Briefing (swarm: 12 tickers × 4 skills + 3 sub-theme compares). The thesis: 18A + EMIB ramp is a supply-chain catalyst across burn-in, substrates, metrology, semicap subsystems, analytics, OSAT, foundry, and AI-server power — happening simultaneously with the AI buildout, so many names get two demand tailwinds. The sharpest finding: the best investments are NOT the obvious Intel pure-plays but durable businesses where 18A is one of several tailwinds, screened hard on management alignment because every name had run 40-420%.

Three head-to-head compares back this: awx-vs-s71-showdown, ats-vs-onto-vs-uctt-and-more-showdown, amkr-vs-umc-vs-6809-and-more-showdown (all in ~/claude/output/compare/).

Cross-swarm conviction ranking:

Rank	Ticker	Sub-theme	Action	Mgmt	Key Catalyst
1	ONTO	Packaging (metrology/inspection)	SCALE BUY $290–300	B+	Q1 earnings May 2026; Semilab + SK Hynix $240M agreement underwrites 46x P/E; net cash $640M
2	ATS (ATS.VI)	Packaging (ABF substrates)	SCALE BUY ≤€85	B/Yellow	ABF shortage widening to 42% by 2028; Kulim FCF inflection; PEG 0.2–0.3x; anchor is AMD not Intel
3	PDFS	Fab analytics software	BUY on May 7	A-	94% recurring, 76% GM; Kibarian $118M stake, zero grants since 2003
4	2308 (2308/2308	2308.TW Delta Taiwan)	AI server power	SCALE post-Apr 30	B+/Green
5	UMC	Mature foundry	STAGE $11.00–$11.30	B+	H2 ASP +10% unpriced; Intel 12nm JDA 2027 optionality

Remaining cluster names: AWX (AEM Holdings, SGX — system-level burn-in, Intel-qualified HDMT; 49x fwd P/E), S71 (Sunright SGX — burn-in + KESM test, SOTP implies operating business at –S$53M), UCTT (Ultra Clean Holdings — semicap subsystems, 35-40 cents incremental GM per WFE dollar, 280% YTD is 65% multiple-driven, securities-fraud class-action tail), 8027 (E&R Engineering, TWO — EMIB tooling, speculative, Intel link indirect, chairman 47.9% share pledge), DELTA (Delta Electronics Thailand, SET — AI-server PSUs, exceptional business but wrong vehicle: CEO holds zero personal DET shares, governance discount), AMKR (Amkor — OSAT / Intel Foveros packaging, Kim family 49.5% stake, but 52x P/E and all 8 analyst targets below price). Preferred combos: ATS.VI + PDFS + ONTO (packaging layers, minimal overlap) and 2308.TW + UMC + 6809.HK (server/foundry). Universal risk: every name at/near 52-week highs as of Apr 26 2026; the Apr 27–30 earnings cluster is the price-discovery event.

Photonics / CPO bottleneck chain (Serenity's hunting ground)

Serenity (@aleabitoreddit, serenity-method, source profile) traces hyperscaler capex upstream until she hits a $100M–$700M-MC bottleneck the market hasn't priced — "the more well known the name ($LITE, $COHR) the more crowded/priced-in; go upstream." Her highest-conviction campaign is SIVE (Sivers Semiconductors): laser supplier for JBL/MRVL/Ayar/Amazon-Celestial, $520M → $2B+ MC target if rerated as a US company, AVGO/MRVL acquisition target (chokepoint M&A logic), EU listing events 5/11 and 5/13 2026. Her thesis-portfolio screenshot ran +1,116% YTD across HPS.A, SIQE, AXTI, SIVE, AAOI, LITE, NBIS, AEHR, TSEM, SOI. Other Serenity names: Shunsin (6451 TWSE, Foxconn ODM for NVDA CPO, ~$1.4B MC), IQE plc (IQIE, $130M MC Taiwan-sale asymmetry), AXTI (substrate), MEC (4971 TYO — molecular-bonding chemistry for FC-BGA substrates, mid-2026 glass-substrate catalyst).

The InP / merchant-optical supply-chain map she uses: DGC (phosphate mine) → NCI (refining) → Sumitomo / JX / AXT (melt) → COHR / LITE (fab) → Innolight / Fabrinet (packaging) → NVDA / GOOGL. "West depends on NCI." The MLCC / PCB-drill chain: Murata (6981), Taiyo Yuden (6976), TDK (6762), Kyocera (6971), Yageo (2327), Walsin (2492), VSH; raw material Sakai Chemical (4078, Barium Titanate); PCB drill Topoint (8021), Union Tool (6278, Pink holds), Kyocera. Her "Made in America" supply premium: AAOI transceivers price above FN (Thailand) / Innolight (China) / Eptolink (China) for geopolitical resilience — analog of INTC-to-TSM, MU-to-Hynix; matters more as export controls tighten.

The Jason's Chips photonics→analog mapping (a pricing-power crib sheet linking each photonics name to its logic equivalent) is reproduced because it ties the photonics multi-baggers to the AI-infra coverage:

Photonics name	Analog (logic equivalent)
TSEM	TSM
LITE	NVDA
COHR	AMD
BESI	AMAT
XFAB	INTC
AAOI	MU
POET	Groq
ALMU	Cerebras
AXTI	SNDK
CIEN	CRDO
SMTC	ALAB
SIVE	LITE

Analog / automotive / RFFE cluster (Doug O'Laughlin / Fabricated Knowledge)

The screenshot archive preserves Fabricated Knowledge's recurring automotive-semi bottom call. The Jan 2022 "Valuations of the Space" list of positive-EPS-growth / negative-price names (the bargain-hunting ground): SITM, ON, MXL, AMKR, AMD, DIOD, LRCX, BESI, NXPI, STM, MTSI, ALGM, MPWR, IFX, MU, QCOM, ADI, MKSI, TXN, POWI, CRUS, WOLF, ENPH, SWKS, QRVO, FSLR, SEDG, OLED. Doug's call: "Automotive, RFFE, and solar is the place to play" — ON, NXPI, STM, ALGM, IFX, ADI, TXN worth a second look. The Mar 2025 bottom call: cheapest names Renesas, ON, STM (low quality of earnings), and IFX; ALGM more expensive because mid-deal. "In the downturn, it never seems like it can be better. Yet I think it will in automotive."

ALGM (Allegro MicroSystems) — magnetic-sensor and power-ASIC IP leader, ~70% automotive revenue, 3+ year design cycles giving 7+ years of revenue visibility; automotive content growing 7-9% CAGR with ALGM at 10-15%. Watching. (Pink held — IBKR snapshot 100 shares ~$25, fwd P/E ~42.) IFX (Infineon) — world's #1 power-semi company by revenue, genuinely diversified across SiC, GaN, IGBT and silicon power; PSS segment ~30% of revenue. Has the Navitas GaN cross-license but Irrational Analysis (Irrational Analysis source) flags the integrated Navitas parts as a performance laggard ("meaningfully worse performance") — Infineon plays the high-V GaN market as the inferior product, capturing the rest while customers pay up for NVTS/TXN. Verdict: diversified power-semi proxy, not the high-conviction GaN vehicle; fully valued at ~27x fwd. The legacy IGBT franchise (~$3B revenue, high margin) is what makes it investable as a proxy.

WOLF (Wolfspeed) carries a going-concern warning per the Nov 2024 read: $5B debt, $1.2B market cap, ~$1B/yr cash burn — "will go bankrupt by the end of 2025 if it doesn't raise cash soon" (plus the question of why 200mm SiC is needed at all).

Source desk

The named sentiment / supply-chain analysts that feed these player views: Doug O'Laughlin / Fabricated Knowledge (_sources/source-doug-olaughlin, paid, the canonical semi Substack — AMAT, GFS, MU, AOSL short); Serenity / Aleabit (_sources/source-aleabit, free X, photonics/CPO bottleneck hunter — SIVE); Daniel Nystedt (_sources/source-dnystedt, free X, Asian-tech breaking news, Taiwan/Korea/China primary-language reporting — vault entry currently empty); Irrational Analysis (_sources/source-insane-analyst, paid, anonymous-engineer semi deep dives — IFX/Navitas performance read); Collyer Bridge Discord (Rigaku/JEM/probe-card cycle); Caughran / FundamentEdge (fundamentedge-method, methodology not picks). Per Serenity's own audit, SemiAnalysis, Global Tech Research, and Fabricated Knowledge were the only paid newsletters in a 23-author experiment to beat reasonable benchmarks.

Monitor

This is the living watch-log for the semiconductors / geopolitics sector — a rolling, dated record of catalysts, policy moves, earnings signals, and standing watch-items consolidated from the vault's screenshot sweeps, swarm briefings, Discord/Substack scans, and source-tracking notes. Read it chronologically for "what happened and when," and use the standing checklist at the bottom for "what to keep an eye on." Company-specific detail lives on the ticker pages; this page keeps the dated thread.

Standing watch-items (the recurring checklist)

These are the things to check every sweep, regardless of headlines:

• Intel Foundry Services external customer win — any IFS major external win re-rates the entire EMIB-linked supply chain (ONTO, ATS.VI, PDFS, AMKR, 8027.TWO) broadly. This is the single biggest binary catalyst for the Intel-18A cluster. Watch alongside any delay disclosure on Intel Panther Lake (18A) production timeline.
• ABF substrate spot price — leading indicator for ATS volume and pricing; the shortage is modeled to widen to 42% by 2028.
• Probe-card / burn-in lead times — AI chip testing probe card lead times ran 6 months to 16-24 months through early 2026, some orders at 2-year visibility (MPI Corp 6223 TWSE, Union Tool 6278). Lengthening lead times confirm the up-phase of the cycle that 6855 (JEM) sits in.
• China WFE memory capacity expansion — combined China memory capacity expansion of c.120K wpm+ projected in 2026, most of it DRAM. Watch domestic-equipment share gains (ACMR targeting 60% long-term share in China cleaning equipment).
• Glass-substrate transition — the shift to glass substrates at 1.4nm / 1.6T speeds is the structural catalyst flagged for mid-2026 that makes traditional organic-substrate roughening obsolete (MEC 4971 TYO thesis). Watch substrate-maker disclosures (SEMCO, Ibiden, Shinko).
• CPO (co-packaged optics) timing — repeatedly flagged as "happening earlier than expected." NVDA's March 2026 $2B-each commitments into COHR / LITE / MRVL is the same capacity-lock playbook as NVDA's 2024 EML preallocation that drove LITE +955%. Watch CW/EML laser and CPO bottleneck names.
• China mineral retaliation escalation — China has weaponized germanium/gallium (2023) and antimony (2024) export controls as retaliation for chip restrictions. The open watch-item is whether China extends to rare-earth magnet exports (critical for defense, EVs, wind).
• Intel 18A as the CHIPS Act acid test — if 18A process technology fails, the US reshoring thesis takes a major hit. Track INTC 18A yield ramp.
• Section 232 semiconductor investigation — Commerce Department national-security investigation that could produce semiconductor-specific tariffs; status and findings pending.

2022 — the trailing-edge thesis takes shape

Jan 2022 (Fabricated Knowledge, "Valuations of the Space," Jan 9): Most prior-cycle winners were driven by multiple expansion, not earnings growth; companies with positive EPS growth AND negative price return were called out as the bargain-hunting ground. Doug's call: "Automotive, RFFE, and solar is the place to play" — flagging ON, NXPI, STM, ALGM, IFX, ADI, TXN.

Jan 25, 2022 (Discord, Doug/Mule): "I am now a 11% for KLAC/AMAT/LRCX + 5/5 ONTO/NVMI + just upped my NXPI/ON." "I have pretty much sold everything off to btfd in semis." Watching WOLF chart.

Feb 1-3, 2022 (Discord, Doug/Mule): NXPI active deal rumors (tend to announce on earnings); STM price raises "nuts" given supply-chain cost absorption in CoGs. TMT rotation call: "semis relative to internet within TMT has never been better." Bought Feb 18 calls on MXL. Earnings-season interest: ACLS, AOSL, ON, ICHR, ONTO. ON Semi reported Q4 EPS $1.09 ex-items.

Mar 2022 (Fabricated Knowledge + Discord #semicap, Mar 19) — the long-tail-of-smaller-fabs thesis: Fastest capex growth is NOT TSMC's ~40% YoY, but the long tail of IDMs and smaller fabs. UMC and STM both guided ~100% YoY capex increases; NXP ~70%; Powerchip and TXN likely faster than 40%. "Lagging edge semiconductor is finally getting the investment it so desperately needs." ACLS flagged as the pure trailing-edge play (80% of business is mature process).

2023 — valuation normalization

Jan 3, 2023 (Substack) — year-end valuation review: Almost every multiple compressed, but much less than stocks actually fell. EPS estimates INCREASED for ASML, TSM, ADI, TXN. Semicap earnings fell less than expected (ex-LRCX). INTC and NVDA estimates plunged. Upper-end valuations still hefty: MPWR, ASML, NVDA, LSCC, AMBA, WOLF. Overall sheet "feels much less insane than a year ago."

Aug 12, 2023 (master-faq): MOSFET/IGBT glut watch — "Chinese foundry and IDMs are expanding capacities rapidly. While BYD continues to gain share in NEV and increases internal sourcing ratio." Standing question logged: "Lagging edge war in the next 5 years."

2024 — subsystems and the AI-content rotation

May 19, 2024 (Fabricated Knowledge) — earnings/subsystems: ONTO, CAMT, AMAT delivered as expected; MKSI did a questionable convertible offering that tanked shares. "Subsystems revenue should continue improving into the second half and that 2025 will be a strong year. All revenue related to HBM continues to shine." "I continue to love the Subsystems trade, and I think that orders will pull in the second half of 2025."

Nov 12, 2024 (Fabricated Knowledge) — quick takes on the semi universe:

• ENTG: memory spending next year, but MSI weak.
• ALAB: second way to play a levered Trainium ramp; extensive Amazon content.
• LSCC: standard kitchen sink; like long-term.
• FN/COHR/LITE/AAOI: incremental telecom strength from ZR and DCI; AAOI talking DOCSIS 4.0. Everything ZR-related hot; positive read-throughs to CIEN and MRVL.
• SITM: transceiver and likely GB200 design wins; expect large beats next year on transceiver revenue.
• QCOM: strong results but most uncertainty (ARM/Apple overhang, potential Intel bid, China overstocking).
• MCHP: "It's over, but hopefully will be back" — Europe drives industrial weakness; stockpiling tools because of CHIPS ITC.
• WOLF: going-concern risk flagged — $5B debt, $1.2B market cap, $1B/yr cash burn. "It will go bankrupt by the end of 2025 if it doesn't raise cash soon."
• ACLS: "Possibly the best funding short in semi-cap forever" — mainly Chinese revenue that will be replaced, near-zero advanced-logic revenue, overstated backlog.
• ARM: doubling+ ASP for CSS products; taking price.
• ANET: doubling TAM because of AI back-end network; sandbag guide as usual.

2025 — automotive bottom call and the supercycle re-rate

Mar 18, 2025 (Substack) — automotive semi bottom call: Cheapest names Renesas, ON, STM (low earnings quality), IFX; ALGM more expensive due to pending deal. Broad bottom call: "In the downturn, it never seems like it can be better. Yet I think it will in automotive." If inventory turn happens, expect double-digit revenue and EPS for 2-3 years.

Jul 4, 2025 — portfolio snapshot (IBKR): NLV $41,405, daily P&L +$414 (+1.01%). Big winners by unrealized: LRCX +39.0%, SKYT +26.7%, IREN +25.9%, PBI +22.1%, TPR +20.9%, AIXA +19.3%, BWXT +18.3%. (Full table on screenshot-insights-archive.)

Jan 2026 — the Intel options post-mortem

Jan 22-23, 2026 (INTC tradelog): Psychological-discipline lesson logged. INTC calls went up 60-70% (potential gains over $150k) then fell to ~10% — held through, then "started gambling by weekly intc 47c" which went down. Root cause: greed + complacency + overtrading into the night reading Discord. Standing framing: "Intel is in the execution phase which will be a long-term game to prove." Core learning: "being able to differentiate between hype and fundamentals to have perspectives on what ST, MT and LT" (see JStore Discord bb chat).

Feb 2026 — Mule's pitch baskets and the gold thesis

Feb 25, 2026 (Mule / @FoolAllTheTime) — "shit bags" basket pitch (cheap/hated names with structural rather than discretionary bull cases):

1. MPTI — makes money from missiles, backlog up 60% before Iran (held)
2. VG — TTF structurally higher (European nat gas; held)
3. KOS — won't go bankrupt because of Brent (held)
4. SM — same quality setup, SM Energy (held)
5. BWXT — nuclear will accelerate (held)
6. CHG / Chemring — "shittier cheaper MPTI" (UK defense) All already in Pink's portfolio — validation, not new ideas.

Feb 25, 2026 (Mule) — gold supply-chain thesis (max-torqued size): "Gold prices could halve and it would still have accelerating demand for the next 5 years." Bull case holds even at $2,500 gold; "2500 enough for these to work; 4k everyone prints money." Names: LYL (Australian EPC, ~8x EBITDA, "cheap because of Africa"), MDI, EHL, ARIS. Easier baskets: PICK ETF or Weir (London). Note the tension with Doug's "no-Africa-assets" Canadian-miner filter — opposite angle (resilience vs value) on the same setup.

Feb 26, 2026 (Fat Tail Capital repost): SNOW CEO "We're seeing this transformation within our customers. They are leveraging agents not just to analyze..." — agentic-AI tailwind for data/analytics platforms.

Mar 2026 — photonics→analog mapping and the CPO playbook

Mar 12, 2026 (Substack, "Jason's Chips") — photonics-to-analog mapping crib sheet:

Photonics name	Analog (logic equivalent)
TSEM	TSM
LITE	NVDA
COHR	AMD
BESI	AMAT
XFAB	INTC
AAOI	MU
POET	Groq
ALMU	Cerebras
AXTI	SNDK
CIEN	CRDO
SMTC	ALAB
SIVE	LITE

The left column (TSEM, COHR, BESI, AAOI, CIEN, SMTC) is where the photonics multi-bagger cases sit; ties directly to the Soitec substrate thesis.

Mar 2026 (Serenity playbook recognition): NVDA's March 2026 $2B-each into COHR / LITE / MRVL recognized as the same capacity-lock playbook as NVDA's 2024 EML preallocation (which drove LITE +955%) — this time CW/EML lasers and CPO bottlenecks. Watch for the downstream effect to repeat.

Late Mar – Apr 2026 — Collyer Bridge Discord and the screenshot sweeps

Mar 30-31, 2026 (Collyer Bridge Discord) — 6855 JEM state of play: JEM + Seikoh Giken + Innotech + GL Techno all ex-div (factor in the drop). CB's top list held: Metasurface, 268A Rigaku, 2337 Macronix, 6855 JEM. CB: "Feb results confirm they're totally in the HBM game, why is it not compelling at these levels?" Sentiment for held 6855: cautiously constructive at lower levels.

Apr 1, 2026 (Collyer Bridge Discord) — 268A Rigaku thesis extension: "2nm or SiPho ramp would see Samsung place major orders for Rigaku kit." Co-development with ONTO referenced in earnings slides. Rigaku hybrid metrology engine: T-SAXS combining X-ray + optical, mass production planned, $100M new market with >50% share, JEP (joint eval) with major memory customers. Debate 3D DRAM vs 3D NAND; lean 3D DRAM given Onto involvement. Material catalyst for held 268A. (Cross-watch: explicit ONTO disclosure on next earnings call.)

Apr 4, 2026 (Collyer Bridge Discord) — FPS dip: FPS <$29 after secondary offering at $29.50. CB consensus bullish (Eric buying size under $30, "phenomenal company and mgmt"; Dropboks "early Vertiv vibes"). Consider add under $29.

Apr (Collyer Bridge Discord) — MPI Corp (6223 TWSE): AI chip testing probe-card lead times at 6 months, some orders 2-year visibility. Targeting double-digit annual revenue growth, new record full year. Acquired land in Zhubei; H2 2026 ramp of vertical + MEMS probe cards, capacity +50%. Complementary to held 6855 in the same probe-card cycle.

Apr (Collyer Bridge Discord) — ACMR / China WFE memory capex: ACMR 60% long-term market-share target in China cleaning equipment; SPM cleaning equipment meaningful progress at a key domestic customer. Combined China memory capacity expansion c.120K wpm+ in 2026, most DRAM; domestic equipment cos expect significant growth.

Apr 5, 2026 weekly screenshot sweep (window 2026-03-29 → 2026-04-05), processed Apr 7 — per-ticker monitor items:

• SOITEC — Asymmetrical Bets thesis (Substack Feb 25): SiPho moat, 95%+ Photonics-SOI share, CPO requires 4x SiPho wafer content vs pluggables, GlobalWafers exiting. Reinforces WATCH, validates pullback entry.
• 268A — Rigaku/Samsung order catalyst (above); Onto co-development on hybrid X-ray/optical metrology; likely 3D DRAM. Material catalyst.
• 6855 — probe-card lead times 16-24 months (Taiwan M-H Corp report); 6278 adding 50% capacity at Zhubei H2 2026.
• ONTO — catalyst added: Rigaku co-development on hybrid X-ray/optical metrology engine; watch for explicit ONTO disclosure on next earnings call.
• MPTI — position drawdown (cost $1.82, low $1.11 = -45%, then bounce to $2.80). Hold through volatility, don't stop out.
• FPS — Discord dip-buy <$30 (sub-secondary).

Apr 5-19, 2026 sweep (103 screenshots, triaged Apr 20) — Serenity (@aleabitoreddit) campaign and others:

• SIVE (Sivers Semiconductors) — central conviction: $520M → $2B+ MC target if rerated as US company (SIVE at 2.232 EUR; also FWB:2DG). Laser supplier for JBL/MRVL/Ayar/Amazon-Celestial. Acquisition target for AVGO or MRVL (AVGO already staked Win Semi). Fear priced into $350-400M MC. European listing events on 5/11 and 5/13 2026. Pink opened IBKR EU trading permissions — set up to buy. Recommend /deep-dive SIVE before any entry given the 5/11-5/13 events.
• CPO happening earlier than expected — Serenity adding to NVDA CPO Asia ecosystem "next week," planning a CPO ETF for subscribers. Timing catalyst flag.
• Taiwan CPO ecosystem — overlooked names: Shunsin (6451 TWSE) Foxconn ODM for NVDA CPO, ~$1.4B MC, capacity can't keep up; PCL (Broadcom-related), Fittech (Browave/Foci), Enplas (Google TPU/NVDA/TSMC).
• InP substrate supply-chain map (merchant optical): DGC (phosphate mine) → NCI (refining) → Sumitomo/JX/AXT (melt) → COHR/LITE (fab) → Innolight/Fabrinet (packaging) → NVDA/GOOGL. "West depends on NCI."
• TSEM — Serenity's home run: +759% YTD from a 3-week discovery post (rallied 90% in 3 weeks post-research). Recurring caution: "Apollo bought out the supplier" — watch for buyouts killing open-market plays.
• MLCC / PCB-drill supply-chain map (Apr 18): MLCC — Murata (6981), Taiyo Yuden (6976), TDK (6762), Kyocera (6971), Yageo (2327 TWSE), Walsin (2492 TWSE), VSH (US); raw material — Sakai Chemical (4078, Barium Titanate); PCB drill — Topoint (8021), Union Tool (6278 — Pink holds), Kyocera.
• Serenity thesis-portfolio screenshot: 11 tickers, YTD +1,116% — HPS.A, SIQE, AXTI, SIVE, AAOI, LITE, NBIS, AEHR, TSEM, SOI, +1.
• Pump-and-dump self-audit: Serenity listed 16 WSB-banned low-cap meme tickers with triple-digit YTD, some overlapping her own names — flag when a conviction-long starts to smell like a pump.

Apr 2026 — other dated items from the sweep:

• MEC (4971 TYO) (Paradis Labs Substack, 4 screenshots): ~$1B MC, sole provider of molecular bonding chemistry for FC-BGA substrates. Phase-2 margin guidance 20% → 30%. Glass-substrate shift at 1.4nm / 1.6T speeds makes traditional roughening obsolete mid-2026 — imminent mid-2026 catalyst. AI-grade substrate demand +50% vs 50% gap. Not held.
• Hammond Power (HPS.A / HMDPF) (Eric Jhonsa, X): 17x 2027E GAAP EPS, Q1 sales +13.8% YoY, backlog supports 100%+ growth. Held — bullish reinforcement.
• PAC-3 defense stockpile rebuild (Mike @BlackSchol): PKE sole supplier of PAC-3 ablative composites. Counter-drone suppliers LASR, EOS.AX, DRS, AVAV, LPTH, ESP. BKSY high-revisit tactical ISR. SYPR $9 target.
• Market sentiment: ISEE call/put ratio signals washed retail; Allbirds rebranding to "Newbird AI" with $50M convertible for GPU acquisition; Cerebras/AWS threat to NVDA but NIXL/EFA software moat intact; Kioxia CY2027 EPS ¥11,880 = 2.3x fwd P/E, Bain + Toshiba reducing stake.
• Broker artifacts (7-Apr accumulation list): SMTC 75+, AEHR $44-, INTC/AMD agentic-CPU, HPS.A/HMDPF power-AI industrial, MPTI defense, LPTH long-term. Active buy order AEHR @ $44 (then $51.32); active buy MPTI RT @ $1.60. Portfolio milestones +40% / +55% YTD, +218% 1Y, +250% since inception. NVDA Jun18'26 $200 call: -64% unrealized.

Apr 26, 2026 — Intel 18A / EMIB supply-chain swarm (12 tickers × 4 skills)

Full idea map at intel-supply-chain. The core monitor takeaway: every name in this cluster was at or near a 52-week high as of April 26, and the earnings cluster Apr 27–30 was the natural price-discovery event — no new money before those prints. Three risk-adjusted standouts: ONTO (net cash $640M underwriting a 46x multiple), ATS (best PEG at 0.2–0.3x, ABF shortage structural), PDFS (A- management, 94% recurring revenue).

Thesis corrections logged: ATS.VI anchor is AMD (not Intel) at Kulim — Intel link pilot-stage only; 8027.TWO TGV customer is Japanese; 6809.HK (Montage Jintide) is an Intel China-strategy play, not an 18A/EMIB supplier.

Intel milestones to watch: Panther Lake (18A) — first internal 18A product, sampling 2025, production 2026; Clearwater Forest — server chip on 18A, volume 2027; EMIB ramp tied to AI accelerator demand; IFS external customer wins (the critical revenue unlock).

Earnings calendar — the Apr 27–Jun 5 cluster (from the swarm):

Date	Ticker	What to watch	Entry trigger
Apr 27	AMKR	Revenue >$1.7B; GM >13.5%; Q2 guide ≥$1.75B	Clean entry $55–70 if disappoints
Apr 28	UCTT	H2 2026 demand step-up commentary	Build to 3–4% on confirmation
Apr 28	DELTA.BK	AI server PSU order backlog	No — wait for THB 250–270
Apr 28	6809.HK	Q1 2026 results; MRDIMM Gen2 ramp	Beat → HK$250–265; miss → HK$175–190
Apr 29	UMC	GM >28%; Singapore Phase 3 timeline; ASP hike	Confirmation → stay in $11–$11.30
Apr 30	2308.TW	Revenue >TWD 110B; Vera Rubin backlog	Beat → initiate TWD 1,700–1,800
May 7	PDFS	Revenue ≥$50M, ≥18% YoY, margins stable	Yes — primary buy trigger
May 11	AWX.SI	New customer identity hints	Beat → consider S$5.00–5.50
May 11	8027.TWO	Q1 revenue >TWD 500M	Yes — required for add
Jun 5	S71/S71	S71.SI	2H FY2026 burn-in services recovery

Verified as of the swarm: ATS.VI Kulim FCF inflection confirmed (operating cash +€331.8M YTD vs –€29.4M prior year); UMC H2 2026 ASP hike ~10% confirmed April 17, not yet in sell-side consensus; AMKR Kim family has NOT diversified (49.5%, $10B+ stake); ONTO $640M net cash + $240M SK Hynix supply agreement through 2027; PDFS Kibarian bought $1.13M open-market at the 52-week low in February 2025.

Open / unverified: AWX.SI's unnamed new customer; ATS.VI Intel EMIB volume order (pilot-stage only); 6809.HK MRDIMM Gen2 ramp timeline/volume; UCTT securities-fraud class-action resolution; 8027.TWO whether Q4 2025 TWD 721M revenue was a one-time AIS recognition or run-rate.

Standing structural watch (longer-horizon items)

• Reshoring at scale — Barclays' manufacturing-footprint read: reshoring works for leading-edge (<7nm) but mature nodes (28nm+) stay in Asia; the CHIPS Act is subsidizing what's economically marginal. Watch whether it sticks when subsidies expire.
• Export-control escalation — how far US chip restrictions go before serious Chinese retaliation beyond minerals (the rare-earth-magnet trigger).
• Japan friend-shoring durability — is the friend-shoring premium permanent or transient? Each JP semi name has different China revenue exposure (japan-semi, jp-semi-supplychain-comparison).
• Nitto Boseki (3110) monopoly — order books full through 2026 (NVIDIA, AMD, Apple, Google, Amazon customers); the entire re-rate hinges on one product line (specialty glass cloth at ~90%+ share). Any AI-capex slowdown causes severe multiple compression; analyst consensus PT ~15,380 sat ~40% below the ~25,470 spot price.
• Qualification-cycle signals — per the Serenity method, watch for qualification orders from hyperscaler supply chains (e.g., AEHR getting qualification orders from a leading optical transceiver company = "early AAOI stage"). The qualification phase, not the volume ramp, is the entry point.

Source-tracking notes (who feeds this monitor)

• Fabricated Knowledge / Doug O'Laughlin (source-doug-olaughlin) — paid; semis core, market structure. X @fabknowledge (pro), @FoolAllTheTime (personal). The recurring source for valuation reviews, subsystems calls, and the trailing-edge thesis above.
• Serenity / Aleabit (@aleabitoreddit) (source-aleabit, serenity-method) — free, X-only; AI/semi supply-chain bottleneck hunting, CPO/SiPh focus. Pink has a premium subscription. Drives the SIVE / TSEM / CPO-timing thread.
• Daniel Nystedt (@dnystedt) (source-dnystedt) — free, X-only; Asian tech/semis breaking news and primary-language Taiwan/Korea/China supply-chain reporting. (Notable-content section still empty — populate as posts are read.)
• Insane Analyst / Irrational Analysis (@insane_analyst) (source-insane-analyst) — paid; anonymous-engineer semi deep dives. (Notable-content section still empty.)
• Collyer Bridge Discord — the recurring Discord source for the JP/Taiwan small-cap probe-card and metrology threads (268A, 6855, 6223, ACMR).
• Methodology overlays: fundamentedge-method (Caughran — how to think: revenue primacy, second derivative, never short into positive revisions) and serenity-method (where to look: trace demand upstream to the small-cap bottleneck, second-order earnings read-throughs, 8-K supply-chain mapping). Both inform how monitor signals get triaged into action.

Sources

The semiconductor sector page draws on a mix of paid Substacks, free X/Twitter analysts, Discord trading communities, broker screenshots, sell-side desks, and standard financial-data aggregators. The authors below are the recurring voices behind the theses on this page. Where a person or publication has a dedicated page under _sources/, it is wikilinked; the rest are listed inline because no source page exists yet.

Named authors and publications (with source pages)

• Aleabit / Serenity (@aleabitoreddit) — AI/semi supply-chain analyst, ex-RISC-V FDN, ex-AI research scientist; Reddit WSB legend now on X since Jul 2025 (206K followers, 5,735 tweets in ~10 months). Free tier (X-only as far as the vault knows), but Pink holds a premium subscription. The methodology engine behind the bottleneck-hunting / CPO / photonics theses. Full digest at serenity-method. Notable: the SIVE (Sivers Semiconductors) campaign (JBL/MRVL/Ayar/Amazon-Celestial laser supplier, AVGO/MRVL acquisition target, events 2026-05-11 and 2026-05-13), TSEM "+759% YTD" glass-substrate call, InP substrate supply chain map (DGC → NCI → Sumitomo/JX/AXT → COHR/LITE → Innolight/Fabrinet → NVDA/GOOGL), Shunsin (6451 TWSE) Foxconn CPO ODM call, IQE plc (IQIE), and the MLCC/PCB-drill supply-chain roll-call.

• Doug O'Laughlin / Fabricated Knowledge — semis core, market structure, technical/market commentary; one of the canonical paid Substacks for semiconductor analysis. X pro @fabknowledge, X personal @FoolAllTheTime ("Mule"), Substack fabricatedknowledge.com. Paid tier (★★ active). The dominant voice in the screenshot archive: "Valuations of the Space" (Jan 2022), "The Long Tail of Smaller Fabs" / trailing-edge thesis (Mar 2022), year-end valuation reviews, subsystems/semicap earnings takes, the Nov 2024 semi-universe quick takes (WOLF going-concern call, ACLS "best funding short in semicap"), the Mar 2025 automotive-semi bottom call, plus his personal-account "shit bags" defense/energy basket and gold-supply-chain pitch (Feb 2026). Connects to AMAT, GFS, MU, AOSL.

• Insane Analyst / Irrational Analysis (@insane_analyst) — semi industry analysis from an anonymous engineer; general investment commentary on X, semi-focused deep dives on Substack (irrationalanalysis.substack.com). Paid tier (★★ active). Source page is a stub (notable content not yet populated).

• Daniel Nystedt (@dnystedt) — Asian tech/semis reporting; primary-language breaking news on Taiwan / Korea / China supply chains. Free tier (X-only). Source page is a stub (notable content not yet populated).

Named authors and analysts cited but without a source page

• Brett Caughran / FundamentEdge (@FundamentEdge) — ex-Maverick, D.E. Shaw, Citadel, Schonfeld (13-yr PM); runs Fundamental Edge analyst-training firm. The methodology counterpart to Serenity (Caughran = how to think; Serenity = where to look). Digest at fundamentedge-method — 742 tweets + 18 threads. No _sources/ page yet.
• Collyer Bridge (Discord) — the most-cited Discord community on this page; source of the Rigaku (268A) hybrid-metrology / T-SAXS thesis, JEM (6855) HBM-game state-of-play, FPS dip-buy consensus, MPI Corp (6223 TWSE) probe-card lead-time intel, and ACMR / China WFE memory-capex notes. Named members referenced: Bob, Zack, Eric, Dropboks/Dropbks. No source page.
• Eric Jhonsa (X) — Hammond Power (HPS.A / HMDPF) call and ISEE call/put sentiment read. No source page.
• Mike (@BlackSchol) (X) — PAC-3 defense-stockpile rebuild thesis (PKE as sole supplier of PAC-3 ablative composites). No source page.
• Paradis Labs (Substack) — MEC (4971 TYO) deep dive: sole provider of molecular-bonding chemistry for FC-BGA substrates, mid-2026 glass-substrate catalyst. No source page.
• "Jason's Chips" (Substack, Mar 12) — the photonics-name → analog-logic-equivalent mapping table (TSEM:TSM, LITE:NVDA, COHR:AMD, etc.). No source page.
• Asymmetrical Bets (Substack, Feb 25) — Soitec SiPho-moat thesis (95%+ Photonics-SOI share, CPO needs 4x SiPho wafer content). No source page.
• Fat Tail Capital (X repost, Feb 26) — SNOW agentic-AI customer-transformation quote. No source page.
• Global Tech Research (Substack) — Nitto Boseki specialty-glass analysis; also one of the three paid newsletters Serenity flags as beating benchmarks. globaltechresearch.substack.com. No source page.
• SemiAnalysis / Dylan Patel — referenced in master-faq (packaging series, HBM/hybrid-bonding) and sector-data-research (AI cost/server analysis, "Gemini Smashes GPT-4 By 5X, The GPU-Poors"); cited by Serenity as one of the few paid newsletters that beat benchmarks. No source page.
• JStore (Discord "bb chat") — cited in the INTC stocklog as the Intel-execution-phase perspective source. No source page.
• Gregory Allen — export-control effectiveness analyst (CSIS); his "true impact of export controls" and US-China AI-race assessments anchor the supply-chain-security hub (Huawei Kirin 9000S / SMIC 7nm, DeepSeek). Referenced via allen-true-impact-export-controls and allen-ai-race-2025. No _sources/ page.

Sell-side, consultancy, and institutional sources

• Barclays — global semiconductor manufacturing-footprint analysis (src-barclays-semi-manufacturing-footprint); leading-edge reshoring works, mature nodes stay in Asia.
• Oliver Wyman — US tariffs on Asia, supply-chain diversion projections (oliver-wyman-us-tariffs-asia).
• Apollo — US consumer tariff impact, pricing pass-through (apollo-us-consumer-tariff-2025).
• Hyundai — Japan/Taiwan field trip (TSMC expansion, HBM localization, China strategy) from a Korean perspective (Hyundai-Field-Trip-Japan-Taiwan-Companies).
• Kerrisdale Capital — Aixtron long thesis (kerrisdale-aixtron-long-thesis).
• Evercore ISI — VRT FCF analysis (cited in screenshot archive).
• Warren Buffett — trade-deficit / tariff macro framing (warren-buffett-trade-deficit).
• Dong/Shim — the Rigaku (268A) "Lab to Fab" thesis (true semi exposure 60%+ of profits vs reported 30% of revenue), via 268a-dong-thesis-analysis.
• BCG — SEA datacenter buildout (bcg-sea-datacenters).
• Citrini Research — "War From Home" defense primer (src-citrini-defense-primer-wfh).
• UTR — defense sector risk / ballistic-missile-defense / drone-warfare / quantum-for-defense primers.

Financial-data aggregators (the JP semi comparison)

The 6-company Japan supply-chain comparison (jp-semi-supplychain-comparison) cites: Yahoo Finance, MarketScreener, Stock Analysis (stockanalysis.com), Simply Wall St, TradingView, Kabutan (Japanese earnings guidance), CNBC, Mordor Intelligence (probe card / T&M market reports), Grand View Research (5G testing equipment), Rigaku Holdings IR, Furukawa Electric IR, plus Global Tech Research and Tom's Hardware for the Nitto Boseki / glass-cloth angle.

Channels and capture mechanisms

• Discord — Collyer Bridge, AI/Semiconductor watchlist channels, and "bb chat" (JStore on INTC). The screenshot-insights-archive and weekly screenshot-sweep files are largely Discord captures.
• X / Twitter — Serenity, Mule (Doug personal), Eric Jhonsa, Mike (@BlackSchol), Fat Tail Capital.
• Substack — Fabricated Knowledge, Irrational Analysis, Paradis Labs, Jason's Chips, Asymmetrical Bets, Global Tech Research, SemiAnalysis.
• Broker screenshots (IBKR) — Pink's own position and portfolio snapshots, processed via the osxphotos screenshot-sweep pipeline (1,128 screenshots in scope, 79 unique tickers).

Consolidated source files on this page

The Semiconductors sector page is assembled from these vault files (none are themselves "sources" in the author sense — they are the vault's own notes):

• china-semi.md — stub (header + thesis line only; "To be populated from source files").
• semi-universe.md — stub (header + thesis line only; "To be populated from source files").
• japan-semi.md — Japan friend-shoring topic hub (the 6-company thesis, cross-domain links).
• jp-semi-supplychain-comparison.md — the quantitative 6-company comp (6754 / 6777 / 6855 / 5801 / 268A / 3110); the comp tables and source list live here.
• export-controls-china-semi.md — stub framework ("Awaiting source material from Dropbox"); BIS, Chip 4, entity list, SMIC/DeepSeek case-studies pending.
• tariffs-trade-policy.md — stub framework ("Awaiting source material from Dropbox"); Section 232, Apollo/Buffett tariff analysis pending.
• supply-chain-security.md — the populated geopolitics topic hub (export controls, reshoring, friend-shoring, defense, critical minerals, SEA).
• master-faq.md — raw semi-terms FAQ / research-to-do list (Tokyo Electron, ASML, AMAT, KLAC, LRCX peer list; BSPDN; MOSFET/IGBT glut; SemiAnalysis packaging series).
• sector-data-research.md — raw finance-fundamentals and AI-capex working notes (Dylan/SemiAnalysis, 1999 fiber-spend vs 2024 AI-capex comparison, NVDA generational TCO note).
• screenshot-insights-archive.md — the master screenshot extraction (Fabricated Knowledge theses, Doug/Mule Discord posts, IBKR position screenshots, Collyer Bridge notes, Apr 5-19 sweep).
• screenshot-sweep-2026-04-05.md — weekly osxphotos sweep (per-ticker updates, photonics→analog mapping, Mule gold/defense baskets).
• serenity-method.md — Serenity methodology digest (the 10 moves; SIVE/CPO/TSEM/InP screenshot additions).
• fundamentedge-method.md — Caughran methodology digest (the 10 moves; revenue primacy, 2nd derivative, TECC, comp-sheet).
• _sources/source-aleabit.md, _sources/source-dnystedt.md, _sources/source-doug-olaughlin.md, _sources/source-insane-analyst.md — the four _sources/ identity pages.
• themes/intel-supply-chain.md — Intel 18A / EMIB swarm-research idea map (12 tickers × 4 skills).
• briefings/2026-04-26-intel-supply-chain.md — the stakeholder-ready handoff briefing for the Intel 18A/EMIB swarm.
• stocklog-2026-01-23-intc.md — INTC options trade-log / lessons note (cites JStore Discord bb chat).

All 21 listed source paths exist and were read in full.

---

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.

• [ ] china-semi.md
• [ ] semi-universe.md
• [ ] japan-semi.md
• [ ] jp-semi-supplychain-comparison.md
• [ ] export-controls-china-semi.md
• [ ] tariffs-trade-policy.md
• [ ] supply-chain-security.md
• [ ] master-faq.md
• [ ] sector-data-research.md
• [ ] screenshot-insights-archive.md
• [ ] screenshot-sweep-2026-04-05.md
• [ ] serenity-method.md
• [ ] fundamentedge-method.md
• [ ] _sources/source-aleabit.md
• [ ] _sources/source-dnystedt.md
• [ ] _sources/source-doug-olaughlin.md
• [ ] _sources/source-insane-analyst.md
• [ ] themes/intel-supply-chain.md
• [ ] briefings/2026-04-26-intel-supply-chain.md
• [ ] stocklog-2026-01-23-intc.md

---

Intel supply chain (folded 2026-06-01, from 2026-04-26-intel-supply-chain)

TLDR

Intel's 18A process node and EMIB (Embedded Multi-die Interconnect Bridge) advanced packaging ramp is a supply chain catalyst spanning burn-in test, IC substrates, optical metrology, semicap subsystems, manufacturing analytics, OSAT packaging, mature foundry, and AI server power infrastructure. Twelve companies were screened across all four layers. The key finding: the most compelling investments are NOT the obvious Intel pure-plays — they are the companies with durable business models that benefit from the 18A ramp as one of several demand tailwinds. Three names stand out on risk-adjusted conviction: ONTO (net cash $640M underwriting a 46x multiple), ATS.VI (best PEG in the group at 0.2–0.3x, ABF shortage structural), and PDFS (A- management, 94% recurring revenue, cleanest founder alignment in small-cap semis). Critical caveat: every name in this cluster is at or near a 52-week high as of April 26. Five companies report earnings April 27–30. No new money before those prints.

---

INVESTMENT THESIS — 3 BULLETS

• Dual tailwind: The 18A / EMIB ramp is happening simultaneously with AI infrastructure buildout. Companies supplying both intel's advanced packaging AND hyperscaler AI chip packaging get two demand curves instead of one — ONTO (SK Hynix $240M agreement), UCTT (WFE leverage), and ATS.VI (ABF substrate shortage widening to 42% by 2028) are the clearest examples.
• Thesis corrections embedded: Three of the twelve names have overstated Intel connections. ATS.VI's anchor is AMD (not Intel) at its Kulim campus. 8027.TWO's TGV customer is Japanese (not Intel directly). 6809.HK (Montage Jintide) is an Intel China strategy play, not an 18A/EMIB supplier. Pricing these correctly changes the risk profile.
• Management alignment as the differentiator: In a cluster where every stock has run 40–420% in recent months, management quality and insider alignment become the most important screen. PDFS (Kibarian $118M personal stake, zero equity grants since 2003) and 2308.TW (Ping Cheng's $3.6B net worth IS the stock) are the standouts. DELTA.BK fails this test — CEO holds zero personal DET shares.

---

THE PICKS

Sub-Theme A — Burn-in / Test

Rank	Ticker	Conviction	One-Line Thesis	Entry	Biggest Risk
1	S71.SI	Low (value)	Operating business implied at –S$53M on SOTP; buying for hard asset floor	SGD 0.48–0.52	Burn-in services recovery slower than expected
2	AWX.SI	Low-Med	Best Intel 18A burn-in pure-play; new unnamed customer is the upside option	S$4.50–5.00	Priced for bull case at 49x forward P/E; customer remains anonymous

Neither is a clean entry today. Both at RSI 81+. AWX May 11 earnings, S71 June 5.

---

Sub-Theme B — Advanced Packaging / EMIB Tools & Substrates

Rank	Ticker	Score	Conviction	One-Line Thesis	Entry	Biggest Risk
1	ATS.VI	3.80	Medium-High	Best PEG (0.2–0.3x); Kulim FCF inflection confirmed; ABF shortage structural	≤€85 starter	AMD concentration; dual-bloc governance conflict history
2	ONTO	3.78	Medium-High	Net cash $640M + SK Hynix $240M agreement underwrites 46x P/E	$290–300 starter	CEO 10b5-1 selling; EMIB-specific revenue not yet material
3	PDFS	3.70	Medium	94% recurring revenue, 76% gross margins, A- management; gap-up Apr 24 = FOMO trap	$41–43 pilot post-May 7 earnings	April 24 gap-up on no specific catalyst; multiple compression if growth disappoints
4	UCTT	2.74	Low-Med	35–40 cents incremental GM per WFE revenue dollar; revenue upside from available capacity	1% starter; Apr 28 earnings	280% YTD is 65% multiple-driven; securities fraud class action (prior CEO)
5	8027.TWO	2.29	Speculative	Option on EMIB-T ramp; priced at 9x EV/Sales on loss-making base	TWD 156 pilot only	Intel connection indirect; chairman 47.9% shares pledged; May 11 binary

Preferred combination: ATS.VI + PDFS + ONTO — three different supply chain layers, minimal correlation, complementary risk profiles.

---

Sub-Theme C — Foundry / OSAT / Server Components

Rank	Ticker	Score	Conviction	One-Line Thesis	Entry	Biggest Risk
1	6809.HK	3.90	Medium	Highest growth; MRDIMM Gen2 ramp undermodeled in consensus	HK$155–180 pullback	Export control binary (TSMC + Intel Jintide supply to China); founders' stake 60% → 4%
2	2308.TW	3.75	Medium-High	Best business quality; CEO alignment exceptional; Vera Rubin power cycle Q3 2026	TWD 1,700 post-Apr 30 earnings	89.9x TTM P/E; intrinsic value estimates 50% below spot
3	UMC	3.65	Medium	Only stock not obviously overpriced; H2 2026 ASP +10% hike confirmed but not in consensus	$11.00–$11.30 limit	SMIC structural pricing pressure; Taiwan geopolitical premium
4	DELTA.BK	3.05	Low	Exceptional business (ROIC 28.5%, 60% AI PSU market share); wrong vehicle	THB 250–270	CEO has zero personal DET shares; parent pledged $525M in shares near peak; 147x P/E
5	AMKR	2.90	Low (timing)	Thesis intact; Arizona + Intel Foveros + Kim family 49.5% stake	$55–70 post-Apr 27 earnings	52x P/E; all 8 analyst targets below current price; FOMO trap

Preferred combination: 2308.TW (2–3%) + UMC (1–1.5%) + 6809.HK (1–1.5%). Own 2308.TW over DELTA.BK — same tech, better governance, better entry thesis.

---

CROSS-SWARM TOP 5 (Conviction Ranking)

Rank	Ticker	Action	Entry	Key Catalyst	Mgmt
1	ONTO	Scale Buy	$290–300 starter; $260–275 add	Q1 earnings May 2026; Semilab revenue ramp	B+
2	ATS.VI	Scale Buy	≤€85 starter; add on May 2026 FY results	ABF supply crunch; Kulim utilization	B/Yellow
3	PDFS	Buy on confirmation	$41–43 pilot post-May 7	Q1 FY2026 revenue ≥$50M	A-
4	2308.TW	Scale post-earnings	TWD 1,700 post-Apr 30	Vera Rubin power cycle Q3 2026	B+/Green
5	UMC	Stage in	$11.00–$11.30 limit	H2 ASP +10% not in consensus; Apr 29 earnings	B+

---

MANAGEMENT QUALITY OVERLAY

Ticker	Grade	Alignment	Guidance	Follow-Through	Key Flag
PDFS	A-	Exceptional	Conservative	~85%	No succession plan
ONTO	B+	Good	Straight	~80%	CEO 10b5-1 selling (scheduled)
2308.TW	B+/Green	Exceptional	Conservative	~86%	45% board independence (TW norm)
UMC	B+	Partial	Conservative	~85%	1.31% insider ownership
AWX.SI	B/Yellow-Green	Good	Straight	~80%	GVT cross-holdings; 3rd CEO
ATS.VI	B/Yellow	Partial	Improving	1 clean beat (Mertin era)	Dual-bloc shareholder conflict
S71.SI	B / C+ capex	Good	Conservative	~75%	Bad FY2022 capex timing; no buybacks at sub-book
6809.HK	B+ (revised)	Partial	Straight	~100%	Founders extracted RMB 1.92B
UCTT	B	Partial	Straight	~75%	Securities fraud class action (prior CEO era)
AMKR	B+	Strong	Conservative	~80%	Arizona execution risk not in comp structure
DELTA.BK	B-	Misaligned	Straight	~75%	CEO zero personal DET shares; parent pledged $525M
8027.TWO	C+/Yellow	Partial	Optimistic	~50–60%	Chairman 47.9% shares pledged

---

EARNINGS CALENDAR — NEXT 6 WEEKS

Date	Ticker	What to Watch	Entry Trigger?
Apr 27	AMKR	Revenue >$1.7B; GM >13.5%; Q2 guide ≥$1.75B	Clean entry $55–70 if disappoints
Apr 28	UCTT	H2 2026 demand step-up commentary	Build to 3–4% on confirmation
Apr 28	DELTA.BK	AI server PSU order backlog	No — wait for THB 250–270
Apr 28	6809.HK	Q1 2026 results; MRDIMM Gen2 ramp	Beat → entry HK$250–265; miss → HK$175–190
Apr 29	UMC	GM >28%; Singapore Phase 3 timeline; ASP hike	Confirmation → stay in at $11–$11.30
Apr 30	2308.TW	Revenue >TWD 110B; Vera Rubin backlog	Beat → initiate TWD 1,700–1,800
May 7	PDFS	Revenue ≥$50M, ≥18% YoY, margins stable	Yes — primary buy trigger
May 11	AWX.SI	New customer identity hints	Beat → consider S$5.00–5.50
May 11	8027.TWO	Q1 revenue >TWD 500M	Yes — required for add
Jun 5	S71.SI	2H FY2026 burn-in services recovery	Recovery confirmed → add at SGD 0.48–0.52

---

KNOWN vs. UNKNOWN

Verified:

• ATS.VI Kulim FCF inflection is confirmed (operating cash +€331.8M YTD vs. –€29.4M prior year)
• UMC H2 2026 ASP hike ~10% confirmed April 17 — not yet in sell-side consensus models
• AMKR Kim family has NOT diversified (49.5% concentrated; $10B+ personal stake)
• ONTO has $640M net cash and a $240M SK Hynix supply agreement through 2027
• PDFS Kibarian bought $1.13M in open market at the 52-week low in February 2025

Open / unverified:

• AWX.SI's unnamed new customer (the bull case depends on this)
• ATS.VI Intel EMIB volume order (currently pilot-stage only — AMD is the confirmed anchor)
• 6809.HK MRDIMM Gen2 ramp timeline and volume (undermodeled in consensus but unconfirmed)
• UCTT securities fraud class action resolution (prior CEO era; not current management)
• 8027.TWO Q4 2025 TWD 721M revenue — AIS one-time recognition event or run-rate?

Watch for:

• Intel Foundry Services external customer win announcement — would re-rate EMIB-linked names broadly
• Any delay disclosure on Intel Panther Lake (18A) production timeline
• ABF substrate spot price movements — leading indicator for ATS.VI volume and pricing

---

RISK OVERLAY

Intel-specific (applies to all names):

• 18A yield ramp delays → reduces near-term demand across the supply chain
• IFS fails to win major external customers → EMIB-linked bull case deflates
• Intel strategic pivot (management change risk — new CEO Pat Gelsinger 2.0 dynamic)

Country/geopolitical:

• Taiwan conflict risk: 2308.TW, 8027.TWO, UMC
• China export controls: 6809.HK (Jintide relies on TSMC + Intel supply — both restricted)
• Thailand political/macro: DELTA.BK

Valuation (universal): Every name is at or near 52-week highs as of April 26. Earnings cluster April 27–30 is the price discovery event. No new positions before those prints.

---

NEXT STEPS

Priority	Action	Why
1	Watch AMKR earnings April 27	Entry decision point at $55–70 if disappoints
2	Watch UCTT/DELTA.BK/6809.HK earnings April 28	Three separate decisions; no pre-earnings buy
3	Set limit order UMC $11.00–$11.30	Only name not obviously overpriced today; H2 ASP hike unpriced
4	Watch UMC April 29 earnings call for ASP hike commentary	Confirmation strengthens the $11 entry thesis
5	Watch 2308.TW April 30 earnings; initiate TWD 1,700 if beats	Best business quality in the server/power sub-theme
6	PDFS — do not buy above $48 before May 7	April 24 gap-up was FOMO; wait for earnings confirmation
7	ATS.VI — initiate ≤€85; add on May 2026 full-year results	Best PEG in group; ABF shortage structural
8	ONTO — initiate $290–300; add on pullback to $260–275	Net cash underwriting the multiple
9	Set alerts: AWX.SI S$4.50, S71.SI SGD 0.48	Pullback entries in burn-in sub-theme
10	Conduct deep-dive analysis on burn-in technology	Background reading before AWX/S71 earnings