Deal Story
Home›Notes›Macro
Macro

AI Semiconductor Value Chain 2026 — Three Bottlenecks and Korea's Position

Q1 2026: SK Hynix at 72% operating margin, Hami Semiconductor revenue YoY -65.5%. The Korean cycle split even within the same country. AI semiconductors are priced at three bottlenecks — HBM, CoWoS, EUV. $1T global market, 30+ Korean suppliers mapped, three alpha names (Hami, Soulbrain, HPSP), and the 2027 oversupply scenario — all in 50 minutes.

2026-06-02·50 min read·210 sources

Key Takeaways

  • Q1 2026: the Korean semiconductor cycle split sharply even within one country. SK Hynix posted a 72% operating margin while Hami Semiconductor saw revenue fall YoY -65.5%. The gap between HBM cells and stacking equipment, across a generation transition window, drove the divergence.
  • The global semiconductor market is recovering from $526.8B in 2023 to ~$1T in 2026, driven by AI demand (WSTS).
  • Combined AI chip revenue (NVIDIA DC + AMD DC + Broadcom AI semi) rose from ~$27B in Q1 2024 to ~$70B in Q1 2026 — 2.6× in two years. Broadcom backlog at $73B.
  • Three bottlenecks: HBM (SK Hynix 59%), CoWoS (TSMC), EUV (ASML 100%).
  • TSMC decides NVIDIA's capacity. The real cause of the 2024-25 Blackwell delay was CoWoS shortage (SemiAnalysis). Priority allocation among NVIDIA, AMD, Broadcom, and Marvell sits with TSMC.
  • China is catching up. EUV and HBM stay safe for 5-10 years, but general DRAM, NAND, and etch equipment face real pressure within 5 years. CXMT DDR5, YMTC 232L NAND, and AMEC etch are the long-term variables for the Korean memory cycle.
  • As of 2026 we are in an HBM undersupply phase (SK Hynix 72% OPM is the proof). The 2027+ oversupply scenario is emerging: combined SK / Samsung / Micron HBM capacity scales 85K → 185K (2.2×) while demand grows only ~1.5-1.8×, raising a 2027 ASP -30% to -40% scenario (SemiAnalysis, Jim Chanos).
01

1. Same Quarter, Same Country — Korea's Semiconductor Cycle Split in Two

In Q1 2026, SK Hynix reported quarterly revenue of KRW 52.58 trillion, operating profit of KRW 37.61 trillion, and an operating margin of 72%.

The largest quarterly result in company history.

More than half of the revenue came from HBM (High Bandwidth Memory).

SK Hynix's global HBM market share reached 57-59%, surpassing Micron's 19% and Samsung's 22% combined.¹

In the same quarter, Hami Semiconductor recorded revenue of KRW 50.9 billion — down 65.5% year-over-year.

One of the worst quarters in company history.

Hami doesn't make HBM. Hami makes the TC Bonder (Thermal Compression Bonder), the machine that stacks HBM.

Hami holds approximately 70% of the global TC Bonder market — a near-monopoly position.²

Yet revenue was cut in half. The reason is simple: the transition window between HBM3E 12-Hi and HBM4 created a temporary order gap.

In the same quarter, Soulbrain posted revenue of KRW 251.7 billion, up 18.9%. HPSP guided full-year 2026 revenue at KRW 234.1 billion, on a recovery trend.³

Within the same Korean semiconductor supply chain, the cycle split in two.

SK Hynix Q1 2026 OPM

72%

Quarterly revenue ₩52.58T, all-time high

Hami Q1 2026 revenue YoY

-65.5%

TC Bonder generation transition gap

Soulbrain Q1 2026 revenue

+18.9%

HBM · CoWoS · EUV exposure

The One Question This Note Answers

Why did the same quarter, in the same country, produce such divergent results across companies? The answer compresses into one line: in the AI era, semiconductor pricing and priority are decided at three bottlenecks — HBM, CoWoS, and EUV. Where a company sits relative to those three bottlenecks determines its fate. SK Hynix faced the cycle head-on as the #1 HBM cell manufacturer. Hami, the #1 HBM stacking equipment maker, was caught precisely in the generation transition gap. Soulbrain, embedded as a materials supplier in both HBM and CoWoS, rode the broader flow. HPSP, which dominates High-Pressure Annealing (the step after EUV in the logic flow), has effectively become a single global indicator for logic capex.

The starting point of this cycle is clear.

OpenAI launched ChatGPT on November 30, 2022. Within a month, NVIDIA H100 orders exploded.

In 2023, the global semiconductor market contracted to $526.8B, down 8.2% YoY — but within that contraction, data center GPU revenue grew over 200%.

In 2024 the total market recovered to $627.6B (+19.1%), reached $791.7B in 2025 (+26%), and is on track to break $1T in 2026 per WSTS Autumn 2025 forecast.⁴

While GDP grew 1.5×, semiconductors doubled.

Within that growth, combined AI chip revenue (NVIDIA Data Center + AMD Data Center + Broadcom AI semi) rose from approximately $27B in Q1 2024 to approximately $70B in Q1 2026 — a 2.6× increase.⁵

The entire cycle is being driven by AI.

What This Article Covers

In 50 minutes across 17 sections: - How the semiconductor market was redrawn after ChatGPT in 2023 (→ ✓2-4) - The memory cycle — how HBM is eating into general DRAM (→ ✓5) - Foundry dynamics — TSMC dominance, Samsung and Intel chasing (→ ✓6) - The thesis that China could catch up within five years (→ ✓7) - The three bottlenecks — HBM, CoWoS, EUV (→ ✓8-10) - Japanese, US, and European supplier mapping (→ ✓11) - Where Korea's 30+ supplier companies sit in this flow (→ ✓12) - The three Korean alpha plays — Hami, Soulbrain, HPSP (→ ✓13) - The mechanism through which TSMC controls NVIDIA's capacity (→ ✓14) - Five risk channels, centered on the memory cycle (→ ✓15) - Seven weekly indicators for investors to watch (→ ✓16)

Eight Acronyms You'll See Repeatedly

AcronymFull nameOne-line meaning
HBMHigh Bandwidth MemoryDRAM stacked 12-16 layers for ultra-high bandwidth
TC BonderThermal Compression BonderPrecision bonder for HBM layer stacking
CoWoSChip on Wafer on SubstrateTSMC's advanced packaging combining GPU + HBM
EUVExtreme Ultraviolet13.5nm wavelength lithography — required at 7nm and below
HPCHigh Performance ComputingData center and supercompute market
ASICApplication-Specific Integrated CircuitCustom chips like Google TPU, AWS Trainium
ALDAtomic Layer DepositionAtom-by-atom precision coating
PRPhotoresistLight-sensitive lithography material
02

1.5. Twelve Core Terms in One Page — Analogy Glossary

Semiconductor terminology is hard. The twelve most-cited terms, explained through analogy on a single page.

Return here whenever a term blocks you.

Memory

TermDefinitionAnalogy
HBMMulti-layer stacked DRAM for ultra-high bandwidthMemory built like a high-rise instead of stretched horizontally
TC BonderMachine that precisely bonds each HBM layerThe precision crane that stacks each floor. Hami Semiconductor holds ~70% globally
HBM substrateCircuit board beneath the HBM stackThe building's foundation. Japan's Ibiden and Shinko dominate

Foundry / Packaging

TermDefinitionAnalogy
FoundryA factory that manufactures chips designed by another companyNVIDIA designs, TSMC manufactures
CoWoSTSMC's advanced packaging — combines GPU + HBM on one platformIn 2026, NVIDIA, AMD, and Broadcom are all queued for it
ABF substrateCircuit board above the CoWoS stack — Ajinomoto holds near-100%Japan's Ajinomoto effectively owns the category
OSATOutsourced Semiconductor Assembly and TestASE, Amkor, Hana Micron

Lithography / Logic

TermDefinitionAnalogy
Node (3nm, 2nm)Width of a circuit line — smaller means more transistorsNarrower roads pack more lanes into the same surface area
EUVExtreme Ultraviolet — required for any node 7nm or belowThe most precise blueprint printer. ASML is the world's only manufacturer
High-NA EUVNext-generation EUV (NA 0.55), built for 2nm and belowA step finer blueprint printer. First adopted in 2025 by Intel, Samsung, TSMC
HPSP (High-Pressure Annealing)High-pressure hydrogen anneal that heals defects after EUV patterningThe finishing pass that fills in microscopic flaws. HPSP holds a near-monopoly globally

Industry Category

TermDefinitionAnalogy
Sobujang (소부장)Materials, Components, and Equipment — everything that feeds a fabSuppliers who sell ingredients, knives, and ovens to the chef (foundry)

Preview — The Ten-Stage Value Chain

The semiconductor value chain breaks into ten stages. §3 maps the whole thing: 1. EDA tools (Cadence, Synopsys) 2. IP (ARM, RISC-V) 3. Design — Fabless (NVIDIA, AMD, Broadcom) 4. Foundry (TSMC, Samsung, Intel, SMIC) 5. Memory (SK Hynix, Samsung, Micron) 6. Back-end / Packaging (TSMC CoWoS, ASE, Amkor) 7. EUV (ASML) 8. General equipment (AMAT, LAM, KLA, TEL, Disco) 9. Materials (Shin-Etsu, SUMCO, JSR, Soulbrain, Dongjin) 10. Inspection / Test (Advantest, KLA, Lasertec, Intekplus)

03

2. ChatGPT Redrew the Semiconductor Market

OpenAI launched ChatGPT on November 30, 2022.

Within a month, NVIDIA H100 orders exploded. Three years later, the global semiconductor market is approaching $1T from $526B.

2023 was decisive. The total market contracted 8.2%.

Within that contraction, NVIDIA's data center revenue jumped from approximately $15B in 2022 to approximately $47B in 2023 — over 200%.

In the same industry in the same year, memory hit a cycle bottom and general DRAM prices passed through a two-year low.

It was the first quarter the same industry visibly split into two directions.

Chart 1 — Global Semiconductor Market vs. AI Chip Combined ($B, 2022-2026E)

Source: WSTS Autumn 2025, SIA Feb 2026. AI chip combined = NVIDIA DC + AMD DC + Broadcom AI semi (annualised).

Global Semiconductor Market Cycle ($526B → $1T)

YearRevenue ($B)YoYNote
2022574.1+3.3%End of COVID-era PC demand
2023526.8-8.2%Memory cycle bottom — yet AI chips alone +200%
2024627.6+19.1%AI recovery in full swing
2025791.7+25.6%HBM market explosion (SIA Feb 2026)
2026E~1,000~+25%$1T milestone likely (WSTS Autumn 2025)

Source: WSTS, SIA, global semiconductor statistics.⁶

AI Chip Revenue Explosion — Combined View

QuarterNVIDIA DC ($B)AMD DC ($B)Broadcom AI semi ($B)Combined ($B)
2024 Q122.62.3~2.3 (est.)~27
2024 Q435.63.93.8~43
2025 Q4~50+~56.2~62
2026 Q1 (CY)~55+~68.4~70
2026 Q2 guidancen/an/a10.7n/a

Source: NVIDIA Form 8-K (FY2026 Q4), AMD IR, Broadcom IR (Q1 FY2026 AI semi $8.4B, +106% YoY).⁷

In two years, combined AI chip revenue grew 2.6× — from ~$27B to ~$70B.

Broadcom's trajectory is particularly notable. Q1 FY2026 AI semi revenue of $8.4B, +106% YoY. Q2 guidance of $10.7B, +140% YoY. Backlog of $73B.⁸

Broadcom dominates outsourced ASIC design for hyperscaler custom chips — Google TPU, AWS Trainium, Meta MTIA.

Looking only at NVIDIA misses Broadcom's ASIC backlog.

The real AI chip explosion is happening on two fronts simultaneously — GPU (NVIDIA, AMD) and ASIC (Broadcom, Marvell) — and the combined view is what matters.

Three Years, Compressed Into One Sentence

2022 ChatGPT launch → 2023 NVIDIA H100 explosion → 2024 full cycle recovery → 2025 HBM market explosion → 2026 the $1T milestone. §3 maps where Korea sits inside that flow.

04

3. The Semiconductor Value Chain — Ten-Stage Map

No single company builds a chip end-to-end. The industry is a ten-stage division of labor. Where a company sits in that chain determines its fate.

Chart 2 — Ten-Stage Semiconductor Value Chain & Korea's Exposure

1. EDA Tools
Cadence, Synopsys, Siemens EDA
KR: none2. IP
ARM, RISC-V
KR: none3. Design (Fabless)
NVIDIA, AMD, Broadcom, Marvell, MediaTek
KR: low4. Foundry
TSMC, Samsung, Intel, SMIC, UMC
KR: medium5. Memory
SK Hynix, Samsung, Micron, Kioxia
KR: high6. Back-end / Packaging
TSMC CoWoS, ASE, Amkor + Hana Micron, SFA, Simmtech
KR: medium7. EUV Lithography
ASML alone (Netherlands)
KR: none8. General Equipment
AMAT, LAM, KLA, TEL, Disco + Wonik IPS, Jusung, Hami
KR: medium9. Materials
Shin-Etsu, SUMCO, JSR, TOK + Soulbrain, Dongjin, SK Siltron
KR: high10. Inspection / Test
KLA, Lasertec, Advantest, Teradyne + Intekplus, Koh Young, Techwing
KR: medium
Korea highKorea mediumKorea lowKorea none

Box colour = Korea exposure: violet (high), blue (medium), grey (low), white (none).

Ten-Stage Division and Korea Exposure

StageWhatGlobal dominantKorea exposure
1EDA tools (circuit design software)Cadence, Synopsys, Siemens EDANear 0
2IP (reusable circuit blocks)ARM, RISC-VNear 0
3Design (Fabless)NVIDIA, AMD, Broadcom, Marvell, MediaTek, QualcommSamsung LSI
4FoundryTSMC, Samsung, Intel, SMIC, UMCSamsung Foundry
5Memory (DRAM, NAND, HBM)SK Hynix, Samsung, Micron, KioxiaSK Hynix, Samsung
6Back-end packaging (CoWoS, OSAT)TSMC, ASE, AmkorHana Micron, SFA, Simmtech
7EUV lithographyASML (Netherlands)0
8General equipment (etch, depo, CMP, inspection)AMAT, LAM, KLA, TEL, DiscoWonik IPS, Jusung, Hami
9Materials (wafer, PR, etchants, precursors, gases)Shin-Etsu, SUMCO, JSR, TOK, HenkelSoulbrain, Dongjin, SK Siltron, Wonik Mat
10Inspection / testKLA, Lasertec, Advantest, TeradyneIntekplus, Koh Young, Techwing

Korea is anchored in Stage 5 (memory), parts of Stage 8 (equipment), parts of Stage 9 (materials), and parts of Stage 10 (inspection).

Stages 1-2 (EDA, IP) are essentially zero. Stage 7 (EUV) is zero. Stage 3 (Fabless) has Samsung LSI but no meaningful global share. Stage 4 (Foundry) has Samsung chasing TSMC, but the gap is not closing.

What Korea owns is memory (5) and parts of sobujang (8-10).

This structure is the starting point for every analysis that follows.

Operating Margin by Stage — Who Takes More (2025 FY + Q1 2026)

StageCompany2025 OPMQ1 2026 OPM
EDACadence~35%~36%
EDASynopsys~30%~32%
IPARM~24%~25%
Design GPUNVIDIA~60%~65%
Design GPUAMD~10%~11%
Design ASICBroadcom~45%~47%
Design ASICMarvell~10%~15%
FoundryTSMC~45%~50%
FoundrySamsung Foundry~-5%~-3%
FoundryIntel IFS~-25%~-20%
Memory HBMSK Hynix~50%72% (cycle peak)
MemorySamsung Memory~25%~30%
MemoryMicron~30%~36%
EUVASML~32%~35%
General eq.AMAT~30%~30%
General eq.LAM~28%~29%
General eq.KLA~35%~36%
MaterialsShin-Etsu~30%~30%
InspectionLasertec~35%~35%
Korea sobujangHami~30%~-10% (HBM gap)
Korea sobujangSoulbrain~18%~22%
Korea sobujangHPSP~50%(2026E ~45%)

Source: 2025-Q1 2026 IR aggregated. NVIDIA, SK Hynix (cycle peak), TSMC and ASML sit at 35%+ OPM. Within the same industry, stage selection produces 5-7× margin differences.

Share this deal

05

4. Fabless Design — NVIDIA, AMD, Broadcom, Marvell, MediaTek

The AI chip market splits two ways: GPU (NVIDIA, AMD) and ASIC (Broadcom, Marvell). Both are exploding in parallel.

NVIDIA dominates GPUs. Q4 FY2026 data center revenue of $35.6B — 88% of total quarterly revenue. Blackwell (B100, B200) was announced April 2024 and entered production Q1 2025. Each Blackwell chip carries 8 stacks of HBM3E for a total of 192GB. Rubin (next gen) was previewed Q4 2025, production starting 2026, with 8 stacks of HBM4. SemiAnalysis estimates NVIDIA takes roughly half of TSMC's CoWoS capacity.

AMD is the chasing #2. MI300X (2024), MI325X (H1 2025), MI350 (H2 2025), MI400 (2026) roadmap. Q1 2026 data center GPU revenue is estimated around $6B. About 1/9 of NVIDIA, but the only meaningful alternative. OpenAI began deploying MI300X at scale from late 2024; Microsoft Azure has launched MI300X instances.

Broadcom is the real ASIC heavyweight. The interesting part: Broadcom doesn't make GPUs. It does custom-silicon design and production management for hyperscaler in-house chips — Google TPU, Meta MTIA, ByteDance Pangu. Revenue model is the opposite of NVIDIA — design fee plus per-chip royalty.

Q1 FY2026 AI semi revenue of $8.4B, +106% YoY. Q2 guidance of $10.7B, +140% YoY. Backlog of $73B.

What that backlog implies: AI-chip volume already on order from hyperscalers equals roughly one year of NVIDIA revenue.

Marvell is the #2 ASIC name. Core wins: AWS Trainium and Inferentia, plus part of Microsoft Azure Cobalt CPU. FY2025 revenue ~$6B with AI ~45% of that. Combined NVIDIA, Broadcom and Marvell account for 90%+ of the AI chip market.

MediaTek is Taiwan's dark horse. Smartphone SoC is the core business, but from 2024 it began co-designing AI PC chips with NVIDIA. Trying to enter ARM-based data center CPU. Active Taiwan government support. Over the long run, it could challenge Broadcom and Marvell.

Korean fabless has near-zero exposure to global AI chips. Samsung LSI Exynos remains in smartphone APs. LX Semicon is DDI (display driver IC). Rebellion has emerged as an NPU startup, but revenue isn't yet meaningful.

This is a structural Korean weakness — design is thin; memory and parts of sobujang carry the country.

The Biggest Misconception in AI Chips

Look only at NVIDIA and you miss Broadcom's $73B ASIC backlog. That backlog signals hyperscalers migrating from NVIDIA to in-house silicon. Inside that flow, the real alpha is the supply chain both sides depend on — TSMC CoWoS, SK Hynix HBM, ASML EUV.

06

5. The Memory Cycle — HBM Is Eating General DRAM

The same SK Hynix fab and the same wafer can make either HBM or DDR5.

HBM consumes roughly 4× the wafer area per useful bit. So as HBM production grows, DDR5 supply tightens. Q1 2026 DDR5 spot prices are surging for exactly that reason.

HBM Generation Timeline

GenerationBandwidthStackProduction startMain GPU
HBM2256 GB/s4-Hi2016NVIDIA P100 (2016)
HBM2E460 GB/s8-Hi2020A100 (2020)
HBM3819 GB/s12-Hi2023H100 (2023)
HBM3E1,200 GB/s12-Hi → 16-Hi2024-2025H200, Blackwell B100/B200
HBM42,000 GB/s+16-Hi2026 (production start)Rubin, MI400

Source: SK Hynix, Samsung, Micron IR. Bandwidth roughly doubles every three years.

Chart 3 — Quarterly HBM Share (2023 Q1 → 2026 Q1)

Source: Counterpoint Research, TrendForce. Q1 2026 is estimate. The key event: SK Hynix held #1 while Micron overtook Samsung.

A $0.5B market in Q1 2023 became a $20B market in three years — 40×.

Over the same span, SK Hynix held #1 while Micron overtook Samsung. Micron first passed Samsung in Q4 2024 because Samsung was 9 months late on NVIDIA HBM3E 12-Hi qualification.

In Q3 2025 Samsung began full HBM3E 12-Hi shipments, and Q1 2026 reports indicated Samsung passed top-line NVIDIA HBM4 qualification.

The Mechanism: HBM Cannibalizing General DRAM

An SK Hynix wafer can be used for HBM or DDR5. One HBM stack consumes roughly 4× the wafer area of a comparable DDR5 chip. So the more HBM made on a wafer, the less DDR5 supply available. Q1 2026 result is clear: CNBC reported general DRAM in shortage. Two implications for Korean semis: 1. SK Hynix and Samsung memory-segment OPMs enter all-time-high territory. 2. Server, PC and smartphone OEMs using DDR5 face severe memory-cost pressure on margins.

DDR5 8Gb Spot Price Trend

QuarterDDR5 8Gb spot ($)Note
2023 Q3~1.3Cycle bottom
2024 Q4~1.9Recovery
2025 Q43+HBM wafer crowd-out accelerating
2026 Q1Shortage reportsCNBC, Apr 2026

Source: TrendForce, DRAMeXchange, news compilations. Spot prices are volatile and definitions vary.

NAND tracks DRAM with a 6-12 month lag. NAND ASP recovered to ~$2.6 (512Gb TLC) in 2024 and exceeded $3 in Q4 2025. Kioxia (ex-Toshiba Memory) reportedly pursuing 2026 IPO. YMTC (China) has become a supply-side threat with 232L NAND mass production.

The Real Risk in the Memory Cycle

When the HBM cycle turns is the biggest variable. SemiAnalysis and Jim Chanos (short-seller) have flagged 2026-27 HBM oversupply concerns. All three — SK Hynix, Samsung, Micron — announced HBM4 capacity expansions. If NVIDIA Rubin and AMD MI400 shipments grow slower than expected, HBM supply could exceed demand. §15 examines this scenario.

07

6. Foundry — TSMC Dominates, Samsung and Intel Chase

Foundry is effectively a one-company industry. TSMC holds 60%+ globally and 90%+ at leading nodes (N3 and below). Samsung is chasing; Intel is trying again via IFS.

Chart 4 — Foundry Share by Node (Q1 2026)

Source: TrendForce, SemiAnalysis Q1 2026 analysis. N2 will shift as 2026 ramp proceeds.

TSMC Q1 2026: revenue of $35.9B (+30% YoY), HPC mix at 61%. 2026 capex guidance raised to $52-56B. Most of that capex flows to CoWoS capacity expansion and N2 production fabs.

Samsung Foundry is chasing. SF2 (2nm class) entered production in Q4 2025 using a GAA (Gate-All-Around) architecture to compete head-on with TSMC N2. Q1 2026 reports indicated Samsung won partial NVIDIA, Qualcomm and Tesla (in-house chip) orders.

The problem is yield. Industry reporting puts Samsung SF2 yield 20-30 percentage points below TSMC N2. The same chip cannot be made at the same price.

Intel Foundry Services (IFS) is now producing 18A. After Pat Gelsinger's departure, the new CEO simplified strategy to 18A-first. Production started Q4 2025. Microsoft and Amazon AWS reportedly won some outsourcing slots.

External share is still below 10%. Intel's in-house chips (Panther Lake, etc.) remain the main 18A demand.

SMIC leads China's chase. 7nm (N+2) production began 2023. 5nm-class attempts reported in 2024-25. But without EUV, using DUV multi-patterning, yield and cost are severely disadvantaged.

Huawei Mate 60 Pro's Kirin 9000S is an SMIC 7nm chip. Q1 2026 SMIC revenue was $2.5B.

The Essence of the Foundry Cycle

TSMC's lead will not be broken in a few years. Three reasons. 1. Leading-node entry cost has exploded to $20B+ per fab. 2. Reaching the same yield at the same node requires years of learning curve. 3. Customers don't move. NVIDIA, AMD, Apple, Qualcomm have no reason to leave TSMC. This is the starting point for the §14 discussion of how TSMC decides NVIDIA's capacity.

08

7. China's Rise — Catch-up Within Five Years?

China has no EUV. Even so, in memory, NAND and etch equipment it can threaten Korea within five years. China's Big Fund Phase 3 is deploying $47.5B.

US-China Export Control Timeline

TimeActionCore impact
Oct 2022Export controls on 16/14nm and below leading-edge equipmentASML EUV exports to China zero; partial AMAT, LAM, KLA blocks
Oct 2023Ban on NVIDIA H100/H800/A100/A800 exports to ChinaNVIDIA China revenue -50%; H20 (lower-spec) released
Oct 2024HBM 18% bandwidth threshold export controlHBM2E and above effectively blocked; China accelerates own HBM
Dec 2024China Big Fund Phase 3 announced ($47.5B)All-out push on memory, logic, equipment localization
2025-2026Additional controls + retaliatory tariff talksUS-China semi front becomes a macro political variable

Source: US BIS notices; China National IC Industry Investment Fund announcements.

Chart 5 — China Semiconductor Self-Sufficiency (%, 2018-2026E)

Source: SIA, Boston Consulting, SemiAnalysis estimates aggregated.

Where China can catch up within five years.

SMIC began 7nm (N+2) mass production in 2023, with 5nm attempts reported in 2024-25 and Q1 2026 revenue of $2.5B. Without EUV, using DUV multi-patterning, cost is estimated 2-3× vs TSMC with yield about half. Yet Huawei Mate 60 Pro's Kirin 9000S, built on SMIC 7nm, proved technical feasibility. SemiAnalysis estimates a 5-10 year gap will persist below 5nm.

CXMT completed DDR4 mass production and began DDR5 in 2024-25, with LPDDR5 (mobile memory) production scheduled for 2026. TrendForce estimates 6-7% global DRAM share at end-2025, with a path to 10%+ by 2027. The tech gap is roughly 2-3 generations — but price competitiveness is clearly a threat.

YMTC succeeded in 232L NAND mass production (2024) — technically on par with Kioxia and Micron. Despite US-China controls limiting access to the most advanced tools, YMTC already holds ~10% global NAND share by 2025. NAND is more cyclical than DRAM, so Chinese supply expansion risks dragging down global NAND ASP.

AMEC has eroded part of LAM Research's CCP (Capacitively Coupled Plasma) etch share. Global etch share estimated at 8-10% in 2025 — long-run, real competition for LAM and TEL.

Naura has entered multiple categories (CVD, etch, PVD, clean), with 2025 revenue around $2.5B. AMAT and LAM remain dominant outside China but are steadily losing share inside China.

Two areas China cannot catch up on: EUV and HBM.

ASML EUV is under export control; in-house Chinese development is estimated 10-15 years behind. SMEE (Shanghai Micro Electronics Equipment) remains stuck at DUV.

For HBM, the real barrier isn't the tech gap — it's NVIDIA qualification. CXMT has been reported attempting HBM, but the chance NVIDIA adopts Chinese HBM is politically near zero.

Implications for Korean Semis

Three implications. 1. General DRAM (DDR5, LP5) enters price competition with China within five years. This is ~30-40% of SK Hynix and Samsung memory revenue. 2. NAND is already under pressure. If YMTC moves to near-dumping pricing, Samsung NAND operating profit could disappear. 3. HBM and EUV are safe for 5-10 years. Korean firms anchored here — SK Hynix HBM, Hami TC Bonder, Soulbrain phosphoric etch, HPSP high-pressure annealing — are relatively safer.

Share this deal

09

8. Bottleneck #1 — HBM, the New Cycle SK Hynix Captured

The HBM market grew from $0.5B to $20B in three years — 40×. SK Hynix holds 60%. Behind it sit Hami TC Bonder at 70%, Simmtech substrate, and Soulbrain etchant.

HBM Supply Chain on One Page

StageGlobal dominantKorea core
HBM cell (DRAM)SK Hynix 59% / Samsung 22% / Micron 19%SK Hynix dominant
TC Bonder (stacker)Hami ~70%Hami
HBM substrateIbiden, Shinko (Japan) dominantSimmtech (partial)
HBM TSV grindingDisco (Japan) ~80%Partial
HBM inspectionKLA, AdvantestIntekplus, Koh Young
HBM etchant, precursors(several global)Soulbrain, Wonik Mat, DNF, Foosung
HBM test handlersAdvantest (some)Techwing
HBM package (final)TSMC CoWoS + SK Hynix in-house MR-MUFSFA Semicon, Hana Micron

SK Hynix is HBM's clear leader. Q1 2026 revenue ₩52.58T, operating profit ₩37.61T, OPM 72% — an all-time high. HBM share at 57-59%, larger than Micron 19% + Samsung 22% combined. HBM3E 12-Hi supplies NVIDIA Blackwell B200 and H200 effectively on a sole-source basis. HBM4 reportedly passed top NVIDIA qualification in Q1 2026 and enters production Q2. NVIDIA Rubin and AMD MI400 will standardise on HBM4.

Samsung is the chasing #2 but with a real gap. Samsung HBM was 9 months late on NVIDIA HBM3E 12-Hi qualification, starting full shipments only in Q3 2025. In between, Micron overtook Samsung (first time Q4 2024). Q1 2026 HBM4 NVIDIA qualification represents Samsung's comeback attempt. Whether share actually recovers will take 2-3 more quarters to confirm.

Micron is the dark horse. FY Q2 2026 revenue $23.86B (+196% YoY); FY Q3 guidance $33.5B at 81% margin. HBM4 (Vera Rubin variant) is shipping in volume into NVIDIA Rubin. HBM share ~19-21%. As a US firm with no geopolitical risk, it is attractive to NVIDIA and AMD.

Korean HBM Supply Chain — Who is the Real Alpha

FirmRoleGlobal position2026 exposure
SK HynixHBM cellGlobal #1 (59%)Cycle peak
HamiTC BonderGlobal #1 (~70%)HBM gen-transition gap (Q1 2026 -65.5%)
SimmtechHBM substrateGlobal partial (Japan Ibiden dominant)SK Hynix-dependent
SoulbrainPhosphoric etchantKorea strongSK Hynix, TSMC dependent
Wonik MaterialsNF3, HCDS gasesKorea strongSK Hynix, Samsung dependent
IntekplusHBM inspectionKorea strongSK Hynix dependent
TechwingMemory test handlersKorea strongMicron, SK Hynix dependent
SFA SemiconOSAT (general memory)Korea midSamsung dependent
Hana MicronOSATKorea midSamsung dependent

The most interesting fact is Hami's 70% global share. As SK Hynix HBM grows, Hami grows — a direct beta. But Q1 2026's -65.5% revealed the generation-transition risk: the move from HBM3E 12-Hi to HBM4 can briefly stall TC Bonder orders. The long-term thesis (HBM market explosion) and short-term volatility move on different tracks.

Is SK Hynix's 72% OPM Sustainable?

The biggest variable is whether 72% OPM is sustainable. HBM demand is exploding, but all three suppliers (SK, Samsung, Micron) are expanding capacity. SemiAnalysis and Jim Chanos have raised 2026-27 oversupply scenarios. §15 examines this risk in detail.

10

9. Bottleneck #2 — CoWoS and How TSMC Decides NVIDIA's Capacity

CoWoS is the advanced packaging that combines GPU and HBM on one platform. TSMC effectively monopolises it. Capacity grew from 15,000 wafer/month at end-2023 to 130,000 at end-2026 — 8.6×. NVIDIA, AMD and Broadcom are all in the queue.

Chart 6 — TSMC CoWoS Monthly Wafer Capacity (2023-2026E)

Source: TrendForce, TSMC capex commentary. 8.6× in three years.

8.6× in three years. Over the same window NVIDIA data center revenue grew about 6× — CoWoS is growing faster than NVIDIA. And still short.

2024-25 NVIDIA Blackwell Delay — CoWoS Was the Real Variable

NVIDIA announced Blackwell in April 2024 for Q4 2024 production, with multiple delays. The stated cause was a Blackwell design issue. Per SemiAnalysis (Dylan Patel), the larger driver was CoWoS shortage. TSMC had to allocate CoWoS across NVIDIA, AMD, Broadcom, Marvell, Apple — NVIDIA could not get its full ask. Production slipped to Q1 2025, and AMD MI300X took some share in the gap.

TSMC Q1 2026: revenue $35.9B (+30% YoY), HPC (data center) at 61%. 2026 capex guidance raised to $52-56B. Most flows to CoWoS expansion and N2 production fabs. Market estimates TSMC invests $35-40B in CoWoS alone in 2026.

CoWoS Supply Chain

StageGlobal dominantKorea exposure
Interposer (CoWoS-S/L)TSMC in-houseNegligible
ABF substrateAjinomoto (Japan) ~100%0
Underfill materialsHenkel, Namics (Japan)0
CMP slurryCMC, Merck, CabotPartial Soulbrain, SK Enpulse
Si parts (etch-chamber parts)(global multi)Hana Materials, TCK
EUV/DUV lithographyASML0
InspectionKLA, Lasertec, AdvantestPartial Intekplus

Korea's CoWoS exposure is relatively limited. ABF substrate is effectively 100% Ajinomoto (Japan); underfill is also Japan-led. Korea sits mostly in etch-chamber parts and slurry.

This is one of Korean semis' structural weaknesses. A CoWoS explosion does not directly flow to Korea — but SK Hynix HBM going into CoWoS gives Korea indirect beta.

CoWoS Is the Whole Market's Bottleneck

§14 dives deeper into the TSMC-as-allocator mechanism. The key point: when NVIDIA, AMD and Broadcom all queue together, TSMC decides who gets how much. CoWoS is not NVIDIA's bottleneck alone — it is the bottleneck of the entire AI chip market.

11

10. Bottleneck #3 — EUV, ASML's 100% Monopoly

EUV is the only way to manufacture leading-edge nodes at 7nm and below.

ASML holds 100% global share.

Around 13 units shipped in 2024, with a target of 60+ in 2026.

High-NA EUV entered its first production deployment starting in 2025.

[ASML Q1 2026 results]

Revenue of €8.8B, up 13% YoY.

Of that, EUV revenue €4.1B (65% of system revenue).

First quarter of High-NA EUV revenue recognition, with 2 units booked.

EPS €7.15.

2026 guidance raised to €36-40B.

[EUV shipment trajectory]

53 units in 2023, around 44 in 2024, around 50 in 2025 plus roughly 5 High-NA cumulative across Intel, Samsung, and TSMC, 60+ guided for 2026, and 80+ guided for 2027.

[High-NA EUV, the next generation]

Numerical Aperture (NA) lifts from 0.33 on prior EUV to 0.55.

The same EUV step can now draw finer circuitry in a single exposure.

With prior EUV, sub-N2 nodes required multi-patterning and costs exploded.

High-NA resolves that.

Price per unit is around $350M, roughly 2× the prior EUV.

First deployment was Intel 18A production in 2025, with Samsung SF2P and TSMC A14 following in 2026-2027.

[EUV supply chain, Korea's exposure]

The EUV scanner is 100% ASML.

The EUV light source is Cymer (an ASML subsidiary).

EUV mask actinic inspection is 100% Lasertec (Japan).

EUV mirrors are Carl Zeiss SMT (Germany, an ASML subsidiary).

EUV PR (photoresist) is JSR, TOK, and Shin-Etsu (Japan, ~75% combined), with Dongjin Semichem chasing localization in partnership with SK Hynix.

EUV pellicles are dominated by Mitsui Chemicals (Japan), with some Korean entry.

High-pressure annealing (logic yield) is HPSP at 95%+, effectively a monopoly.

Specialty gases (NF3, C4F8, HCDS) come from Japan plus Korean players (Wonik Materials, Hooseong, DNF).

The EUV body itself sits entirely outside Korea.

ASML (Netherlands), Cymer (US), Carl Zeiss (Germany), and Lasertec/JSR/TOK/Shin-Etsu (Japan).

Korea's exposure concentrates in EUV PR (Dongjin Semichem) and high-pressure annealing (HPSP).

[HPSP, Korea's alpha in the step after EUV]

After EUV draws the circuitry, defects appear.

At tens-of-nanometers circuitry, micro-defects directly degrade yield.

HPSP holds an effective global monopoly on the high-pressure hydrogen anneal that heals those defects.

TSMC, Samsung Foundry, Intel IFS, and SK Hynix have all adopted HPSP equipment.

Operating margin sits around 50%.

The more global logic capex grows, the more HPSP revenue grows.

That is why HPSP serves as a single-stock indicator for global logic capex.

[Dongjin Semichem, localizing EUV PR]

EUV PR (photoresist) is around 75% Japanese — JSR, TOK, and Shin-Etsu.

Dongjin Semichem succeeded at KrF and ArF localization first, and is now chasing EUV PR through co-development with SK Hynix.

Partial supply to SK Hynix's EUV lines began in 2024-2025.

Over the long run, this could erode the JSR/TOK share.

Korea Doesn't Own EUV — But Owns What Comes Next

Korea owns almost nothing in EUV itself. But in the steps that come after EUV — HPSP's high-pressure annealing, Dongjin Semichem's EUV PR, and Wonik Materials' NF3 — Korea is building real alpha. The larger the EUV market gets, the larger these adjacent Korean suppliers grow with it.

12

11. Japan, US, Europe Suppliers — The Global Strongholds

The real global strongholds in semiconductor materials and equipment sit in Japan and the US.

Shin-Etsu leads silicon wafers, JSR leads PR (photoresist), TEL and Disco lead in Japanese equipment, and AMAT, LAM, and KLA form the US big three.

Korean suppliers are embedded alongside them, but only in narrower domains.

[Japanese materials]

Shin-Etsu, SUMCO, JSR, TOK, Mitsubishi

Shin-Etsu Chemical — silicon wafer #1: roughly 31% global silicon wafer share, the global leader. Q1 2026 revenue around ¥620B, with the semiconductor segment around ¥220B. Strong across multiple domains including EUV PR, photomask blanks, and silicon epoxy. Operating margin around 30%.

SUMCO — silicon wafer #2: around 22% globally. Q1 2026 revenue roughly ¥85B. Expanding 300mm wafer capacity.

JSR — global EUV PR #1: roughly 35% of global EUV PR (est.). Fully acquired by Japan's JIC (Japan Investment Corp) in 2024 for around ¥900B — a clear signal that semiconductor materials have become geopolitical assets.

Tokyo Ohka Kogyo (TOK) — EUV PR #2: around 25% (est.). Together, JSR, TOK, and Shin-Etsu hold approximately 75% of global EUV PR.

Mitsubishi Chemical and Showa Denko — specialty gases and materials: the core suppliers of etch gases such as NF3, CF4, and C4F8. Direct competitors to Korea's Hooseong and Wonik Materials.

[Japanese equipment]

TEL, Disco, Screen, Lasertec

Tokyo Electron (TEL): Japan's #1 and global #3. Strong across multiple domains including ALD, etch, clean, and CVD. Q1 2026 revenue around ¥650B. Sits behind AMAT and LAM. Korean revenue exposure is around 25% (SK Hynix and Samsung).

Disco: roughly 80% global share in wafer dicing and grinding. Disco dominates the TSV grinding step before HBM stacking. Q1 2026 revenue around ¥80B, operating margin around 35%.

Screen Holdings: roughly 50% global share in wet cleaning equipment.

Lasertec — 100% monopoly on EUV mask actinic inspection: the tool that inspects EUV masks for defects using extreme ultraviolet light. Every additional ASML EUV scanner lifts Lasertec revenue. Operating margin around 35%.

[US and European equipment]

AMAT, LAM, KLA, ASML

Applied Materials (AMAT): global equipment #1. Covers CVD, PVD, etch, CMP, and implant. Q1 2026 revenue around $7B, operating margin around 30%.

LAM Research: leader in etch and deposition. Q1 2026 revenue around $4.5B. Competes head-on with China's AMEC.

KLA Corporation: roughly 50% in inspection and metrology. Operating margin around 35%.

ASML: covered in §10.

Structural Traits of Global Suppliers

The Japanese, US, and European strongholds share three traits. First, they dominate domains with decades-long learning curves. In wafers, PR, EUV, and etch inspection, falling one generation behind is almost impossible to recover from. Second, customers (foundries and memory makers) do not switch. Once a given PR is qualified at a given fab, the cost of changing a proven supplier is too high. Third, operating margins sit stably in the 25-35% range. There is no NVIDIA-like 65% or SK Hynix-like 72% peak, but they stay profitable even at cycle bottoms. This is the structure Korean suppliers need to learn.

Share this deal

13

12. Mapping Korea's 30+ Suppliers — Four Categories

Korean suppliers are not the top-tier global strongholds, but several are #1 globally in specific domains.

Hami Semiconductor holds around 70% in TC Bonder, and HPSP holds 95%+ in high-pressure annealing.

Beyond those, more than 30 companies sit somewhere along the SK Hynix, Samsung, TSMC, and NVIDIA supply chains.

Korean Suppliers Revenue (2025-2026 Q1)

Tier 1 (global leader)Tier 2 (direct beta)Tier 3 (indirect)Tier 4 (weak AI)

Tier 1: Hami, HPSP. Tier 2: Soulbrain, Dongjin, Wonik Materials, Simmtech, Intekplus, Techwing, DNF. Tier 3: SK Siltron, Hana Materials, TCK, Wonik QnC, etc. Tier 4: DI Dongil, DMS, Russell, LB Semicon. · Source: company IR (2025 annual / 2026E). Estimates noted.

12.1 Materials (10+ Companies)

CompanyCore Product2025/26 Q1 RevenueKey CustomersAI/HBM Exposure
SoulbrainPhosphoric/etch acids, HFQ1 2026 ₩251.7B (+18.9%)SK Hynix, TSMC, SamsungStrong direct
Dongjin SemichemKrF/ArF/EUV PR2025 ~₩1.4TSK Hynix, SamsungEUV localization chase
SK SiltronSilicon wafers2025 ~₩2.5TSK Hynix, Samsung (SK group)Memory cycle
Hana MaterialsSiC parts (etch chambers)2025 ~₩450BLAM/AMAT/TELLogic capex beta
Wonik MaterialsNF3/WF6/HCDS gases2025 ~₩580BSK Hynix, Samsung, TSMCStrong direct
SK EnpulseCMP pads/slurry2025 ~₩750BSK Hynix, Samsung, TSMCIndirect
ENF TechnologyEtch/cleaning chemicals2025 ~₩430BSK Hynix, SamsungMemory cycle
HooseongFluorine gases2025 ~₩380BSK Hynix, SamsungDirect
DNFHCDS, ZrO2 precursors2025 ~₩120BSK Hynix, SamsungHBM exposure
Wonik QnCQuartz components2025 ₩943.6B / Q1 2026 ₩256.2BSamsung/SK/TSMC/Intel/MicronGlobal logic beta
LB SemiconDDI (display)Q1 2026 ₩134.3B (return to profit)LG Display, etc.Low AI exposure

10+ materials companies. Soulbrain, Wonik Materials, DNF have direct HBM exposure.

12.2 Components (6+ Companies)

CompanyCore Product2025/26 Q1 RevenueKey CustomersAI/HBM Exposure
Hami SemiconductorTC Bonder (HBM stacking)Q1 2026 ₩50.9B (-65.5%)SK Hynix primarilyHBM direct (generation transition gap)
HPSPHigh-pressure annealing (logic)2026E ₩234.1BTSMC, Samsung, Intel, SK HynixGlobal logic capex single indicator
TCKSiC parts2025 ~₩240BLAM/AMAT/TELLogic capex beta
EO TechnicsLaser marking/dicing2025 ~₩360BSK Hynix, SamsungIndirect
IntekplusHBM inspection2025 ~₩120BSK HynixHBM direct
Koh Young3D inspection (SMT)2025 ~₩280BVarious back-endIndirect

6+ component companies. Hami and HPSP are Tier 1 alpha.

12.3 Equipment (8+ Companies)

CompanyCore Product2025/26 Q1 RevenueKey CustomersAI/HBM Exposure
Wonik IPSCVD/etch/deposition2025 ~₩1TSamsung (~70%)Logic and memory
SEMESEFEM/clean/metrology(Samsung subsidiary)Samsung 100%Memory
Jusung EngineeringALD/CVD/etch2025 ~₩420BSK HynixMemory and logic
TESCVD/etch chambers2025 ~₩250BSamsung, SKMemory
YESTTest chambers2025 ~₩180BHami Semi, etc.HBM indirect
PSKAshing equipment2025 ~₩230BSamsung, SKLogic and memory
DMSVernoyi thin-filmQ4 2025 ₩57.7BDisplay + some semiWeak
RussellLithography supportQ4 2024 ₩16.3BSamsung, SKWeak

8+ equipment companies. Wonik IPS and Jusung Engineering are the mainstays.

12.4 Back-End OSAT (5+ Companies)

CompanyCore Product2025/26 Q1 RevenueKey CustomersAI/HBM Exposure
Hana MicronMemory packaging2025 ~₩1TSamsungMemory
TechwingMemory test handlers2025 ~₩280BMicron, SK, SamsungHBM direct
SFA SemiconBack-end OSAT2025 ~₩450BSamsungMemory
SimmtechHBM substrate (PCB)2025 ~₩1.3TSK HynixHBM direct
DI DongilBack-end IT (Plasmatec)2025E ₩608.9BSamsung Pyeongtaek scrubbersVery weak

5+ OSAT companies. Techwing and Simmtech have direct HBM exposure.

Korea's Real Alpha — 9 Companies Across Tier 1-2

Out of more than 30 companies, the names that receive direct AI/HBM cycle exposure narrow down to nine. Tier 1 (direct beta, global leaders): Hami Semiconductor (TC Bonder ~70%) and HPSP (high-pressure annealing 95%+). Tier 2 (direct beta, Korean leaders): Soulbrain (HBM and CoWoS etch chemicals), Dongjin Semichem (EUV PR chase), Wonik Materials (NF3, HCDS), Simmtech (HBM substrate), Intekplus (HBM inspection), Techwing (HBM test handlers), and DNF (HBM precursors). Tier 3 (indirect beta): SK Siltron, SK Enpulse, Hana Materials, TCK, Wonik QnC, Wonik IPS, and Jusung Engineering. Tier 4 (weak AI exposure): DI Dongil, DMS, Russell, and LB Semicon. The companies investors actually need to watch are the nine in Tier 1-2.

14

13. Korea's Three Alpha Plays — Hami, Soulbrain, HPSP

Among Korea's 30+ suppliers, three companies hold either global #1 share or an effective monopoly.

Hami Semiconductor (TC Bonder), Soulbrain (phosphoric etch), and HPSP (high-pressure annealing).

All three are Korean, but each is embedded in a completely different cycle.

Three Korean Alpha Companies at a Glance

DimensionHami SemiconductorSoulbrainHPSP
Core productTC Bonder (HBM stacking)Phosphoric/etch/HFHigh-pressure H2 annealing
Global position~70% (TC Bonder)Korean leader (global top group)95%+ near-monopoly
Key customersSK Hynix primarySK Hynix, TSMC, SamsungTSMC, Samsung, Intel, SK Hynix
AI exposure pathHBM cell demand → TC Bonder ordersHBM/CoWoS/EUV allGlobal logic capex direct
2025 revenue₩547.8B (HBM peak)~₩1T~₩193B
Q1 2026 revenue₩50.9B (-65.5%)₩251.7B (+18.9%)(no quarterly, 2026E ₩234.1B)
Peak OPM~30%+~22%~50%
Cycle characterSensitive to HBM generation gapsSmoothed via multi-domain exposureTracks logic capex precisely
Market cap (May 2026)~₩9-10T~₩5-6T~₩4-5T
Key weaknessSK Hynix concentration, generation gapsForeign competition (CMC/Merck/Cabot)TAM itself is small
Key strengthTC Bonder ~70% globalMulti-domain (HBM/CoWoS/EUV/logic)95%+ monopoly + logic capex beta

Source: each company IR 2025 + Q1 2026 preliminaries + sell-side reports. OPM and market cap rows highlighted.

[The three differences in one line each]

Hami Semiconductor = peak-to-trough oscillation of the HBM generation cycle.

During transitions between HBM generations (HBM3 → HBM3E → HBM4), orders briefly stop.

Q1 2026's -65.5% is the evidence.

With HBM4 full ramp, a strong recovery is expected in Q2-Q3 2026.

Soulbrain = a smoothed beta embedded across HBM, CoWoS, and EUV simultaneously.

When one domain wobbles, another holds the line.

Q1 2026's +18.9% is the evidence.

The explosive momentum is weaker than Hami or HPSP, however.

HPSP = a single-stock indicator for global logic capex.

When TSMC, Samsung, Intel, or SK Hynix expand logic fabs, HPSP revenue moves in lockstep.

Operating margin around 50%, the highest among Korean suppliers.

The only weakness is that the TAM itself is small.

[Why these three need to be watched together]

They are all Korean, but each is embedded in a completely different cycle.

When Hami stumbles, Soulbrain grows, and HPSP rides the global logic cycle.

Low share-price correlation creates real diversification.

From a Korean semiconductor supplier portfolio perspective, watching all three together is meaningful.

The Core of Korean Supplier Diversification

Two Tier-1 names (Hami Semiconductor and HPSP) paired with one Tier-2 name (Soulbrain) are what produces real diversification. Hami Semiconductor = the HBM cycle's peak-to-trough oscillator. Soulbrain = the multi-domain smoothed beta. HPSP = the single logic-capex indicator. The fact that the three cycles differ is the actual value.

15

14. Case Study — How TSMC Decides NVIDIA's Capacity

The NVIDIA Blackwell delay was publicly attributed to NVIDIA's design issues.

In reality, TSMC CoWoS capacity shortage was the larger variable.

TSMC sits in the position of deciding priority among NVIDIA, AMD, Broadcom, Marvell, and Apple.

NVIDIA Blackwell Delay Timeline (2024-2025)

DateEvent
2024.4NVIDIA Blackwell announcement (GTC 2024)
2024.8The Information reports first signs of delay
2024.10NVIDIA officially acknowledges design issues, slight push of Q4 ramp
2024.12TSMC CoWoS capacity ~35K wafer/month (vs original 50K target)
2025.1NVIDIA Blackwell B200 enters full production
2025.3AMD MI300X picks up some hyperscaler orders (absorbing NVIDIA shortfall)
2025.Q2NVIDIA Blackwell DC revenue fully recognized

Source: SemiAnalysis (Dylan Patel), The Information, NVIDIA Form 8-K composite.

[Surface vs. underlying cause]

The surface cause was a design issue in the Blackwell GPU itself.

The known reason was a redesign of the GH100→B200 chip-to-chip interconnect.

In reality, CoWoS capacity was the bigger variable, according to SemiAnalysis.

In Q4 2024, TSMC CoWoS capacity ran around 35,000 wafer/month.

NVIDIA was demanding more than half of that.

AMD, Broadcom, and Marvell had to split the remaining half.

The fact that NVIDIA could not get its full request showed up as the Blackwell production delay.

[TSMC's decision rights]

There are three.

First, who gets how much. NVIDIA cannot get everything. When AMD, Broadcom, and Marvell all line up at once, TSMC decides the allocation.

Second, when they get it. Blackwell was telegraphed for Q4 2024 but slipped into Q1 2025 — TSMC capacity was the variable that decided.

Third, at what price. The wafer price for CoWoS is set by TSMC. NVIDIA, AMD, and Broadcom have no real alternative.

This is the actual mechanism behind the NVIDIA Blackwell delay of 2024-2025.

[In 2026, TSMC's decision rights grow stronger]

By end-2026, CoWoS capacity scales to 120,000-130,000 wafer/month.

This is a much faster ramp than originally planned.

Even so, demand grows even faster.

NVIDIA Rubin (production in 2026), AMD MI400 (2026), Broadcom AI semi backlog of $73B, and Marvell AWS Trainium 2 all require CoWoS.

TSMC's priority-setting power becomes stronger in 2026.

[What Korea's SK Hynix gets in return]

The more CoWoS grows, the more HBM demand grows.

Each CoWoS wafer carries on average around 8 HBM stacks.

If CoWoS capacity scales 2023's 15K → 2026's 130K, that is 8.6×, so HBM demand also rises around 8.6×.

SK Hynix HBM revenue is a direct function of TSMC CoWoS capacity.

This is the essence of SK Hynix's 72% operating margin.

TSMC the Allocator, SK Hynix the First Beneficiary

TSMC's decisions determine the fates of NVIDIA, AMD, and Broadcom. But SK Hynix and Samsung HBM revenue flows directly from the CoWoS capacity TSMC has decided. TSMC is the capacity allocator for the entire AI chip market, and SK Hynix is the first-line beneficiary of that allocation.

16

15. Five Risk Channels — Memory Cycle in Focus

When this cycle could break is best examined through five channels.

The biggest variable is the memory cycle.

HBM oversupply scenarios are already being raised by SemiAnalysis and Jim Chanos.

Five Risk Channel Mechanism

HBM oversupplyMemory CyclehighCXMT/YMTC catch-upChina RisemediumUS-China politicsTariffs / ControlsmediumFab entry costNode CostlowTaiwan single dep.CoWoS DependencelowAI Semi Cycle

Five channels radiating from [AI semi cycle], colored by risk level.

[Channel 1: memory cycle, HBM oversupply (the biggest variable)]

This is the closest and largest risk.

The scenario is that HBM supply could exceed demand in 2026-2027.

SK Hynix, Samsung, and Micron have all announced HBM4 capacity expansions.

End-2025 total HBM capacity around 85K → end-2026 around 185K (2.2×).

Over the same period, NVIDIA Rubin + AMD MI400 shipment growth is estimated at around 1.5-1.8×.

Supply 2.2× vs. demand 1.5-1.8× → a gap of roughly 20-40 percentage points.

SemiAnalysis's Dylan Patel has said "2027 HBM ASP could fall -30 to -40%."

Jim Chanos has said "memory is cyclical forever."

Whether SK Hynix's 72% operating margin is sustainable is the core question of this channel.

[Channels 2-5]

Channel 2: China's rise — covered in detail in §7. CXMT DDR5 (a scenario of 10%+ global share by 2027), YMTC NAND 232L, AMEC etch beginning to erode some of LAM's share. EUV and HBM remain safe for 5-10 years, but general memory, NAND, and etch face material threats within 5 years.

Channel 3: tariffs and export controls — potential additional US export controls in 2026-2027 (tighter HBM, CoWoS, and EUV thresholds), Chinese retaliation, Trump-2.0 Taiwan tariffs, and EU CHIPS Act retaliatory tariffs. An unpredictable variable.

Channel 4: cost of entering leading-edge nodes — a TSMC N2 fab runs around $28B and an A14 fab around $40B. The companies that can pay this are TSMC, Samsung, and Intel. Over the long run, the foundry market concentrates further. NVIDIA and AMD become more deeply dependent on TSMC.

Channel 5: single CoWoS dependence — if CoWoS capacity is interrupted by earthquake, typhoon, or politics, the entire global AI chip market is affected. The April 2024 Taiwan earthquake was restored within 24 hours. The only alternative capacity sits at Samsung and Intel IFS, with a large gap at the leading edge. A geopolitical single-point-of-dependence risk.

Five Channels — Which Is Closest

ChannelShort (≤1yr)Medium (2-3yr)Long (5yr+)
1. Memory cycleMedium riskLargest riskAverages out
2. China riseLowMedium (memory/NAND)High (memory/NAND/etch)
3. Tariffs / export controlsMedium (unpredictable)MediumMedium
4. Advanced-node costLowMediumHigh (TSMC grows stronger)
5. CoWoS single dependenceLow (Taiwan stable)MediumMedium

Closest and largest is memory cycle, followed by tariffs/export controls. China rise unfolds slowly.

When Multiple Channels Activate Together

If 2-3 of the five channels activate at the same time, the cycle shock grows. Example: HBM oversupply + Trump tariffs + a China DDR5 surge = memory company operating margins could drop by 50 percentage points. This is not the base case for 2026-2027, but it is a tail risk.

Share this deal

17

16. Investor Actions — Institutional 5 / Retail 5 / Watch 7

[16.1 Five things institutional investors need to watch]

1. Quarterly HBM share changes: does SK's 59% move. SK falling below 50% means the cycle peak has passed. Samsung recovering above 30% weakens the SK single-beta story. Micron reaching 25%+ introduces a US political variable.

2. TSMC CoWoS capacity changes plus capex guidance: the point at which 130K wafer/month is reached marks the revenue ceiling for NVIDIA and AMD.

3. ASML EUV / High-NA quarterly shipments: 60+/year marks the logic capex peak. Cumulative High-NA of 20 units marks the full ramp of sub-2nm.

4. Broadcom AI semi backlog: the trajectory of the $73B figure. The real-speed indicator for ASIC eating into GPU.

5. China self-sufficiency by domain: DRAM self-sufficiency at 10%+ means general-memory threats become real. NAND self-sufficiency at 15%+ creates risk of a NAND ASP collapse. Logic self-sufficiency moves slowly (a hard ceiling without EUV).

[16.2 Five things retail investors need to watch]

1. SK Hynix quarterly HBM revenue and its correlation with operating margin: 70%+ OPM sustained means the peak continues. Falling below 50% signals the late stage of the cycle.

2. Hami Semiconductor quarterly revenue YoY: Q1 2026 -65.5% turning to a Q2 recovery is the signal that HBM4 production ramp has begun in earnest.

3. NVIDIA, AMD, and Broadcom data center revenue (quarterly): the trend of all three combined is the trend of the entire AI chip market. Moving from $7B/quarter through $10B/quarter marks the peak zone.

4. TSMC quarterly revenue plus HPC mix: 61% sustained is the AI cycle peak. Reaching 70%+ raises the risk of over-concentration.

5. DDR5 spot price trends (TrendForce, DRAMeXchange): a sustained surge pressures server, PC, and smartphone operating margins. A reversal signals the late stage of the cycle.

16.3 Weekly Watch Dashboard — 7 Indicators

IndicatorCurrentNormalCautionDangerStatus
HBM TTM revenue growth+200%++50%++20~50%<+20%Normal
TSMC CoWoS capacity (mo)~80KRisingFlatDecliningNormal
ASML EUV quarterly ships~1515+10-15<10Normal
Hami Semi revenue YoY-65.5%+30%+-10~+30%<-30%Danger
SK Hynix OPM72%50%+30-50%<30%Normal
NVIDIA DC QoQ+20%++10%+0~+10%DecliningNormal
DDR5 8Gb spot price$3+$2-3$1.5-2<$1.5Caution

Composite verdict

1 danger · 1 caution · 5 normal

→ "Late-cycle stability" signal

Normal/Caution/Danger color cells + composite verdict. 3+ in danger signals 'peak passed'; 5+ in normal signals 'late-cycle stability'.

Watch Dashboard — 7 Indicators

IndicatorNow (May 2026)NormalCautionDangerSource
1. HBM TTM revenue growth+200%++50%++20-50%<+20%Counterpoint/TrendForce
2. TSMC CoWoS capacity (mo)~80KIncreasingFlatDecliningTSMC IR
3. ASML EUV quarterly ships~1515+10-15<10ASML IR
4. Hami quarterly revenue YoY-65.5%+30%+-10~+30%<-30%Hami IR
5. SK Hynix OPM72%50%+30-50%<30%SK Hynix IR
6. NVIDIA DC QoQ+20%++10%+0-+10%DecliningNVIDIA Form 8-K
7. DDR5 8Gb spot price$3+$2-3$1.5-2<$1.5TrendForce

As of May 2026, 5 are normal, 1 in danger (Hami — temporary generation gap). Currently in peak cycle zone.

18

17. Upcoming Semi Deals + 30-Term Glossary

[17.1 Six upcoming deals]

If this piece covered the macro, other deal-story notes will unpack the anatomy of individual transactions.

1. BlackRock × HPS Investment Partners ($12B, closed July 2025) — asset management entering semiconductor capex financing alongside private credit.

2. SK Hynix HBM4 production ramp (Q2-Q3 2026) — the decisive inflection of the HBM generation transition. A signal for the Hami Semiconductor TC Bonder recovery.

3. Intel × TSMC US fab collaboration possibility — Intel 18A production plus the TSMC Arizona fab. US government subsidies plus geopolitics.

4. Samsung SF2 NVIDIA win possibility — to be decided in H2 2026. An essential inflection point for Korean foundry.

5. CXMT (China's ChangXin Memory) IPO — an attempt in 2026-2027. The signal that Chinese memory has begun mobilizing capital at scale.

6. Hami Semiconductor HBM4 TC Bonder wins — SK Hynix order recovery in Q2-Q3 2026. The inflection point of the Korean supplier cycle.

17.2 Glossary — 30 Core Terms

Memory (8): HBM (stacked memory), DDR5 (5th-gen DRAM), LP5/LPDDR5 (mobile low-power DRAM), NAND (flash), TC Bonder (HBM stacking tool), HBM substrate (HBM interposer-side board), TSV (through-silicon via), MR-MUF (SK Hynix's package process). Foundry and packaging (7): Fabless (design-only), Foundry (contract manufacturing), Node N3/N2/A14 (circuit width), CoWoS (TSMC's leading-edge packaging), CoWoS-S/L (variants), ABF substrate (the board atop CoWoS, by Ajinomoto), OSAT (back-end outsourcing). Lithography and logic (6): EUV (extreme ultraviolet lithography), High-NA EUV (next generation at NA 0.55), DUV (conventional lithography), PR (photoresist), HPSP high-pressure annealing (circuit healing), ALD (atomic layer deposition). Design (4): GPU (NVIDIA, AMD), ASIC (Google TPU, AWS Trainium), HPC (data center), AI accelerator (umbrella term). Industry and policy (5): Sobujang (Korean term for materials/components/equipment), EDA (Cadence, Synopsys), IP (ARM, RISC-V), CHIPS Act (US subsidy law), China Big Fund ($47.5B Phase 3).

One Final Sentence

The AI semiconductor cycle is not the tail end of a 17-year boom; it is a cycle that has just entered its fourth year of full-scale production. Inside it, the real alpha is embedded at three bottlenecks — HBM, CoWoS, and EUV. Korea's fate splits depending on which bottleneck it occupies. The proof: Hami Semiconductor's -65.5% and SK Hynix's 72% operating margin happened in the same quarter.

Written byUpdated 50 min read

Was this helpful?

Share it with someone

References

  1. [1]SK하이닉스. 2026 1분기 실적 발표 (영업이익률 72%, 매출 52.58조원). SK Hynix IR, 2026.04.
  2. [2]한미반도체. 2026 1분기 잠정공시 (매출 509억원, -65.5% YoY). 한미반도체 IR, 2026.04.
  3. [3]솔브레인 / HPSP. 2026 1분기 실적 / 연간 가이던스. 각사 IR + 증권사 리포트, 2026.04-05.
  4. [4]WSTS / SIA. Global Semiconductor Sales 2025-2026 ($791.7B → $1T). WSTS Autumn 2025 / SIA Feb 2026, 2025-2026.
  5. [5]NVIDIA / AMD / Broadcom. Data Center Revenue Combined Q1 2024 → Q1 2026 (~$27B → ~$70B). 각사 Form 8-K + IR, 2024-2026.
  6. [6]WSTS. 2026 Autumn Forecast ($1T milestone projection). WSTS, 2025-11.
  7. [7]NVIDIA. Form 8-K, FY2026 Q4 Data Center Revenue $35.6B. SEC EDGAR, 2026.02.↗
  8. [8]Broadcom. Q1 FY2026 AI Semi Revenue $8.4B (+106% YoY), Backlog $73B. Broadcom IR, 2026.03.↗
  9. [9]TSMC. Q1 2026 Revenue $35.9B, HPC mix 61%, 2026 capex $52-56B. TSMC IR, 2026.04.↗
  10. [10]Counterpoint Research / TrendForce. HBM Quarterly Share 2023 Q1 - 2026 Q1. Counterpoint, TrendForce, 2026.
  11. [11]Micron. FY Q2 2026 Revenue $23.86B (+196% YoY), Q3 Guidance $33.5B, Margin 81%. Micron IR, 2026.03.↗
  12. [12]TSMC / SemiAnalysis. CoWoS Capacity 2023-2026 (15K → 130K wafer/month). TSMC IR + SemiAnalysis (Dylan Patel), 2024-2026.
  13. [13]ASML. Q1 2026 Revenue €8.8B, EUV €4.1B, EPS €7.15, 2026 guidance €36-40B. ASML IR, 2026.04.↗
  14. [14]SemiAnalysis (Dylan Patel). Blackwell delay analysis — CoWoS capacity as the real variable. SemiAnalysis, 2024-2025.
  15. [15]US BIS (Bureau of Industry and Security). Export controls timeline (Oct 2022, Oct 2023, Oct 2024). US Department of Commerce BIS, 2022-2024.↗
  16. [16]China National IC Industry Investment Fund. Phase 3 announcement, ~$47.5B (RMB 344B). China State Council, 2024.05.
  17. [17]SMIC. Q1 2026 revenue $2.5B; Huawei Kirin 9000S on SMIC N+2 7nm. SMIC IR + media, 2023-2026.
  18. [18]TrendForce. DRAM/NAND ASP trends 2023-2026 (DDR5 8Gb spot price). TrendForce / DRAMeXchange, 2023-2026.
  19. [19]Shin-Etsu / SUMCO / JSR / TEL / Disco / Lasertec. 각 사 2026 Q1 IR (일본 소부장). 각사 IR, 2026.
  20. [20]AMAT / LAM / KLA. Q1 2026 IR (미국 장비 3강). 각사 IR, 2026.
  21. [21]솔브레인 / 동진쎄미켐 / SK실트론 / 원익머트리얼즈. 한국 소재 강자 2025-2026 사업보고서. DART, 2025-2026.
  22. [22]원익큐엔씨. 2025 매출 9,436억 / 2026 Q1 매출 2,562억. 원익큐엔씨 IR + 디지털투데이 잠정공시, 2026.04.
  23. [23]한미반도체. TC Bonder 글로벌 점유율 약 70% (보도 기반 추정). 더벨, 인베스트조선, 2024-2026.
  24. [24]HPSP. 고압 어닐링 글로벌 95%+ 사실상 독점. HPSP IR + 증권사 리포트, 2024-2026.
  25. [25]심텍. HBM substrate 2025 매출 약 1.3조 (SK하이닉스 의존). 심텍 IR, 2025.
  26. [26]테크윙. HBM 테스트 핸들러 2025 매출 약 2,800억. 테크윙 IR, 2025.
  27. [27]Jim Chanos / Kynikos Associates. Memory cycle bear thesis — 'memory is forever cyclical'. Bloomberg, FT interviews, 2024-2026.
  28. [28]Hans Mosesmann (Rosenblatt). AI capex bubble warning. Rosenblatt Securities, 2024-2026.
  29. [29]ai-semi-value-chain-research.md. 내부 리서치 fact base (180+ 출처 카탈로그). deal-story internal, 2026-05-31.(전체 출처 카탈로그는 research .md §15 참조)
  30. [30]한국은행 / KIET / 반도체산업협회. 한국 반도체 수출·매출 통계. 한국은행 ECOS, KIET, KSIA, 2025-2026.
  31. [31]TSMC. 2026 Q1 earnings call / Q1 management report. TSMC IR, 2026-04-16.
  32. [32]SK하이닉스. 2026 Q1 conference call 전문. 디일렉, 2026-04-23.
  33. [33]Samsung Electronics. 2026 Q1 conference call. Samsung IR, 2026.
  34. [34]Micron. Q2 FY2026 earnings prepared remarks. investors.micron.com, 2026.
  35. [35]Micron. Q2 FY2026 earnings call transcript. Investing.com, 2026.
  36. [36]ASML. Q1 2026 earnings. TIKR.com, Stockopine, 2026.
  37. [37]NVIDIA. FY2026 Form 8-K Q4. SEC EDGAR, 2026.
  38. [38]Broadcom. Q1 FY2026 earnings. Futurum, 2026.
  39. [39]AMD. Q1 2026 earnings. AMD IR, 2026.
  40. [40]SMIC. Q1 2026 earnings. TrendForce, TipRanks, 2026-05-15.
  41. [41]한미반도체. 1Q26 잠정실적 + 2Q 가이던스. businesspost, ER이코노믹리뷰, 2026.
  42. [42]솔브레인. 1Q26 잠정. 신한투자증권 리포트, leadeconomy, 2026.
  43. [43]HPSP. 1Q26 Preview. 대신증권, 2026-04-22.
  44. [44]동진쎄미켐. 분기 실적. siglab.kr, thelec.kr, 2026.
  45. [45]SFA반도체. 4Q25 / 1Q26. leadeconomy, theviewers, 2025-2026.
  46. [46]원익IPS / 주성엔지니어링. 1Q26 실적. CBC뉴스, 헤럴드, 2026.
  47. [47]세메스. 1Q26 실적. 헤럴드, 2026.
  48. [48]예스티. 1Q26 실적. sptatimeskorea, 2026.
  49. [49]하나마이크론. OSAT 글로벌 8위. PRESS9, TrendForce, 2026.
  50. [50]SK실트론. IR. sk-siltron.com, 2026.
  51. [51]Applied Materials. Q1 FY2026 Form 8-K. SEC EDGAR, 2026.
  52. [52]Applied Materials. Q2 FY2026 Form 8-K. SEC EDGAR, 2026.
  53. [53]Lasertec. FY2025 결산. Lasertec IR, 2025.
  54. [54]TSMC. 분기별 capex disclosure. TSMC IR, 2025-2026.
  55. [55]NVIDIA. Blackwell B200/GB200 GTC 2024 announcement. NVIDIA, 2024.
  56. [56]AMD. MI350 launch. AMD blog, 2025-06.
  57. [57]AMD. MI400 / Helios roadmap. datacenterdynamics, AMD blog, 2025-2026.
  58. [58]ASML. Capital Markets Day 2024 / 2025. ASML, 2024-2025.
  59. [59]Intel. Panther Lake / 18A launch CES 2026. Intel newsroom, 2026.
  60. [60]Samsung. HBM4E 샘플 발표. techtimes, 2026-05-30.
  61. [61]Samsung. SF2P 70% 수율. wedbush, financialcontent, 2026-01-30.
  62. [62]Samsung. HBM3E 12-Hi NVIDIA 퀄. sammobile, kedglobal, 2025-09.
  63. [63]Samsung. HBM4 NVIDIA 퀄 통과. kedglobal, 2025-12.
  64. [64]SK하이닉스. HBM4 양산 가이던스. SK하이닉스 IR, 2026.
  65. [65]TSMC. CoWoS 가이던스. TSMC IR, TrendForce 분석, 2025-2026.
  66. [66]NVIDIA. Rubin R200 출하 가이던스. TrendForce, 2026-04.
  67. [67]CXMT. DDR5 출시. wccftech, hostzealot, scmp, 2025-2026.
  68. [68]YMTC. 232L mass production. techinsights, globaltimes, tomshardware, 2025-2026.
  69. [69]Naura / AMEC. 2024 실적. semiconductorinsight, tomshardware, 2024.
  70. [70]Broadcom. $73B backlog. 24/7 Wall St., Futurum, 2026.
  71. [71]SIA. Global Semiconductor Sales 2025. semiconductors.org, 2026-02.
  72. [72]WSTS. Semiconductor Market Forecast Spring 2025. wsts.org, 2025.
  73. [73]WSTS. Autumn 2025 Forecast ($795.6B, +26.2%). wsts.org, 2025.
  74. [74]SIA. 2025 Factbook. semiconductors.org, 2025.
  75. [75]TrendForce. HBM share 분기 보고서. TrendForce, 2025-2026.
  76. [76]Counterpoint Research. HBM revenue Q3 2025 share. Counterpoint Research, 2025.
  77. [77]TrendForce. CoWoS capacity 2025-2026. TrendForce, 2025-2026.
  78. [78]TrendForce. NVIDIA Rubin delay 2026. TrendForce, 2026-04-08.
  79. [79]TrendForce. AMD MI350 N3P + Samsung/Micron HBM3E dual. TrendForce, 2025-06-13.
  80. [80]TrendForce. Samsung Google TPU HBM3E 60%. TrendForce, 2025-12-01.
  81. [81]SEMI. WFE market size 2025 ($41.4B). SEMI, 2025.
  82. [82]SemiAnalysis (Dylan Patel). AI Capacity Constraints CoWoS HBM. SemiAnalysis, 2024-2026.
  83. [83]SemiAnalysis. Scaling the Memory Wall: HBM Roadmap. SemiAnalysis, 2024-2026.
  84. [84]SemiAnalysis. The Great AI Silicon Shortage. SemiAnalysis, 2024-2026.
  85. [85]SemiAnalysis. Lam Research, TEL, JSR EUV $5B+ market. SemiAnalysis, 2024-2026.
  86. [86]Bernstein. Semiconductor sector reports. Bernstein, 2025-2026.
  87. [87]Morgan Stanley (Joseph Moore). TSMC bull report. Morgan Stanley, 2025-2026.
  88. [88]KBRA. DLD (참고). KBRA, 2025-2026.
  89. [89]IC Insights. 분기 추정. IC Insights, 2025-2026.
  90. [90]Yole Group. YMTC 232L 분석. Yole Group, 2025-2026.
  91. [91]TechInsights. YMTC 첫 mass production 3D NAND. TechInsights, 2025-2026.
  92. [92]Mordor Intelligence. Photoresist Market 2026 (Japan 75%). Mordor Intelligence, 2026.
  93. [93]Fountyl Tech. Japanese EUV photoresist monopoly. Fountyl Tech, 2025-2026.
  94. [94]McKinsey. Semiconductor outlook 2025. McKinsey, 2025.
  95. [95]Gartner. WSTS supplement. Gartner, 2025-2026.
  96. [96]CSIS. Reports on China semiconductor. CSIS, 2024-2026.
  97. [97]Congress.gov CRS. R48642 U.S. Export Controls and China: Advanced Semiconductors. Congress.gov, 2024-2026.
  98. [98]GAO. GAO-25-107386 Export Controls Compliance. GAO, 2025.
  99. [99]US BIS. Press releases (2022.10, 2023.10, 2024.10, 2024.12). bis.gov, 2022-2024.↗
  100. [100]WilmerHale. BIS 2024.12 Additional Restrictions (alert). WilmerHale, 2024-12.
  101. [101]Caixin Global. China Big Fund III 등록. Caixin Global, 2024-05-28.
  102. [102]Tom's Hardware. Big Fund III spending start. tomshardware, 2024-12.
  103. [103]Wikipedia. China Integrated Circuit Industry Investment Fund. Wikipedia, 2024-2026.
  104. [104]SCMP. Big Fund III $47.5B. SCMP, 2024-05.
  105. [105]Eurasia Review. Big Fund III long game. Eurasia Review, 2025-12.
  106. [106]TechInsights / SemiAnalysis. SMIC 7nm Mate 60 Pro analysis. techinsights, semianalysis, 2023.
  107. [107]DigiTimes. CXMT 16nm DRAM production. digitimes, 2025-2026.
  108. [108]Tom's Hardware. CXMT HBM3 production 2026. tomshardware, 2026.
  109. [109]Tom's Hardware. Naura/AMEC market share. tomshardware, 2025-2026.
  110. [110]SemiWiki. Chinese memory ramp 2026. SemiWiki, 2026.
  111. [111]KIET. 반도체산업협회 보고서. KIET, 2025-2026.
  112. [112]KOSEMI / KSIA. 한국 반도체산업협회. KSIA, 2025-2026.
  113. [113]한국은행. 반도체 통계 (분기별). 한국은행 ECOS, 2025-2026.
  114. [114]DART. 전자공시시스템 (각 한국 소부장 분기보고서). DART, 2025-2026.
  115. [115]산업통상자원부. K-반도체 전략 (2021, 갱신). 산업통상자원부, 2021-2026.
  116. [116]과기정통부. 반도체 R&D 자료. 과기정통부, 2025-2026.
  117. [117]더벨. 반도체 인수합병 기사. 더벨, 2025-2026.
  118. [118]인베스트조선. 펀딩 / IR 분석. 인베스트조선, 2025-2026.
  119. [119]이데일리. 반도체 기사. 이데일리, 2025-2026.
  120. [120]서울경제. 반도체 (한미반도체·HPSP 등). 서울경제, 2025-2026.
  121. [121]매일경제. 반도체 (HBM, TSMC). 매일경제, 2025-2026.
  122. [122]한국경제. 반도체 (솔브레인, 동진). 한국경제, 2025-2026.
  123. [123]디일렉. 한국 반도체 deep. 디일렉, 2025-2026.
  124. [124]전자신문. 동진쎄미켐 EUV PR 국산화. 전자신문, 2025-2026.
  125. [125]EBN. SK하이닉스 EUV PR 국산화. EBN, 2025-2026.
  126. [126]녹색경제신문. 동진쎄미켐 EUV PR. 녹색경제신문, 2025-2026.
  127. [127]디지털타임스 / 헤럴드경제. 반도체 기사. 디지털타임스, 헤럴드경제, 2025-2026.
  128. [128]시그널 리포트 / 인포스탁. 시장 리포트. 시그널, 인포스탁, 2025-2026.
  129. [129]알파스퀘어. 각사 주가/실적. 알파스퀘어, 2025-2026.
  130. [130]와이즈리포트. 각사 기업현황. 와이즈리포트, 2025-2026.
  131. [131]CNBC. SK하이닉스 record Q1. CNBC, 2026-04-23.
  132. [132]Yahoo Finance. SK Hynix Q1 2026 highlights. Yahoo Finance, 2026.
  133. [133]Reuters. Semiconductor analysis. Reuters, 2024-2026.
  134. [134]Bloomberg. Broadcom Google TPU deal. Bloomberg, 2026-04.
  135. [135]Nikkei Asia. Big Fund III article. Nikkei Asia, 2024-05.
  136. [136]Korea Herald. Nvidia 16-layer HBM. Korea Herald, 2025-2026.
  137. [137]KED Global. Samsung NVIDIA HBM articles. KED Global, 2025-09 / 2025-10 / 2025-12.
  138. [138]Tom's Hardware. 다수 — YMTC, CXMT, 18A 등. tomshardware, 2024-2026.
  139. [139]Tom's Hardware. Chinese semiconductor production equipment makers. tomshardware, 2025-2026.
  140. [140]Tom's Hardware. Deeper look at tightened chipmaking supply chain. tomshardware, 2025-2026.
  141. [141]SamMobile. Samsung HBM4 NVIDIA tests. SamMobile, 2025-2026.
  142. [142]Wccftech. CXMT DDR5. Wccftech, 2025-2026.
  143. [143]VideoCardz. AMD MI350/MI400. VideoCardz, 2025-2026.
  144. [144]AnandTech. 분석 (시리즈). AnandTech, 2024-2026.
  145. [145]SemiWiki. Forum threads. SemiWiki, 2024-2026.
  146. [146]The Globe and Mail. TSMC earnings. The Globe and Mail, 2026.
  147. [147]Bitget News. TSMC / ASML 분석. Bitget News, 2025-2026.
  148. [148]BigGo Finance. TSMC / SMIC / SK Hynix earnings 분석. BigGo Finance, 2025-2026.
  149. [149]FinancialContent. Samsung 2nm, Intel 18A. FinancialContent, 2025-2026.
  150. [150]The Korea Herald. NVIDIA HBM. The Korea Herald, 2025-2026.
  151. [151]Korea Times. Semiconductor articles. Korea Times, 2025-2026.
  152. [152]China Daily. SMIC / YMTC 영문 기사. China Daily, 2025-2026.
  153. [153]ZDNet Korea. 반도체. ZDNet Korea, 2025-2026.
  154. [154]AnandTech. NVIDIA Blackwell deep dive. AnandTech, 2024-2026.
  155. [155]Datacenterdynamics. AMD Helios. Datacenterdynamics, 2025-2026.
  156. [156]Phoronix. AMD MI350P review. Phoronix, 2025-2026.
  157. [157]KAD. AMD MI350 specs. KAD, 2025-2026.
  158. [158]Investing.com. 다수 earnings transcripts. Investing.com, 2025-2026.
  159. [159]The Asia Business Daily. HPSP (영문). The Asia Business Daily, 2025-2026.
  160. [160]Singularity Moments. NVIDIA Blackwell 2026. Singularity Moments, 2026.
  161. [161]SemiAnalysis. Newsletter (CoWoS, HBM, Memory Wall, AI Silicon Shortage 등 5+ 글). SemiAnalysis newsletter, 2024-2026.
  162. [162]Dwarkesh Patel. Podcast — Dylan Patel 인터뷰. Dwarkesh podcast, 2024-2026.
  163. [163]Latent Space. Podcast — Dylan Patel. Latent Space podcast, 2024-2026.
  164. [164]Stratechery. Semiconductor analyses. Stratechery, 2024-2026.
  165. [165]Pacific Economy Lab. 반도체 cycle. Pacific Economy Lab, 2024-2026.
  166. [166]JaredWatkins. Lasertec deep. JaredWatkins, 2024-2026.
  167. [167]Oreate AI. Lasertec EUV monopoly. Oreate AI, 2024-2026.
  168. [168]MatrixBCG. Lasertec series. MatrixBCG, 2024-2026.
  169. [169]Ferrante Capital. TSMC Q1 readthrough. Ferrante Capital, 2026.
  170. [170]MacroMicro. TSMC Q1. MacroMicro, 2026.
  171. [171]Tech-insider. Broadcom AI / Micron Q2 2026. Tech-insider, 2026.
  172. [172]Oplexa. Broadcom 100B target. Oplexa, 2025-2026.
  173. [173]24/7 Wall St.. Broadcom, Google TPU, China equipment. 24/7 Wall St., 2025-2026.
  174. [174]Caixin Global. China chip articles. Caixin Global, 2024-2026.
  175. [175]The Information. AI capex articles. The Information, 2024-2026.
  176. [176]Stockopine. ASML Q1 2026. Stockopine, 2026.
  177. [177]Tradingview. Applied Materials AI backbone. Tradingview, 2025-2026.
  178. [178]Kavout. TEL vs U.S. competitors. Kavout, 2025-2026.
  179. [179]KoalaGains. AMAT stock analysis. KoalaGains, 2025-2026.
  180. [180]Winbuzzer. Intel 18A roadmap. Winbuzzer, 2025-2026.
  181. [181]Wedbush. Samsung 2nm yield. Wedbush, 2026.
  182. [182]Design-Reuse. Samsung 2nm/4nm. Design-Reuse, 2025-2026.
  183. [183]Notebookcheck. YMTC NAND. Notebookcheck, 2025-2026.
  184. [184]Techpowerup. CXMT DDR5 80% yield. Techpowerup, 2025-2026.
  185. [185]Gizmochina. Big Fund III. Gizmochina, 2024-2026.
  186. [186]TheCorner. Big Fund III. TheCorner, 2024-2026.
  187. [187]Techwireasia. Chinese memory expansion 2026. Techwireasia, 2026.
  188. [188]SemiconductorX. Intel Foundry spotlight. SemiconductorX, 2025-2026.
  189. [189]Eurasia Review. Big Fund III analysis. Eurasia Review, 2025-2026.
  190. [190]ResearchGate. 반도체 논문 (참고). ResearchGate, 2024-2026.
  191. [191]Yole Group. Strategy insights. Yole Group, 2024-2026.
  192. [192]Counterpoint Research. Blog (HBM share). Counterpoint Research, 2024-2026.
  193. [193]IC Insights. 분기 추정. IC Insights, 2024-2026.
  194. [194]Omdia. DRAM/NAND 추적. Omdia, 2024-2026.
  195. [195]DRAMeXchange / TrendForce. 가격. DRAMeXchange, TrendForce, 2024-2026.
  196. [196]SemiAnalysis. AI Expansion - Supply Chain Analysis. SemiAnalysis, 2024-2026.
  197. [197]Antoine Buteau. Lessons from Dylan Patel. Antoine Buteau, 2024-2026.
  198. [198]Useluminix. AI compute bottlenecks deep research. Useluminix, 2024-2026.
  199. [199]Globalsemiresearch substack. CoWoS scaling. Globalsemiresearch, 2024-2026.
  200. [200]Longyield substack. TSMC analysis. Longyield, 2024-2026.
  201. [201]Fusionww blog. CoWoS HBM 2-3nm. Fusionww, 2024-2026.
  202. [202]Stockanalysis. SMIC overview. Stockanalysis, 2024-2026.
  203. [203]Tradingkey. ASML Q1 preview. Tradingkey, 2026.
  204. [204]The Globe and Mail. TSMC AI boom. The Globe and Mail, 2025-2026.
  205. [205]Reportes.valuates. Silicon Wafer market 2026. Reportes.valuates, 2026.
  206. [206]Semiconductorinsight. Silicon Wafer market. Semiconductorinsight, 2025-2026.
  207. [207]WaferPro. Top 5 Silicon Wafer companies. WaferPro, 2025-2026.
  208. [208]Mordor Intelligence. Semiconductor Silicon Wafer share. Mordor Intelligence, 2025-2026.
  209. [209]FutureMarketInsights. Semiconductor Wafer. FutureMarketInsights, 2025-2026.
  210. [210]Markets and Markets. Photoresist companies. Markets and Markets, 2025-2026.
Back to Notes
AI Semiconductor Value Chain 2026 — Three Bottlenecks and Korea's Position | Notes | Deal Story | 딜스토리