Korea Semiconductors on Korea Invest Insights

Sam-Ha-Ma Parity Follow-Up: Micron Deserves a Premium, But Korean Memory Looks Too Cheap

Fri, 05 Jun 2026 12:34:00 +0900

Follow-up context: this note follows Sam-Ha-Ma parity, SK hynix vs Micron, and Samsung HBM4E 12-high sample. Related hubs: AI HBM hub, Korea Daily Market Hub, and Korean stocks for foreign investors.

TL;DR

This follow-up expands Sam-Ha-Ma from Samsung, SK hynix and Micron into a broader AI chip and memory forward P/E map including ARM, Marvell, NVIDIA, AMD and SanDisk.
Samsung and SK hynix sit in the lowest P/E cluster of the AI chip basket. That is not automatically bullish. The market is still asking whether memory earnings are peak earnings.
Micron deserves part of its premium: U.S. listing, dollar exposure, AI-memory scarcity, HBM4, SOCAMM2 and data-center SSD positioning all matter.
But Samsung and SK hynix have not suffered an EPS breakdown. The relative discount is mostly the result of Micron re-rating first as the U.S. AI-memory proxy.
Relative-value ranking remains: SK hynix first, Samsung as the larger HBM catch-up option, Micron as the benchmark rather than the cheapest stock.

Core Takeaway

Micron became expensive first as the U.S.-listed AI-memory proxy. Samsung Electronics and SK hynix did not see EPS damage, yet their relative P/E fell. This follow-up shows that the gap appears not only inside Sam-Ha-Ma, but also against the broader AI chip basket.

AI Chip And Memory Map

The map uses Thesis OS local close prices and a mix of BusinessQuant and FnGuide EPS consensus. It should be read as a rank and cluster tool, not as a single-decimal absolute valuation model. Fiscal-year conventions, adjusted EPS definitions and source vendors differ across U.S. and Korean stocks.

Ticker	2026E P/E	2027E P/E	2028E P/E	2026-2028 EPS CAGR	Read
ARM	195.7x	146.3x	116.7x	29.5%	Highest platform premium
MRVL	85.1x	67.8x	50.7x	29.5%	Custom silicon and networking premium
MU	28.1x	23.1x	20.3x	17.6%	U.S.-listed AI-memory proxy premium
SNDK	49.2x	20.5x	24.1x	43.0%	High-beta NAND/storage recovery
Samsung Electronics	8.2x	6.3x	6.2x	14.8%	Lowest large-cap AI-memory P/E cluster
SK hynix	7.7x	5.7x	5.6x	17.1%	HBM leader, still in the lowest P/E cluster
NVIDIA	29.2x	21.2x	18.9x	24.4%	Balanced growth and declining P/E
AMD	83.3x	51.3x	40.1x	44.1%	Strong growth expectation, high P/E

The striking point is simple: Samsung and SK hynix are the lowest P/E names in this AI chip and memory basket.

That can mean either opportunity or a value trap. The key question is whether 2027-2028 EPS holds.

Micron Premium Is Not A Technology Premium

Micron is a very good company in this cycle. It also gives U.S. investors the cleanest large-cap memory proxy.

Micron guided FY3Q26 revenue to $33.5 billion, gross margin to about 81%, and non-GAAP diluted EPS to $19.15. (Micron IR) In its prepared remarks, Micron said it continues to expect DRAM and NAND supply-demand conditions to remain tight beyond calendar 2026. It also described very sharp sequential price increases in DRAM and NAND. (Micron Prepared Remarks)

That supports a premium.

But the premium should not be misread as HBM technology leadership over SK hynix. The cleaner interpretation is:

Micron premium
= U.S. listing premium
+ AI-memory scarcity premium
+ dollar asset premium
+ HBM4 / SOCAMM2 / data-center SSD platform narrative
+ strong FQ3 guidance

Korean Memory Discount Looks Too Wide

Korean memory also deserves some discount.

Samsung is not a pure memory company. It carries smartphones, appliances, displays, foundry and system LSI. It also still has an HBM catch-up discount.

SK hynix is cleaner, but it is still a Korean-listed, won-denominated stock with foreign-access and Korea-discount friction.

Even so, the discount looks wide.

Using the user-provided local run-rate operating profit snapshot:

Company	Earnings base	Annualized OP	Equity value / market cap	Run-rate OP multiple
Samsung Electronics	2Q26E OP about KRW 90tn	about KRW 360tn	about KRW 2,176tn	about 6.0x
SK hynix	2Q26E OP about KRW 66.5tn	about KRW 266tn	about KRW 1,532tn	about 5.8x
Micron	FY3Q26 OP about $25.7bn	about $102.8bn	about $1.137tn	about 11.1x

Micron trades at roughly 1.8-1.9x the Korean memory run-rate operating-profit multiple. Some U.S. premium is fair. But that spread is hard to justify if SK hynix remains the HBM leader and EPS estimates stay intact.

Stock Read

SK hynix

SK hynix is the clearest relative-value name. It has the most visible HBM leadership, a 2028E P/E of 5.6x, and a 2026-2028 EPS CAGR of 17.1%, close to Micron’s 17.6% in the map.

The buy case is not “memory is cheap” in a generic sense. It is that the HBM leader trades at a deeper discount than the U.S. AI-memory proxy should command.

The caveat is flow. Full sizing needs foreign selling to slow, 2Q earnings to confirm the run-rate, and 3Q DRAM/HBM price guidance to hold.

Samsung Electronics

Samsung’s discount is more justified than SK hynix’s discount. It is a conglomerate-like technology platform, not a memory pure-play, and HBM credibility is still being rebuilt.

But the upside path is still meaningful. Samsung does not need to trade at Micron’s multiple. A move from roughly 6x run-rate OP to 7-8x would already matter if HBM4E qualification and DS earnings momentum improve.

Micron

Micron remains the benchmark. It is the U.S.-listed AI-memory proxy and has strong near-term numbers. But after the re-rating, it is no longer the cheapest relative-value expression of the same memory cycle.

Bottom Line

Micron deserves a premium. But the Korean memory discount has likely gone too far if EPS durability holds.

The practical ranking is:

1. SK hynix
- HBM leader
- lowest P/E cluster
- excessive discount versus Micron

2. Samsung Electronics
- more justified discount
- larger HBM catch-up option

3. Micron
- high-quality benchmark
- less relative upside after re-rating

In one sentence:

Memory stocks look cheap again not because earnings collapsed, but because the U.S. AI-memory proxy got expensive first. As long as that gap remains, the Sam-Ha-Ma parity trade is still alive.

Sources And Data

AI Chip & Memory Forward P/E Map: Thesis OS local close prices plus BusinessQuant and FnGuide consensus, as of June 4, 2026.
Micron FY2Q26 results and FY3Q26 guidance: Micron IR
Micron FY2Q26 prepared remarks: Micron Prepared Remarks
Prior notes: Sam-Ha-Ma parity, SK hynix vs Micron

Broadcom Reconfirmed a $100B+ AI Semiconductor Run-Rate for 2027: How Korea Should Read the Value Chain

Fri, 05 Jun 2026 09:30:00 +0900

📚 Connected context This is a follow-up to the Marvell/Broadcom Korea bottleneck preview, the Jensen Huang-Marvell read-through, the laser-source follow-up, and the Samsung-Hynix-Micron parity note. Relevant hubs are the Korea semiconductor value chain hub, AI HBM hub, and AI PCB/substrate hub.

TL;DR

Broadcom’s Q2 FY2026 result is not negative for Korean semiconductors. The better read is that AI custom silicon demand continues to flow into HBM, package substrates, networking, and power-integrity components.

The stock fell because expectations were even higher, not because the underlying AI demand collapsed.

Core View

Broadcom reconfirmed the FY2027 AI semiconductor revenue frame of more than $100B. The market wanted an upward revision, not just confirmation. Korean investors should treat the selloff as an expectation reset and re-check the durability of HBM, FC-BGA, AI networking, silicon capacitors and test/packaging bottlenecks.

The Korean translation is:

Priority	Layer	Korean names to watch	View
1	HBM / server DRAM / eSSD	Samsung Electronics, SK hynix	Core layer. Custom ASIC broadens the HBM customer base beyond NVIDIA GPUs.
2	FC-BGA / Si-Cap / MLCC	Samsung Electro-Mechanics	Most direct non-memory read-through, but valuation discipline matters.
3	AI networking PCB / high-speed boards	Isu Petasys, Korea Circuit and peers	Second-order beta to AI Ethernet and switch ASIC demand.
4	TC bonders / sockets / inspection	Hanmi Semiconductor, ISC, Leeno and peers	Follow-through depends on real orders.

One-line conclusion: Broadcom’s fall is not a bearish signal for Korean AI infrastructure. It is a reset in over-heated expectations. The first name to revisit is Samsung Electronics, followed by SK hynix and Samsung Electro-Mechanics.

What Broadcom Actually Confirmed

[Fact] Broadcom reported Q2 FY2026 revenue of $22.187B, up 48% year over year. GAAP net income was $9.310B, non-GAAP net income was $12.074B, and free cash flow was $10.262B. (Broadcom Q2 FY2026)

[Fact] AI semiconductor revenue was the key. Q2 AI semiconductor revenue was $10.8B, up 143% year over year, and Q3 AI semiconductor revenue is guided to $16.0B, implying more than 200% year-over-year growth. Q3 total revenue guidance is $29.4B. (Broadcom Q2 FY2026)

Q2 AI semiconductor revenue mix = $10.8B / $22.187B = about 48.7%
Q3 AI semiconductor guidance mix = $16.0B / $29.4B = about 54.4%

Broadcom is no longer just a diversified semiconductor/software company with VMware attached. AI custom silicon is approaching half of revenue.

[Fact] Earnings-call summaries indicate Broadcom still frames FY2026 AI semiconductor revenue at about $56B and FY2027 at more than $100B. It also cited Q2 AI semiconductor bookings of more than $30B. (MarketBeat)

The important word is “still.” Broadcom did not introduce a new, higher 2027 frame. It maintained the existing one. That is where the market was disappointed.

Why the Stock Fell

The absolute numbers were strong. The problem was expectation.

First, investors expected more than the Q3 AI semiconductor guide of $16B. Some reports put market expectations around $16.3B to $17.2B. (Reuters)

Second, the FY2027 $100B+ frame was a confirmation, not an upgrade.

Third, margin mix became a debate. The AI custom silicon mix is rising, and investors questioned whether that could pressure gross margins even as operating leverage remains strong.

So the message is not “AI semiconductor demand is rolling over.” It is: AI ASIC leaders are now priced for very large upside, and anything short of another step-up can trigger a reset.

Why It Matters for Korea

Broadcom’s AI revenue is not only custom XPU. It also includes AI networking. More custom AI chips mean more HBM, high-performance packages, power-integrity parts, high-speed boards, test sockets and optical/networking hardware.

[Inference] The custom AI compute stack is widening from NVIDIA GPUs into Google TPU, Broadcom XPU, OpenAI accelerators, Meta MTIA, Anthropic compute and other hyperscaler designs. Korea participates mostly through the physical bottlenecks: memory, substrates, passive components, test and high-speed interconnect.

Korean Stock Read-Through

Samsung Electronics is the balanced read-through. It has HBM catch-up, server DRAM, eSSD/NAND and a long-term foundry/base-die option. The key check is HBM4E/HBM4 customer approval and whether DS earnings estimates keep rising.

SK hynix remains the cleanest HBM exposure. But the investment question is no longer simply “HBM is good.” It is whether the stock is attractive versus Micron and Samsung Electronics after the recent move. That relative-value angle is covered in the Samsung-Hynix-Micron parity note.

Samsung Electro-Mechanics is the strongest non-memory translation. FC-BGA, MLCC and silicon capacitors sit directly under AI ASIC and AI networking growth. The thesis is right, but the stock already reflects a large part of the re-rating, so repeated orders and margin proof matter.

Hanmi Semiconductor, ISC, Leeno and testing/socket names benefit only when the package/test capex turns into orders. Here the right posture is watchlist, not automatic buying.

Execution

The event supports Korean AI infrastructure, but it does not say “buy everything.”

Layer	What to verify
Samsung Electronics	HBM4E/HBM4 approval, DS earnings revision, foreign flow stabilization
SK hynix	Relative PER vs Micron, HBM price/mix, foreign selling pressure
Samsung Electro-Mechanics	Si-Cap repeat orders, package-solution margin, AI customer diversification
High-speed PCB	AI networking orders, ASP, capacity allocation
Packaging/test	Real orders, not only TAM expansion narratives

Final View

Broadcom reconfirmed a large AI ASIC market. The market sold the stock because it wanted an even larger upgrade. For Korea, that distinction matters. The demand signal remains positive, but the investable alpha should be sorted by physical bottleneck and valuation.

The cleanest path is Samsung Electronics as the core, SK hynix on relative-value pullbacks, and Samsung Electro-Mechanics only with valuation discipline and repeated order evidence.

Sources

Sam-Ha-Ma Parity: Samsung and Hynix Look Cheap Again Versus Micron

Wed, 03 Jun 2026 15:10:00 +0900

Follow-up context: this note follows Samsung vs SK hynix forward PER inversion, SK hynix vs Micron, and Samsung HBM4E 12-high sample. Related hubs: AI HBM hub, Korea Daily Market Hub, and Korean stocks for foreign investors.

TL;DR

The relative discount since mid-May is not about a deterioration in Samsung or SK hynix. It is mostly about Micron re-rating faster.
As of June 1, 2026, Samsung/Micron forward PER is 0.82x and SK hynix/Micron is 0.80x.
Samsung is now roughly 20% below its Samsung/Micron average ratio of 1.03x. SK hynix is roughly 13% below its SK hynix/Micron average ratio of 0.93x.
SK hynix is buyable on a first tranche from a relative-value angle, but foreign selling still argues for scaling in.
Samsung has a larger catch-up path if HBM4E/HBM4 credibility and foreign flow improve.

Forward PER Parity

Name	Latest forward PER	Period average PER	Latest vs Micron	Average vs Micron
Samsung Electronics	8.10x	6.43x	0.82x	1.03x
SK hynix	7.90x	5.87x	0.80x	0.93x
Micron	9.84x	6.49x	1.00x	1.00x

Data basis: Research OS local DB, KR FY2026 EPS for Samsung and SK hynix, US NEXTY EPS for Micron, 49 trading days from April 14 to June 1, 2026. Download: daily panel CSV, baseline summary CSV.

What Changed Since Mid-May

Name	May 14 PER	June 1 PER	Change
Micron	7.62x	9.84x	+29.0%
SK hynix	6.75x	7.90x	+17.1%
Samsung Electronics	7.04x	8.10x	+15.0%

Micron’s price rose faster than Korean memory leaders while all three saw EPS estimates improve. That makes the current discount a relative-value gap, not an earnings-damage signal.

Investment Read

SK hynix remains the cleaner HBM leader. At 0.80x Micron, a return to the 0.89x mid-May ratio implies roughly 10% multiple catch-up; a return to the 0.93x post-April average implies roughly 15%.

Samsung is less proven in HBM, but the relative discount is larger. If Samsung returns to its average Samsung/Micron ratio of 1.03x while Micron’s PER holds near 9.84x, Samsung’s implied relative PER is about 10.1x versus the current 8.1x. That is a multiple catch-up path, not a price target.

The main risks are a Micron derating, Samsung HBM qualification delay, persistent foreign selling in Korea, and EPS downgrades.

Bottom Line

SK hynix is the memory leader that is already buyable in a first tranche. Samsung is the larger catch-up candidate if HBM credibility improves. The common condition for sizing up is foreign-flow stabilization.

Sources: Micron FY2Q26 results, Micron prepared remarks.

Marvell and Broadcom Earnings Preview — Korea's AI Bottleneck Check Before the Calls

Sat, 23 May 2026 09:30:00 +0900

Related series: ARM rally and Korea’s next AI bottlenecks / NVIDIA Q1 FY27 and Korean AI infrastructure / Samsung Electro-Mechanics silicon capacitor contract / Samsung Electronics re-rating thesis

Marvell and Broadcom’s next earnings calls are not just U.S. semiconductor events. They are a check on whether the AI infrastructure bottleneck is moving from GPU-only demand into custom ASICs, Ethernet networking, optical links, package substrates, silicon capacitors, high-speed PCBs and test sockets. If the numbers are strong, Korea’s read-through expands from SK hynix and Samsung Electronics HBM into Samsung Electro-Mechanics, FC-BGA, power-integrity components and selected back-end names.

Summary

Broadcom is the larger macro signal. In Q1 FY2026, Broadcom reported revenue of $19.311B, AI revenue of $8.4B, Q2 revenue guidance of $22.0B, and Q2 AI semiconductor revenue guidance of $10.7B. On that guidance, AI semiconductor revenue alone would be 48.6% of total Q2 revenue. Broadcom is no longer just “semis plus VMware.” It is increasingly an AI custom-chip and AI networking company with a VMware cash-flow engine attached. (Broadcom)

Marvell is smaller, but the signal is cleaner. Q4 FY2026 revenue was $2.219B, Q1 FY2027 guidance is $2.4B ±5%, and Q4 data-center revenue was $1.651B, or 74% of revenue. Marvell’s call will test whether custom silicon, electro-optics and CXL/PCIe switching are moving from narrative into revenue acceleration. (Marvell)

For Korean investors, the question is not “Are Marvell and Broadcom AI winners?” The question is whether their comments confirm a broader physical bottleneck: XPU + HBM + Ethernet fabric + optical links + advanced packaging + power integrity.

If that signal is strong, the Korean priority stack becomes:

SK hynix / Samsung Electronics — HBM, server DRAM, LPDDR/SOCAMM.
Samsung Electro-Mechanics — silicon capacitor, FC-BGA, MLCC and package-level power integrity.
High-speed PCB and test sockets — Isu Petasys, Daeduck, Simmtech, TLB, Korea Circuit, ISC, LEENO, TSE, with direct revenue proof required.

The conclusion is simple: Marvell and Broadcom earnings are a test of whether Korea’s AI semiconductor thesis should broaden beyond HBM into ASIC networking, packaging and power-delivery bottlenecks.

1. Dates and Baselines

Item	Marvell	Broadcom
Next earnings	Q1 FY2027, May 27, 2026 at 1:45 p.m. PDT	Q2 FY2026, June 3, 2026 at 5:00 p.m. EDT
Korea time	May 28, 2026 05:45 KST	June 4, 2026 06:00 KST
Latest reported revenue	Q4 FY26 $2.219B	Q1 FY26 $19.311B
Next guidance	Q1 FY27 $2.4B ±5%	Q2 FY26 $22.0B
AI / data-center benchmark	Q4 data-center $1.651B, 74% mix	Q1 AI revenue $8.4B, Q2 AI semiconductor $10.7B guide

Marvell’s Q1 guide implies roughly +8.2% QoQ growth at the midpoint. Broadcom’s Q2 AI semiconductor guide implies +27.4% QoQ from Q1 AI revenue. These are the two numbers to watch first.

2. Broadcom — The Bigger Macro Signal

The key Broadcom question is not only whether AI semiconductor revenue beats $10.7B. The more important question is whether AI networking is rising inside the AI revenue mix.

Broadcom’s AI infrastructure exposure includes custom XPUs, Ethernet switches, NICs, SerDes, DSPs, retimers and optical connectivity. As clusters expand, the bottleneck shifts from individual accelerators to the fabric connecting chips, racks and data centers.

Broadcom checkpoint	Why it matters	Korea read-through
Q2 AI semiconductor revenue above $10.7B	Confirms custom-chip demand strength	HBM, FC-BGA, test, package components
Rising AI networking share	Confirms Ethernet and optical bottlenecks	High-layer PCB, low-loss materials, sockets
Google TPU and long-duration customer visibility	Extends the custom-ASIC cycle	HBM customer diversification, packaging duration
Margin defense	Revenue growth must convert into profit	Supplier pricing power check
Packaging and supply-chain comments	Shows where the physical constraint sits	Samsung Electro-Mechanics, memory, substrates, test

Broadcom’s OpenAI collaboration is also important. The companies announced a 10GW custom AI accelerator collaboration, with deployment expected to begin in the second half of 2026 and run through the end of 2029. The system includes accelerators as well as Ethernet, PCIe and optical connectivity. (Broadcom OpenAI)

3. Marvell — The Cleaner Connectivity Signal

Marvell’s market cap and revenue base are smaller than Broadcom’s, but it is a cleaner signal for the “connectivity bottleneck” thesis.

The call should answer four questions:

Marvell checkpoint	Why it matters	Korea read-through
Q1 FY27 revenue vs $2.4B guide and Q2 guide	Confirms FY27 acceleration	AI ASIC and interconnect demand durability
Custom-silicon design wins	Shows hyperscaler ASIC diversification	HBM, packaging and test demand broadening
Electro-optics / DSP / 1.6T-3.2T roadmap	Confirms data-movement bottleneck	Limited direct Korea exposure; long-term option
XConn / CXL / PCIe	Connects to memory pooling	CXL memory, SOCAMM, server DRAM
NVLink Fusion collaboration	Shows custom ASIC and NVIDIA ecosystems can coexist	HBM diversification, test and package complexity

In April 2026, Marvell announced the Polariton acquisition to strengthen its optical roadmap toward 3.2T and beyond. As AI workloads expand, data-center bandwidth demand rises and optical links become a power-efficiency problem, not only a speed problem. (Marvell Polariton)

Marvell’s XConn acquisition points to PCIe/CXL switching and scale-up fabric. That matters for Korean memory because CXL and memory pooling broaden the hierarchy from “HBM next to GPU” into rack-level memory architecture. (Marvell XConn)

4. Korea Translation — What Comes After HBM?

The first-order Korea frame has been:

AI = NVIDIA GPU = HBM = SK hynix.

That remains true, but it is no longer enough. If Broadcom and Marvell print strong numbers, the frame broadens:

AI = GPU + custom ASIC + HBM + Ethernet + optics + packaging + power integrity + test.

Memory Still Comes First

Custom ASICs may compete with NVIDIA GPUs for some workloads, but they do not remove the memory bottleneck. Google TPUs, Broadcom XPUs, OpenAI accelerators and Marvell custom silicon all need high-bandwidth memory. SK hynix remains the cleaner exposure. Samsung Electronics has a larger option set, but needs HBM4, HBM4E and customer qualification proof.

Samsung Electro-Mechanics Is the Cleanest Non-Memory Bottleneck

Samsung Electro-Mechanics has three relevant exposures: FC-BGA, MLCC and silicon capacitors. The company announced an approximately KRW 1.5T silicon capacitor supply contract for 2027-2028 and described the product as a component used inside high-performance packages such as AI server GPUs and HBM to stabilize power supply. (Samsung Electro-Mechanics)

This is the cleanest Korean non-memory read-through from AI ASIC and AI networking. As chips get larger and more power-dense, power integrity becomes a package-level constraint.

PCB and Test Need Proof

Isu Petasys, Daeduck, Simmtech, TLB and Korea Circuit are plausible high-speed board and module substrate candidates. ISC, LEENO and TSE are plausible test-socket candidates. But this group needs proof: customer qualification, AI-related revenue, ASP uplift and margin preservation.

5. Practical Watchlist

Priority	Layer	Korean names	View
1	HBM / server DRAM / SOCAMM	SK hynix, Samsung Electronics	Core layer. Custom ASICs expand the customer base
2	Silicon capacitor / FC-BGA	Samsung Electro-Mechanics	Clearest non-memory bottleneck with an official large contract
3	High-speed PCB / low-loss materials	Isu Petasys, Daeduck, Simmtech, TLB, Korea Circuit, Doosan Electronics BG	Positive if AI networking grows; direct proof required
4	Test sockets / inspection	ISC, LEENO, TSE, Intekplus	Benefits from ASIC SKU growth and I/O complexity
5	Optical interconnect	Korean optical modules and materials candidates	Directionally right, but direct Korea exposure is less clear

6. What To Listen For

For Broadcom, bullish lines would be:

AI semiconductor revenue tracking above $10.7B.
AI networking share rising.
Better visibility into 2027.
Supply chain and packaging constraints under control.

For Marvell, bullish lines would be:

Q1 FY27 above the $2.4B midpoint and Q2 guidance showing acceleration.
Record custom-silicon bookings or design wins.
Electro-optics, optical DSP and CXL/PCIe switching turning into FY27/FY28 revenue.
Data-center growth and gross margin holding together.

Final Read

Marvell and Broadcom earnings are not a blanket buy signal for Korean semiconductors. They are a bottleneck map.

If Broadcom beats the AI semiconductor guide and talks up AI networking, Korea’s read-through expands beyond HBM into FC-BGA, high-speed PCBs, test sockets and silicon capacitors. If Marvell confirms custom silicon and optical interconnect acceleration, AI servers should be viewed less as GPU boxes and more as memory, connectivity, package and power-delivery systems.

The safest conclusion is still disciplined: SK hynix and Samsung Electronics remain the memory core; Samsung Electro-Mechanics is the clearest non-memory bottleneck; PCB and test-socket names need revenue and margin proof before they deserve full re-rating.

Fact / Inference / Blocked

[Fact]

Marvell’s Q1 FY2027 call is scheduled for May 27, 2026 at 1:45 p.m. PDT. (Marvell)
Marvell’s Q4 FY2026 revenue was $2.219B and Q1 FY2027 guidance is $2.4B ±5%. (Marvell)
Marvell’s Q4 FY2026 data-center revenue was $1.651B, or 74% of revenue. (Marvell)
Broadcom’s Q2 FY2026 earnings call is scheduled for June 3, 2026 at 5:00 p.m. EDT. (Broadcom Events)
Broadcom’s Q1 FY2026 AI revenue was $8.4B and Q2 AI semiconductor guidance is $10.7B. (Broadcom)
Samsung Electro-Mechanics announced an approximately KRW 1.5T silicon capacitor supply contract. (Samsung Electro-Mechanics)

[Inference]

Custom ASIC growth is more likely to broaden HBM demand than reduce it.
For Korea, the more actionable read-through is memory, package substrates, silicon capacitors and test sockets rather than generic optical themes.
Samsung Electro-Mechanics’ silicon capacitor contract is evidence that package-level power integrity is moving into commercial supply.

[Blocked]

Customer-level proof that specific Korean suppliers ship directly into Broadcom or Marvell ASIC programs.
The exact customer, margin and package placement for Samsung Electro-Mechanics’ silicon capacitor contract.
AI customer revenue split for Korean PCB and test-socket suppliers.

This article is for research and commentary only, not investment advice. Earnings dates and guidance are based on company IR materials as of May 23, 2026 KST. Korean supply-chain exposure includes inference because customer-level disclosures are limited.

After Google I/O 2026 — Search Isn't Dead. The Real Issue Is the AI Infrastructure Bottleneck Created by the Agentic OS

Thu, 21 May 2026 10:35:00 +0900

📚 Related Series Google I/O + NVIDIA Earnings Preview / Korea AI Infrastructure After NVIDIA Q1 FY27 / Samsung Electronics TSMC-Style Rerating / Samsung Electro-Mechanics Hybrid Challenger / May 17 Market Snapshot

Google I/O 2026 was not a simple model announcement. It was an event that bundled Gemini 3.5, Gemini Omni, Gemini Spark, Antigravity, AI Search, Universal Cart, and Android XR together to reengineer Google’s entire product portfolio into an agentic OS. Search moves from a window displaying links to an interface that executes tasks; shopping becomes a cart agent; developer tools become a runtime that runs multiple agents in parallel. From an investment perspective, the key is not Alphabet itself but the infrastructure bottlenecks this transformation creates: token throughput, inference traffic, data center CapEx, HBM, packaging, power integrity, high-layer PCBs, and power equipment. Post-Google I/O, Korean alpha comes not from models but from bottlenecks.

Key Takeaways

The Essence of Google I/O 2026: Google bundled Gemini 3.5 Flash, Gemini Omni Flash, Search agents, Gemini Spark, Antigravity 2.0, Managed Agents API, Universal Cart, and Android XR into a single agentic product layer.
The Search Collapse Thesis Is Not Yet Confirmed by the Numbers: Alphabet’s Q1 2026 Search & other revenue was $60.399 billion, up approximately 19.1% year-over-year from $50.702 billion in the prior year period. So far, this looks less like AI encroaching on Search and more like Google absorbing Search into an AI interface.
The Most Important Number Is Token Throughput: Google disclosed that monthly processed tokens grew from 9.7 trillion in May 2024, to approximately 480 trillion at the 2025 I/O, to over 32 quadrillion in May 2026. The agentic OS shifts the unit of usage from search queries to tokens, tasks, and agent execution counts.
Alphabet Shares Confirm Defensiveness; Korea Should Track the Bottleneck: Alphabet’s case as a full-stack AI company has strengthened, but its market cap in the $4 trillion range and an earnings multiple of around 30× already reflect significant expectations. Fresh alpha is more visible in HBM, packaging, power integrity, PCBs, and data center power.
Korean Priority: Samsung Electro-Mechanics > Samsung Electronics > SK Hynix > Test/Socket/HBM Equipment > ISU Petasys · Daeduck Electronics · TLB · Korea Circuit > Power Equipment.
Strongest Single Idea: Samsung Electro-Mechanics. The KRW 1.557 trillion silicon capacitor contract is evidence that power stabilization within AI packages has emerged as a distinct bottleneck.
Macro Conditions: If the seven macro gates from the May 17 market snapshot have not been cleared, high-P/E AI infrastructure names can be shaken even with a good thesis. Scaling in after confirmation beats chasing.

1. The Essence of Google I/O 2026 — All of Google Becomes an Agentic OS

Summarizing this I/O as “Gemini got better” understates what actually happened. What Google actually did was reengineer its entire product portfolio into an agent interface.

Google’s official I/O collection presented Gemini Omni and Gemini 3.5 as the core new models, and grouped Search agents, Gemini Spark, Universal Cart, intelligent eyewear, Workspace, and Antigravity all under agentic experiences. In other words, this was not an update to one or two features — it was a repackaging of Search, the Gemini app, YouTube, Gmail, Docs, Shopping, Android, and Cloud developer tools into a single agent layer.

Reclassifying the key announcements from an investment perspective:

Category	Announcement	Investment Significance
Models	Gemini 3.5 Flash, 3.5 Pro upcoming	AI competition axis shifts from one-shot responses to long-running task execution
Multimodal Generation	Gemini Omni Flash	Video generation/editing, YouTube Shorts, and Flow become creative infrastructure
Search	AI Mode as default model, intelligent Search box, Search agents	Search box transforms from a query box to a task interface
Personal Agent	Daily Brief, Gemini Spark	Personal OS tying Gmail, Calendar, Docs, and Gemini data together
Developer Tools	Antigravity 2.0, CLI, SDK, Managed Agents API	Beyond Cursor-style coding assistance into enterprise agent runtime
Commerce	Universal Cart	Reassembling the purchase funnel across Ads, Search, YouTube, and Gmail
XR	Android XR, smart glasses	Post-smartphone ambient computing form factor option

In this structure, Google’s true asset is not any single model. It is the distribution that encompasses Search, YouTube, Gmail, Android, Chrome, Workspace, Shopping, Maps, Cloud, and TPU. Even when OpenAI or Anthropic lead on model performance, Google can convert the surfaces where users already spend time into an agent layer. That is what “agentic OS operator” means, as opposed to merely a “Search defensive name.”

2. Search Isn’t Dead — But the Unit of Search Is Changing

The market narrative over the past two years has been simple.

AI answers replace search results
→ Users click fewer links
→ Google Search ad revenue erodes
→ Alphabet multiple contracts

This logic is intuitive, but current numbers tell the opposite story. In Alphabet’s Q1 2026 results, Google Search & other revenue was $60.399 billion, up approximately 19.1% from $50.702 billion in the prior year period.

Search & other growth rate
= 60.399 / 50.702 - 1
= approx. 19.1%

Google Cloud revenue also came in at $20.028 billion, up approximately 63.4% from $12.260 billion in the prior year period. Cloud operating income was $6.598 billion, with an OPM of 32.9%.

Cloud growth rate = 20.028 / 12.260 - 1 = approx. 63.4%
Cloud OPM = 6.598 / 20.028 = approx. 32.9%

The facts to date do not support “AI killed Search.” A more accurate framing is: “Google is defending usage by transforming Search into an AI interface.”

That said, there is no room for complacency. The economic unit of search is changing.

Legacy Search	AI Search / Agentic Search
Keyword input	Natural language, files, images, video, tab input
Link clicks	Agent collects information, summarizes, and suggests actions
CPC-centric	Intent/action auction potential
Search query count	Token throughput · agent execution count
SEO publisher traffic	Task completion within Google surfaces

Google stated that AI Mode surpassed 1 billion monthly users within a year of launch and that AI Mode queries have more than doubled quarter over quarter since launch. Google also presented a framework in which Search agents continuously monitor web, news, social, financial, shopping, and sports data in the background.

This is not the destruction of search — it is the redefinition of search. But redefinition comes with costs. Search where an agent reads the web, summarizes, reasons, and sends notifications consumes far more compute than search that displays ten links. That is the critical point for Korean semiconductors and AI infrastructure.

3. The Most Important Numbers — Token Throughput and CapEx

The most important numbers from this I/O are not model benchmarks — they are usage figures.

Google disclosed that monthly processed tokens have grown as follows.

Period	Monthly Processed Tokens	Significance
May 2024	9.7 trillion	Early-stage AI usage
2025 I/O	approx. 480 trillion	Gemini, Cloud, and developer usage expansion
May 2026	32 quadrillion+	Approx. 7× year-over-year increase

The Gemini app also grew from 400 million monthly users at the 2025 I/O to over 900 million in May 2026. Google stated that daily requests also increased more than 7× over the past year.

More important is CapEx. Sundar Pichai stated that while Google spent $31 billion in annual CapEx in 2022, this year it expects approximately $180–190 billion. At that scale, Google I/O is not a consumer product event — it is a global AI infrastructure demand event.

AI agentic OS
→ Search, Gemini, YouTube, Workspace, Android, Shopping usage increases
→ Token throughput increases
→ Inference cluster expansion
→ HBM / packaging / network / power / cooling bottlenecks widen
→ Korean AI infrastructure value chain benefits

The first-order interpretation that “Google using TPUs reduces NVIDIA’s benefit” is too shallow. Whether it’s a TPU, GPU, or ASIC, high-performance AI accelerators require HBM, advanced packaging, high-speed substrates, power stabilization components, and data center power. Korean investors should focus on shared bottlenecks rather than any specific US chip vendor.

4. Alphabet Stock — Defensiveness Confirmed, Fresh Alpha Limited

Alphabet reinforced three defensive arguments through this I/O.

First, it demonstrated that Search is not simply a platform being encroached upon by AI, but rather a platform that absorbs AI. The 19% growth in Q1 2026 Search & other revenue lends force to this argument.

Second, Google is a full-stack AI company with models, TPU, Cloud, Search, YouTube, Android, Workspace, and Ads all under one roof. Its monetization surface is broader than companies that have only a model API.

Third, Google Cloud demonstrated both high growth and high margins simultaneously. A Cloud revenue growth rate of 63% and Cloud OPM of 32.9% reduce concerns that “AI Cloud grows revenue but not profits.”

However, investment discipline demands a cold eye. Alphabet is already a company with a market cap in the $4 trillion range, and its input P/E is around 30×. Even on a simple annualized Q1 operating income basis, market cap to operating profit is approximately 30×. The Q1 2026 net income figure also includes significant other income effects such as unrealized gains on equity investments, so concluding it is cheap based on GAAP EPS alone would be a mistake.

The investment view on Alphabet is as follows.

Item	View
Business Quality	Very high
Search Defensiveness	Reinforced by I/O and Q1 2026 results
Cloud Growth	Strong
Valuation	Already largely priced in
New Entry	Watchlist / on correction
Better Alpha	AI infra second-order beneficiaries

Alphabet is a good company. But the better trade after Google I/O is not chasing Alphabet — it is finding the beneficiaries of the CapEx bottlenecks Google will create.

5. Korean Value Chain Mapping — From HBM to Power Integrity

The Korean investment implication of Google I/O is not simply “HBM demand is good.” It is that as agentic workloads increase, bottlenecks propagate in the following sequence.

AI Search / Omni / Antigravity / Spark / Universal Cart
→ Persistent inference traffic increases
→ TPU/GPU/ASIC cluster expansion
→ HBM · DRAM · eSSD demand increases
→ Packaging · substrate · power stabilization component demand increases
→ Optical/networking · data center power demand increases

Mapping the Korean value chain by node:

Node	Bottleneck	Korean Candidates	View
HBM / Memory	Bandwidth, capacity	SK Hynix, Samsung Electronics	Structural benefit continues; leaders are crowded
Advanced Packaging	HBM stack, bonding, inspection	Hanmi Semiconductor, Intest Plus, ISC, Leeno Industrial	Beneficiary thesis is valid but pricing and customers need confirmation
Power Integrity	Silicon capacitors, MLCC	Samsung Electro-Mechanics	Clearest new alpha
Substrate / PCB	FC-BGA, high-layer PCB, memory module PCB	Samsung Electro-Mechanics, ISU Petasys, Daeduck Electronics, Simmtech, TLB, Korea Circuit	Earnings confirmation needed over theme-chasing
Power Infrastructure	Transformers, distribution, power quality	HD Hyundai Electric, LS ELECTRIC, Hyosung Heavy Industries, Iljin Electric	Demand is clear; valuation discipline required
Optical / Network	Optical, SerDes, interconnect	OESOLUTION etc.	Potential beneficiary; direct revenue verification needed

The key here is not finding “direct Google suppliers.” Even if it cannot be confirmed whether Korean companies directly supply Google TPUs, the bottlenecks commonly consumed by Google, NVIDIA, AWS, Meta, and Microsoft’s AI infrastructure expansion exist. Alpha lies in identifying where those bottlenecks accrue.

6. Samsung Electro-Mechanics — The Clearest Korean AI Infrastructure Alpha After Google I/O

The most important Korean name after this I/O is Samsung Electro-Mechanics. The reason is simple: the fact that AI infrastructure bottlenecks are shifting from HBM to power stabilization within the package has been confirmed through a real order.

Samsung Electro-Mechanics disclosed on May 20, 2026, that it had entered into a silicon capacitor supply agreement worth KRW 1.5570 trillion with a major global company. The contract period runs from January 1, 2027 to December 31, 2028.

Annualized contract value = KRW 1.5570tn / 2 years = KRW 778.5bn
Total contract value as % of recent revenue per disclosure = 13.8%
Annualized revenue contribution = 13.8% / 2 = approx. 6.9%/year

More important than the revenue figure itself is the classification shift this contract represents. Samsung Electro-Mechanics was previously classified as an MLCC, camera module, and substrate company. However, silicon capacitors are components placed inside or close to AI semiconductor packages to reduce power noise and stabilize power delivery. In other words, Samsung Electro-Mechanics is entering the AI server package BOM.

Samsung Electro-Mechanics’ official technical documentation also explains that MLCC-centric power architectures are expanding due to growing power demands from AI servers and GPU modules. AI GPUs have large instantaneous power fluctuations, making low ESR/ESL, high-frequency ripple response, and space efficiency critical. This is not a simple MLCC cycle — it is a power integrity cycle.

Samsung Electro-Mechanics’ P×Q×C framework can be viewed as follows.

Axis	Variable	Significance
P	Silicon capacitor, AI MLCC, FC-BGA ASP	Higher value-add mix versus legacy IT components
Q	GPU/TPU/ASIC package count, HBM stack count, AI server board count	Directly linked to AI CapEx
C	Yield, material costs, expansion depreciation	Early ramp risk

Why this may not yet be fully priced in.

The market still strongly tends to view Samsung Electro-Mechanics as a smartphone MLCC and camera module company.
Silicon capacitors are a separate category, and it takes time for them to be incorporated into sell-side models.
Contract revenue recognition begins in 2027–2028, so it feeds into near-term EPS with a lag.
The customer is undisclosed, so it cannot be determined whether the order is from NVIDIA, Google, or a Broadcom/Marvell ASIC.

The conclusion is therefore not “chase the move” but rather “pullback buy.” Samsung Electro-Mechanics is the clearest reclassification candidate in Korean AI infrastructure following this I/O and NVIDIA’s results, but chasing a sharp spike is inefficient when the macro gates from the May 17 market snapshot have not yet opened.

7. Samsung Electronics and SK Hynix — The Other Implication of Google TPU Multipolarity

Google I/O might appear to be an event that weakens NVIDIA’s monopoly. Google has its own TPUs, and the ASIC ecosystem is broadening to include Broadcom, Marvell, and even MediaTek. But from a Korean memory perspective, the conclusion is different.

TPUs consume HBM. ASICs consume HBM. Even as inference TPUs multiply, memory bandwidth and data movement bottlenecks do not disappear.

SK Hynix

SK Hynix remains the highest-quality beta to HBM. Google I/O’s growing agentic workloads, NVIDIA Rubin ramp, and hyperscaler CapEx expansion are all positive for HBM demand.

However, the alpha has diminished. The market already views SK Hynix as the lead HBM name. Being a good company and offering a good entry price right now are different things.

View: Hold / Buy on pullback
Positives: HBM pure beta, customer trust, profitability
Watch: crowding, elevated expectations, intensifying HBM4 competition

Samsung Electronics

Samsung Electronics is more interesting. The market tends to view Samsung Electronics solely as a “late HBM follower to NVIDIA.” However, as Google/Broadcom/Marvell/ASIC multipolarity advances, the HBM customer structure also moves away from a single-NVIDIA frame. This dynamic is favorable for a variant perception on Samsung Electronics.

Samsung Electronics is an integrated IDM with HBM, DRAM, eSSD, base die, foundry, and advanced packaging all under one roof. The issue is execution. HBM4 customer qualification, base die, foundry profitability, and customer diversification must be confirmed in actual numbers before it can escape its approximately 5× P/E discount.

View: Watchlist / Conditional Buy
Positives: HBM4, base die, eSSD, foundry optionality
Watch: execution discount, foundry losses, HBM qualification delays

This is the same conclusion as the prior Samsung Electronics rerating piece. Samsung Electronics is not cheap because the market doesn’t understand it. The market is demanding proof of downcycle defensibility and customer neutrality. Google I/O showed one path to meeting that demand: not through the NVIDIA single-chain, but through TPU/ASIC multipolarity in which Samsung’s HBM and foundry optionality can come alive.

8. Broadcom vs. Marvell — Korean Investors Should Not Be Looking for a Winner

In the US, the relative performance of Broadcom and Marvell post-Google I/O was also read as significant. The key is not whether Google is abandoning Broadcom. It is a signal that Google intends to multi-source its AI ASIC supply chain.

Broadcom is Google’s core TPU partner, and there are reports of a long-term agreement with Google. Marvell drew attention from reports of discussions with Google on a new AI chip, memory processing unit, and inference TPU. This movement does not mean “Google TPU demand is weak” — it means “Google is expanding its supply chain leverage.”

What matters to Korean investors is not which US ASIC vendor wins. It is the bottlenecks that are needed in common regardless of whether Broadcom or Marvell prevails.

US ASIC Shift	Korean Common Bottleneck
TPU generation increases	HBM, base die, advanced packaging
Inference TPU expansion	HBM/LPDDR, eSSD, networking
Memory processing unit	Memory bandwidth, substrate, test socket
AI rack growth	High-layer PCB, optical, power equipment
Multi-sourcing	Value increase for companies with customer qualifications

The Korean priority is therefore not any specific US chip designer but vendor-agnostic bottlenecks. The order should be: Samsung Electro-Mechanics, Samsung Electronics, SK Hynix, test/socket, high-layer PCB, power infrastructure.

9. Practical Trading Strategy

Priority

Rank	Stock / Group	View	Rationale
1	Samsung Electro-Mechanics	Buy on pullback	Silicon capacitor contract confirms AI package power integrity bottleneck via real order
2	Samsung Electronics	Watchlist / Conditional Buy	TPU/ASIC multipolarity can narrow the late-HBM discount
3	SK Hynix	Hold / Buy on pullback	HBM direction is right but consensus is crowded
4	Hanmi Semiconductor · ISC · Leeno Industrial · Intest Plus	Watchlist	Benefit from growing HBM/ASIC SKUs and test/packaging complexity
5	ISU Petasys · Daeduck Electronics · Simmtech · TLB · Korea Circuit	Watchlist	AI server PCB beneficiary potential, but customer/product/margin confirmation needed
6	HD Hyundai Electric · LS ELECTRIC · Hyosung Heavy Industries · Iljin Electric	Wait	AI data center power demand is clear, but valuation may be stretched

Entry Conditions

Samsung Electro-Mechanics: Confirmation of follow-on silicon capacitor customers, expansion plans, and potential disaggregation of AI MLCC/FC-BGA revenue.
Samsung Electronics: HBM4 customer qualification, HBM4 base die, volume toward Google/Broadcom/ASIC customers, or signs of foundry recovery.
SK Hynix: HBM4 pricing holds, NVIDIA Rubin ramp, 10–15% correction, or easing of supply-demand overheating.
PCB / Substrates: When AI-oriented revenue mix, ASP premium, and OPM maintenance are confirmed in quarterly results.
Test / Socket: Confirmation of TPU/ASIC/HBM customer references and new socket revenue.
Power Equipment: When North American data center order backlog growth and OPM maintenance are confirmed simultaneously.

Catalyst

Gemini 3.5 Pro launch and AI Mode/Search agent adoption.
Google Cloud next-quarter revenue, backlog, and CapEx update.
Broadcom and Marvell earnings commentary on Google/ASIC pipeline.
NVIDIA Rubin / Vera Rubin NVL72 ramp.
Samsung Electro-Mechanics follow-on silicon capacitor disclosure.
Samsung Electronics HBM4 qualification, base die, and foundry customer commentary.
Korean PCB/substrate companies disaggregating AI server-oriented revenue in Q2–Q3.

Invalidation

Google Search & other growth rate decelerates sharply after AI Mode expansion.
Google Cloud growth remains high but OPM falls below the low-20% range.
AI CapEx increases translate into depreciation, power costs, and networking costs that erode Alphabet margins.
HBM ASP declines or HBM4 oversupply signals emerge.
Samsung Electro-Mechanics’ silicon capacitor contract is confirmed to be low-margin or a single-customer event.
PCB/substrate names where AI revenue increases but OPM declines.
May 17 macro gates remain uncrossed, compressing high-P/E AI infrastructure multiples.

10. Frequently Asked Questions

“Is Google I/O Negative for NVIDIA?”

In the near term, some negative interpretation is possible. If Google expands its TPU footprint and the Broadcom/Marvell ASIC ecosystem grows, NVIDIA’s monopoly premium could compress somewhat. But from a Korean supply chain perspective, multipolarity is actually positive. Whether it is a GPU, TPU, or ASIC, HBM, packaging, PCBs, and power infrastructure are all required.

“Wouldn’t It Be Better to Just Buy Alphabet?”

Alphabet is a good company. Search defensiveness, Cloud growth, and its full-stack AI structure are all strong. However, the current market cap and earnings multiple already reflect much of the winner premium. The conclusion of this piece is not that Alphabet is a bad investment, but that immediately after Google I/O, fresh alpha is larger in Korean AI infrastructure bottlenecks than in Alphabet itself.

“Why Is Samsung Electro-Mechanics the Top Pick?”

HBM is already well known to the market. By contrast, the power integrity thesis — covering power stabilization within AI packages, silicon capacitors, AI MLCC, and FC-BGA — is still not finely classified by the market. To that, a real order of KRW 1.557 trillion is attached. This is a rare moment when both the story and the numbers are present simultaneously.

“Can I Buy All the PCB/Substrate Names?”

No. This is the segment with the greatest risk of theme excess. The “AI server-oriented” label is easy to apply, but without confirmed customers, products, ASP premium, and OPM, it can end as a short-term trading theme. From here, only companies that simultaneously show “AI-oriented revenue growth + margin maintenance” should remain in the portfolio.

11. The Final Word

Google I/O 2026 was not an event that declared search dead. On the contrary, it was an event in which Google declared its intention to convert search, mail, documents, shopping, developer tools, video, and even eyewear into agent interfaces. The search box moves from a query box to a task box; developer tools move from an IDE to an agent runtime; and shopping moves from a keyword funnel to Universal Cart.

Alphabet’s Q1 2026 numbers show that this transition has not yet damaged Search. Search & other revenue grew approximately 19% year-over-year; Cloud revenue grew approximately 63%; and Cloud OPM rose to 32.9%. So far, “Google absorbs AI into its existing distribution” is a more accurate frame than “AI kills Google.”

But the real investment point is the AI infrastructure bottleneck, not Alphabet itself. Google disclosed that monthly processed tokens grew from 9.7 trillion in May 2024 to over 32 quadrillion in May 2026. It also indicated that CapEx could expand from $31 billion in 2022 to approximately $180–190 billion this year. This shift is a demand function for HBM, packaging, power integrity, PCBs, and power infrastructure.

Korean priorities are clear. Samsung Electro-Mechanics has confirmed the AI package power integrity bottleneck through a real large-scale silicon capacitor contract. Samsung Electronics is an option where TPU/ASIC multipolarity can narrow the late-HBM discount. SK Hynix remains the strongest HBM name but is already a crowded trade. PCB/substrates, test, and power equipment have significant beneficiary potential but require earnings confirmation.

The conclusion is simple. Post-Google I/O alpha comes not from “which model is smarter” but from “where does growing agent usage force spending on bottlenecks.” The answer is not Alphabet alone. It is HBM, packaging, silicon capacitors, high-layer PCBs, and data center power. Of these, the freshest Korean alpha right now is Samsung Electro-Mechanics. But until the macro gates open, scaling in after confirmation beats chasing.

This article is for research and commentary purposes only and does not constitute investment advice. Google I/O 2026 announcements (Gemini Omni, Gemini 3.5, Search agents, Gemini Spark, Universal Cart, Antigravity, Android XR) are based on the official Google I/O collection and Google’s official blog. Alphabet Q1 2026 results (revenue $109.896 billion, operating income $39.696 billion, Search & other $60.399 billion, Google Cloud $20.028 billion, Cloud operating income $6.598 billion) are based on Alphabet’s Q1 2026 SEC filings. Monthly processed tokens of 32 quadrillion+, Gemini app monthly users of 900 million+, and CapEx guidance of $180–190 billion are based on Sundar Pichai’s statements at Google I/O 2026. Samsung Electro-Mechanics’ KRW 1.5570 trillion silicon capacitor supply agreement (term: 2027/01/01–2028/12/31; 13.8% of recent revenue) is based on the company’s May 20, 2026 regulatory disclosure and major domestic news coverage. Statements related to Samsung Electro-Mechanics’ AI server power architecture are based on Samsung Electro-Mechanics’ official technical documentation. Information on Broadcom, Marvell, and Google ASIC is based on reporting by Reuters and other foreign media; confirmed contracts and negotiation reports are mixed and should be distinguished accordingly. Stock priorities, value chain mapping, entry conditions, and invalidation conditions represent the analyst’s current framework and may differ from actual outcomes. Global macro conditions (interest rates, oil prices, exchange rates, VIX, foreign investor flows) may have additional impact on stock prices. This analysis may be wrong. Data as of May 21, 2026 KST.

Disclaimer: For research and information purposes only. Not investment advice. Names cited are for analytical illustration; readers should perform their own due diligence and consult licensed advisors before any investment decision.

Korea Alpha in the CloudMatrix Expansion Scenario — Find Where Inefficiency Accrues, Not the Direct Huawei Suppliers

Sun, 17 May 2026 23:58:00 +0900

📚 US-China Decoupling Follow-Up Series US-China Summit Decoupling Analysis / NVIDIA Korea 2nd–3rd Tier Value Chain / AI Back-End 11-Stock Data Comparison / Semiconductor Value Chain Hub

Huawei’s CloudMatrix 384 is not a simple story of China catching up to NVIDIA. It is a system that lashes together 384 Ascend chips, draws 559kW of power, and routes thousands of optical transceivers to match the GB200 NVL72 in aggregate throughput — not in efficiency. China lost on efficiency but countered with system scale. The real alpha that this inefficiency points to is not in direct Huawei suppliers. Because China cannot reach NVIDIA, it must scale up its own fab, equipment, and consumables investment — and that process generates demand for inspection, metrology, and high-difficulty consumables. Treating 티씨케이, 넥스틴, and 코미코 as the same China theme is a trap. The three companies differ in investment thesis, confirmed results, and buy timing.

Key Takeaways

CloudMatrix 384’s essence is aggregate bypass, not an efficiency victory. 384 Ascend chips, roughly 559kW of power, and large-scale optical connectivity allow it to match the GB200 NVL72 on select metrics.
The real alpha is not in direct Huawei suppliers. HBM export controls and China’s domestic substitution policy limit direct supply routes for Korean companies. What matters instead is the indirect accrual structure: as China invests in its own fabs, demand for inspection, metrology, and consumables grows.
티씨케이 is a confirmed recovery play. 1Q26 revenue of ₩95.4bn, operating income of ₩28.6bn, and an operating margin of roughly 30% reflect China customer normalization and high-layer NAND conversion benefits already in the numbers.
넥스틴 is an inflection-point bet. The China inspection equipment PO resumption and Wuxi subsidiary ramp are the key catalysts — but 1Q26 was an operating loss, and the May 15 tape showed simultaneous foreign and institutional selling. This is still a pre-confirmation zone.
코미코 is global foundry quality with a China option, not a China play. The core thesis is TSMC and global foundry cleaning/coating demand plus ESC product mix improvement.
The May 15 tape differentiated the three names. 티씨케이 saw foreign buying near the 20-day moving average despite the selloff. 넥스틴 broke below all major moving averages with retail absorbing the supply.
Current priority order is 티씨케이 > 코미코 > 넥스틴. 넥스틴 has the highest thematic purity, but purity and buy timing are different things.

1. What CloudMatrix Is Actually Pointing To

Viewing CloudMatrix 384 purely as a “Chinese GB200” is half right and half wrong. Based on reporting from Tom’s Hardware and SemiAnalysis, CloudMatrix 384 links 384 Ascend chips to deliver roughly 300 PFLOPs of BF16 performance — higher than the GB200 NVL72’s roughly 180 PFLOPs. On some metrics, its memory capacity and bandwidth are also larger.

But the cost is steep. System power for CloudMatrix 384 is reported at approximately 559kW, compared to roughly 145kW for the GB200 NVL72 — a ratio of 3.85x. China did not beat NVIDIA on efficiency; it matched aggregate throughput by using more chips, more power, and more interconnects.

Metric	Huawei CloudMatrix 384	NVIDIA GB200 NVL72	Interpretation
Accelerators	384 Ascend	72 GPU + 36 CPU	Huawei uses far more chips
BF16 Performance	~300 PFLOPs	~180 PFLOPs	Huawei leads in raw aggregate
System Power	~559kW	~145kW	NVIDIA leads in efficiency
Optical Links	6,912 × 800G-class LPO reported	Relatively fewer	China’s aggregate bypass drives sharply higher optical dependency

What investors should extract from this table is not “Huawei won.” It is where the inefficiency accrues. Weaker chips require more chips. More chips require more power, more cooling, more substrates, more optical connectivity, more inspection, and more process stability. The more inefficient China’s AI infrastructure, the higher the total input volume of components and equipment.

This is where Korean alpha emerges — not because Korean companies can freely supply HBM or key equipment directly to Huawei. That route is constrained by sanctions and domestic substitution policy. The opposite is true: precisely because direct routes are blocked, China must scale investment in SMIC, CXMT, YMTC, and the broader Huawei ecosystem through its own fabs, equipment, and consumables. The resulting demand for yield-improving inspection and metrology equipment and high-difficulty process consumables flows outward from there.

The inference chain should be reframed as follows.

Flawed Chain	Better Chain
CloudMatrix success → Huawei revenue growth → find Huawei direct suppliers	H200 withheld, sanctions sustained → China scales own fab investment → yield requirements rise, process difficulty increases → inspection, metrology, consumables demand grows

This article focuses on the second chain.

2. The Three Names Are Not the Same China Theme

Markets tend to group 티씨케이, 넥스틴, and 코미코 together as “China semiconductor self-sufficiency beneficiaries.” But the investment thesis for each is entirely different.

Item	티씨케이	넥스틴	코미코
Investment Character	Confirmed consumables recovery	2H26 China PO inflection bet	Global foundry quality + China option
CloudMatrix Directness	Low	Medium–High	Low–Medium
China AI Self-Sufficiency Purity	Medium	Highest	Medium
1Q26 Confirmed	Yes	No	Partial
Core Product	SiC Ring and other high-difficulty consumables	Wafer defect inspection equipment	Cleaning/coating, ESC
Key Catalyst	China customer normalization, high-layer NAND	China PO resumption, Wuxi subsidiary ramp	TSMC spillover effect, ESC volume ramp
Current View	Conditional Buy	Watch	Watchlist

티씨케이 is not a “direct CloudMatrix component supplier.” More precisely, it benefits from China fab utilization normalization and NAND high-layering through higher consumables usage. Inventory destocking by Chinese customers ended, urgent orders picked up, and 1Q26 results confirmed the inflection. Revenue of ₩95.4bn, operating income of ₩28.6bn, and an operating margin of roughly 30% are numbers already in the record, not mere expectations.

넥스틴 is the highest-purity bet. China building its own AI chips requires improving yield, and improving yield requires more wafer defect inspection. That is where the 넥스틴 China subsidiary and PO resumption narrative comes from. But the numbers have not been confirmed yet. 1Q26 was an operating loss, and the 2H26 China PO resumption must actually materialize.

코미코 loses its logic if it is purchased purely as a China theme. The core is global foundry cleaning and coating quality. The key drivers are TSMC and global foundry utilization, operations across US, Taiwan, and Korean facilities, and ESC product mix improvement at Mico Ceramics. The China ESC option exists, but it is difficult to reduce the entire company to a single China narrative.

This distinction matters. 티씨케이 is a current-results confirmation play, 넥스틴 is a future inflection play, and 코미코 is a global quality play. Trading all three under the same theme can cause mistimed entries and exits across all three names.

3. What the May 15 Tape Said

In the sharp selloff on May 15, the signals from the three names diverged — particularly the gap between 티씨케이 and 넥스틴.

Item	티씨케이	넥스틴
May 15 Close	₩305,000	₩64,100
May 15 Change	-12.5%	-13.4%
20-Day Return	+12.3%	-15.2%
60-Day Return	+38.0%	-21.1%
120-Day Return	+87.9%	-0.2%
vs. 52-Week High	-12.5%	-36.2%
Gap to 20-Day MA	+0.1%	-12.6%
Gap to 50-Day MA	+15.7%	-13.8%
RSI 14	53.9	36.7

Both names fell roughly 13%, but the meaning is different. For 티씨케이, it was the first support test — a pullback to near the 20-day moving average within a strong 120-day uptrend, with RSI cooling to a neutral 53.9. For 넥스틴, it is trading below the 20-day, 50-day, and 120-day moving averages simultaneously. An RSI of 36.7 may look like oversold territory, but a low RSI within a downtrend can signal trend deterioration rather than a buy opportunity.

The supply/demand picture also differed.

Item	티씨케이	넥스틴
May 15 Foreign	+₩4.91bn	-₩2.31bn
May 15 Institutional	-₩3.39bn	-₩1.95bn
May 15 Retail	-₩1.51bn	+₩4.26bn
May 15 Program	-₩7.65bn	-₩3.45bn
Short-Sale Ratio	15.16%	14.03%
Stock Lending Balance	₩212.4bn	₩56.1bn

티씨케이 attracted foreign buying even on the selloff day. Institutional and program selling was present, but the cumulative foreign accumulation trend remained intact. 넥스틴, by contrast, saw foreigners and institutions selling together while retail absorbed the supply. The tape pattern to avoid most in an event-driven setup is exactly “foreign + institutional dual selling absorbed by retail.”

Looking at 5-day and 20-day cumulative flow, the divergence is even more pronounced.

Period	티씨케이 Foreign	티씨케이 Institutional	넥스틴 Foreign	넥스틴 Institutional
5D	+₩14.06bn	+₩12.97bn	+₩5.02bn	-₩3.53bn
20D	+₩74.33bn	-₩15.68bn	+₩11.92bn	-₩15.76bn

티씨케이 pulled back with ₩74.3bn of 20-day foreign accumulation underneath it. 넥스틴 had just ₩11.9bn of 20-day foreign buying, of which ₩2.3bn left on May 15 alone. The base is entirely different. 티씨케이’s move can be read as a healthy consolidation; 넥스틴 warrants waiting for trend recovery confirmation before taking a position.

4. Four Common Misreads

First misread: connecting CloudMatrix to a search for direct Huawei suppliers. The pathway for Korean core components and equipment to reach Huawei directly is constrained by sanctions and China’s domestic substitution policy. Alpha comes from indirect accrual, not direct supply. The more inefficient China’s AI systems, the greater the demand for domestic fab investment and process stabilization — which flows into inspection, metrology, and consumables.

Second misread: treating all inspection and metrology names as one block. 넥스틴 focuses on front-end pattern inspection, Park Systems on AFM metrology, OROS on overlay metrology, and Intekplus on FC-BGA and packaging inspection. They all carry the “inspection/metrology” label, but their catalysts, customer sets, and revenue recognition timing are different.

Third misread: assuming a low RSI is always a buy signal. In an uptrend, RSI at 30–40 can mark a pullback opportunity. In a downtrend, RSI at 30–40 can signal continued weakness. 넥스틴 is trading below its 20-day, 50-day, and 120-day moving averages simultaneously. In that context, RSI 36.7 is a wait signal, not a buy signal.

Fourth misread: treating thematic purity as the primary buy criterion. 넥스틴 has the highest purity for the China AI self-sufficiency theme. But purity alone is insufficient. Buy attractiveness requires assessing purity, price positioning, supply/demand flow, and fundamental confirmation together. On current price and tape, 티씨케이 has the advantage. 넥스틴 is high-purity but unconfirmed, with weak flow.

5. Name-by-Name Entry Guide

티씨케이 can be considered a pullback buy candidate on confirmation of ₩300,000 support. A first check is support at ₩300,000–305,000 with sustained foreign net buying. Secondary targets are recovery to ₩320,000, then ₩335,000. Conversely, a break below ₩295,000 accompanied by continued foreign selling warrants trimming. If 2Q26 operating margin falls below 25%, or if signals emerge that China normalization was a one-off event, the investment thesis weakens.

넥스틴 is still better observed than bought. A recovery to ₩69,500, a resumption of foreign buying, news on China PO or Wuxi subsidiary ramp, quarterly revenue recovery, and a visible path to operating profit breakeven need to appear together. Buying near ₩64,000 on RSI alone risks catching a falling knife. If China PO is not confirmed by 3Q26 or the Wuxi ramp is delayed, the thesis itself needs to be downgraded.

코미코 is a quality company but not a standalone CloudMatrix buy. To initiate a position, confirm the ESC volume production transition, utilization rate improvement across US, Taiwan, and Anseong facilities, TSMC revenue share trends, and the degree to which Mico Ceramics profits accrue to parent shareholders. The China option is a secondary catalyst.

6. Portfolio Construction

For a balanced portfolio, 40% 티씨케이 / 30% 코미코 / 20% 넥스틴 / 10% hedge is the most straightforward allocation. 티씨케이 is the core with confirmed results and foreign accumulation. 코미코 provides global quality and ESC optionality. 넥스틴 is a satellite position for convexity when 2H26 China PO confirmation arrives.

An aggressive portfolio can raise 넥스틴’s weight — but conditions apply. Recovery to ₩69,500, foreign re-entry, and China PO news must appear simultaneously. Without those conditions, a 넥스틴-heavy aggressive allocation carries unnecessary risk.

A conservative portfolio fits closer to 60% 티씨케이 / 30% 코미코 / 10% 넥스틴. What is confirmed today is 티씨케이; 코미코 is a quality holding; 넥스틴 remains an option.

Type	티씨케이	코미코	넥스틴	Hedge
Balanced	40%	30%	20%	10%
Aggressive	35%	20%	45%	0%
Conservative	60%	30%	10%	0–10%

7. Scenario-Based Action

The highest-probability scenario is sustained US sanctions combined with accelerating Chinese self-sufficiency. If H200 shipments remain blocked and China continues building its own systems, fab and process stabilization investment across the SMIC, CXMT, YMTC, and Huawei ecosystems is likely to persist. In this scenario, 티씨케이 responds fastest. 넥스틴 can strengthen later when China PO actually resumes, and 코미코 follows more slowly through ESC and global foundry revenue.

If US sanctions ease and China returns to NVIDIA, 넥스틴’s high-purity thesis takes the hardest hit. 티씨케이 has NAND and consumables drivers that provide some cushion. 코미코’s global business is the primary driver, so the relative impact is smaller.

A scenario of further US sanction tightening and increased ally-pressure pressure is also worth considering. In this case, restrictions on Korean equipment and component exports to China could intensify. 넥스틴 faces direct exposure. 티씨케이 is relatively better positioned given the consumables nature of its products, but is not fully insulated. 코미코’s global revenue share provides defensive depth.

The possibility of a Xi Jinping visit to Seoul in September and the October truce expiration are also catalysts. A truce extension would bring near-term relief; re-escalation would create near-term shock; a major deal would generate near-term euphoria. But in any of these cases, the underlying force driving China to build its own AI systems does not disappear. The framework of inefficiency accrual created by CloudMatrix therefore remains valid across scenarios.

8. Investment Conclusion

CloudMatrix 384 is not evidence that China beat NVIDIA on efficiency. It is a system that bypasses GB200 NVL72 in aggregate by using more chips, more power, and more interconnects. That inefficiency itself is the investment point. Inefficiency creates cost, and cost accrues into equipment and consumables demand.

What Korean investors should seek is not direct Huawei suppliers. That path is narrow given HBM export controls and China’s domestic substitution policy. The focus should be on the inspection, metrology, and high-difficulty consumables needed as China scales its own fab investment and pushes yields higher.

At current prices and tape, 티씨케이 leads. 1Q26 results are confirmed, and foreigners were buying on the May 15 selloff. 코미코 is global foundry quality with a China option attached. 넥스틴 has the highest thematic purity but remains unconfirmed. A simultaneous appearance of 2H26 China PO resumption and flow recovery is needed before its buy priority rises.

The real alpha is not in the word “CloudMatrix” itself. It lies in matching where inefficiency accrues and whether the market tape is confirming that accrual structure. Right now, the cleanest combination of those two conditions sits in 티씨케이. The largest option value is in 넥스틴. The longest-duration quality is in 코미코.

Sources and Evidence Classification

Fact

CloudMatrix 384 has been reported as a structure using 384 Ascend chips, approximately 559kW of power, and large-scale optical transceivers.
티씨케이 disclosed 1Q26 revenue of approximately ₩95.39bn and operating income of approximately ₩28.6bn, implying an operating margin of roughly 30%.
May 15 supply/demand and price tape figures are based on backfill data provided by the user.

Inference

CloudMatrix’s aggregate bypass structure was judged to require expanded Chinese investment in domestic fab, equipment, and consumables.
티씨케이 was classified as a confirmed consumables recovery play, 넥스틴 as a China PO inflection bet, and 코미코 as global quality with a China option.
At current price and May 15 tape, 티씨케이 was judged to have greater buy attractiveness than 넥스틴.

Speculation

넥스틴’s 2H26 China PO resumption, Wuxi subsidiary ramp, and 코미코’s ESC revenue expansion are scenarios based on sell-side estimates and industry trend analysis.
The scenario-by-scenario impact of a September Xi visit and October truce expiration are analyst estimates.

Primary sources referenced: Tom’s Hardware CloudMatrix 384 report, SemiAnalysis CloudMatrix 384 analysis, 티씨케이 1Q26 results disclosure coverage, 티씨케이 Shinhan Investment report summary PDF, 코미코 SK Securities estimate summary.

This article is intended for research and commentary purposes only and does not constitute investment advice. CloudMatrix performance and power figures are based on published reports and analyst sources and may differ under actual operating conditions. 티씨케이, 넥스틴, and 코미코 results and outlook references draw on company disclosures, press reports, sell-side report summaries, and backfill data provided by the user. China PO resumption, ESC volume ramp, and September summit diplomacy scenarios are all subject to significant uncertainty. Name-level priority rankings and portfolio weightings reflect the analyst’s subjective judgment and carry no guarantee. This analysis may be wrong. Data as of May 17, 2026 KST.

KOSPI May 12 Selloff and Rebound: AI Citizen Dividend, Tax Windfall, and the Samsung-SK Hynix Profit Debate

Tue, 12 May 2026 13:30:00 +0900

🔗 Related reading: Why Korea Part 3 — Samsung Electronics and SK Hynix are upgrading Korea’s economic weight · Samsung Electronics Citi KRW 460,000 target analysis · Korea Semiconductor, HBM and KOSPI Hub

The May 12 intraday selloff and rebound in KOSPI should not be dismissed as ordinary semiconductor volatility. The surface explanation was familiar: a pullback in global AI beta, oil and geopolitical noise, and profit-taking after a very sharp Korea semiconductor rally. But the speed of the move and the way the market rebounded point to another trigger: the policy question of how Korea should distribute the gains from the AI semiconductor supercycle.

Kim Yong-beom, head of the Presidential Policy Office, did not appear to be launching a direct attack on Samsung Electronics or SK Hynix. His comments are better read as an attempt to frame the AI semiconductor boom as a fiscal, distributional and national-strategy event. But markets do not parse political philosophy slowly. When “excess profits,” “citizen dividend,” “structural return” and Norway’s oil fund appear in the same policy context, foreign investors hear a different phrase: Korea AI semiconductor windfall-tax risk.

That is why the May 12 move matters. This was not only a question of whether Samsung and SK Hynix had risen too fast. It was the first market test of a deeper issue: will the excess margins created by Korea’s AI infrastructure bottleneck accrue to shareholders, or will policymakers increasingly view those gains as a social and fiscal resource?

TL;DR

Kim’s comments look closer to an excess-tax-revenue argument than a direct corporate-profit levy. His recent sequence is consistent: KOSPI re-rating, semiconductor tax windfall, AI-era distribution design, and a broader state-strategy frame.
The communication problem was real. By mixing “excess tax revenue” with “excess profits,” the message was easy to read as a potential AI or semiconductor windfall tax on Samsung Electronics and SK Hynix.
The morning selloff and rebound looks like a policy-headline risk unwind. The first read was “AI profits shared with citizens”; the later clarification was closer to “excess tax revenue, not corporate profit-sharing.”
The core Samsung-SK Hynix earnings thesis is not broken. There is no confirmed special levy, windfall-tax bill or forced profit-sharing mechanism. But the policy-risk premium has not disappeared.
The debate is unavoidable. If Samsung Electronics and SK Hynix can generate more than KRW 300 trillion of annual profit in an AI cycle, the question of who receives the gains — shareholders, workers, the state, regions and non-beneficiaries — becomes a national issue.

1. What the Market Priced First

The market did not price a tax bill on the morning of May 12. No semiconductor special levy had been announced. No corporate profit-sharing mechanism had been legislated. What the market priced was policy language.

The sensitive sentence was the idea that the fruits of the AI infrastructure era should not remain only with specific companies and asset owners, but should be structurally shared with citizens. Seoul Economic Daily’s English edition reported that Kim argued Korea may need a new social contract as AI chips, power grids and data centers reshape the structure of the Korean economy. Asia Economy carried the Korean-language report that Kim had raised the principle of a national dividend in the AI era. (Seoul Economic Daily, Asia Economy)

As policy philosophy, that is a legitimate topic. If Korea is gaining a structurally stronger position in AI memory, power, data centers and advanced manufacturing, it is reasonable for policymakers to ask how the gains should spread across society.

But the equity market hears accounting first.

AI demand expands
 -> HBM, DRAM, power and substrate bottlenecks
 -> Samsung and SK Hynix margins rise
 -> The state may view part of that excess return as a social resource
 -> EPS or multiple discount

The key issue is not the current tax code. The issue is uncertainty over who captures incremental margin. The Korea semiconductor rally is not just a revenue-growth story. It is a scarcity-margin story. If investors begin to question whether that incremental margin fully belongs to shareholders, they sell first and ask questions later.

That is the best way to understand the morning move. It was not a judgment that AI memory demand had deteriorated. It was a repricing of the perceived ownership of AI semiconductor excess returns.

2. Kim Yong-beom’s Frame: Market-Friendly, but Fiscal

It would be inaccurate to describe Kim as a politician who does not understand markets. He is a career economic policymaker who served across finance and fiscal agencies, including senior roles at the Financial Services Commission and the Ministry of Economy and Finance. He was involved in crisis-response work during the Covid period. (Yonhap profile)

His recent comments show a consistent structure.

2.1 January: Korea Discount, Capital Markets and Tax Capacity

In a January interview with SisaIN, Kim spoke positively about a Korean equity-market re-rating. His logic was that a more advanced stock market can support corporate financing, investment, dividends, corporate taxes, securities-transaction taxes and income taxes. He also emphasized the importance of tax capacity for policy programs. (SisaIN)

This is not anti-market thinking. It is fiscal thinking.

Capital-market re-rating
 -> Higher corporate value
 -> More investment, dividends and transactions
 -> More tax revenue
 -> More policy capacity

He is not rejecting rising equity markets. He is trying to connect rising markets to fiscal capacity.

2.2 April: AI and Basic-Income Logic

The AI distribution frame had already entered the policy conversation in late April. In President Lee Jae-myung’s meeting with Google DeepMind CEO Demis Hassabis, the Korean side raised the question of basic income in the AI era. The reported answer emphasized the need to think about basic services and new economic models. (Daum/MBC report)

So the May 12 citizen-dividend language did not come out of nowhere. It was part of a broader government conversation about what happens when AI changes labor, productivity and wealth distribution.

2.3 May 8: Semiconductor Boom and Excess Tax Revenue

On May 8, Kim argued that Korea’s KOSPI 7,500 and 10,000 discussion should not be viewed only through a conventional cyclical lens. He suggested that if the semiconductor boom continues through 2027, Korea could see historically large tax revenues in 2026 and 2027. Chosunbiz’s English edition and Yonhap Infomax both reported the “historic tax windfall” framing. (Chosunbiz English, Yonhap Infomax)

That is the direct bridge to May 12.

Semiconductor supercycle
 -> Samsung and SK Hynix corporate taxes rise
 -> High-income semiconductor workers pay more income tax
 -> Trade surplus and asset-market effects improve
 -> Excess tax revenue
 -> More fiscal flexibility

2.4 May 12: From Tax Windfall to Citizen Dividend

The May 12 comment takes that logic one step further. If AI infrastructure creates structural excess returns and if that produces excess tax revenue, what principle should guide the social use of those gains?

That looks much closer to an excess-tax-revenue framework than a direct corporate-profit seizure. The problem is that the language was sloppy for markets. “Excess tax revenue” and “excess profits” are very different concepts, but in the trading window they were heard as one.

3. Why Foreign Investors Heard Windfall Tax

For foreign investors, the Korea AI semiconductor story has three major attractions.

Pillar	Investment Logic
Price	Samsung and SK Hynix trade at lower multiples than global AI infrastructure bottlenecks
Earnings	HBM and DRAM pricing plus supply constraints drive operating leverage
Index	Their KOSPI weights attract both active and passive foreign capital

The most important pillar is earnings. More precisely, it is incremental margin. The bull case is not merely that Korea sells more memory. It is that shortage conditions allow Korea to sell critical AI memory at much better margins.

When a senior policy official uses the language of AI excess profits, citizen dividends and structural redistribution, foreign investors translate it like this:

Korean Policy Language	Market Translation
Excess profits	Excess profit
Citizen dividend	Fiscal transfer from AI gains
Structural return to citizens	Institutionalized redistribution
Norway’s oil fund	Resource-rent capture
Not solely the result of specific companies	Shareholder claim may be politically limited

This combination naturally raises the specter of a windfall tax. Bloomberg’s original headline framed the issue around using AI profits for a citizen dividend, which was enough for global investors and algorithms to react. Subsequent clarification that the idea was based on excess tax revenue rather than corporate profit-sharing helped reverse part of the move. (Bloomberg)

That makes the intraday rebound understandable: windfall-tax risk priced in, excess-tax-revenue clarification priced back out.

4. Why This Debate Is Unavoidable

The debate is uncomfortable, but it is not avoidable. The reason is the scale of the numbers.

As argued in Why Korea Part 3, when Samsung Electronics and SK Hynix are capable of generating more than KRW 300 trillion in annual profits, the macro impact goes far beyond equity investors. Tax receipts, bonuses, pension-fund gains, regional investment, local taxes and supplier revenue all move together.

At that scale, the gains become a social question.

Channel	Social Question
Corporate tax	Is this a temporary windfall or a durable fiscal-capacity shift?
Bonuses	How wide can the gap become between semiconductor workers and the rest of the labor market?
Stock gains	How should society think about the gap between shareholders and households without equity exposure?
Regional investment	How do Pyeongtaek, Yongin, Icheon and Cheongju gains spread beyond semiconductor clusters?
AI automation	If Korea supplies AI infrastructure, how does it support workers displaced by AI adoption?

Kim’s “citizen dividend” language attempts to compress all of these questions into one phrase. The phrase is market-sensitive, perhaps too market-sensitive. But the underlying question will not go away.

For Korea, AI is not only a software story. It is a physical infrastructure story: memory, power equipment, substrates, batteries, displays, shipbuilding, robotics and precision manufacturing. That means the AI boom shows up as manufacturing profits, tax revenue, regional capex and high-income engineering compensation.

The social debate is therefore inevitable. The market communication has to be much more precise.

5. Excess Tax Revenue Is Not a Windfall Tax

This is the distinction the market had to process in real time.

Concept	Meaning	EPS Impact	Market Reaction
Use of excess tax revenue	Government spends tax revenue collected under the existing tax code	Limited	Fiscal-policy issue
Corporate tax-rate hike	The tax rate on semiconductor profits rises	Directly negative	PER discount
Special levy	A sector-specific levy on excess profits	Directly negative	Foreign selling pressure
Profit-sharing scheme	Part of corporate profit is transferred to a social fund	Directly negative	Windfall-tax risk
Public growth-fund participation	Citizens gain exposure through investment vehicles	Limited	Capital-market expansion

If Kim’s proposal remains an excess-tax-revenue discussion, Samsung and SK Hynix EPS are not directly impaired. The companies pay taxes under existing rules, and the government debates how to use unexpectedly large revenue.

If the proposal evolves into a special levy, windfall tax or forced profit-sharing structure, the story changes. Then semiconductor margins and shareholder returns are directly affected, and the upper bound of the Korea semiconductor re-rating falls.

As of now, the confirmed facts point more toward the first interpretation. But the market’s initial reaction shows how quickly investors will price the second.

6. Investment Read-Through

The Samsung-SK Hynix thesis is not broken by this event. HBM, DRAM pricing, AI inference infrastructure, sovereign AI, physical AI, robotics and data-center power remain the same structural drivers. Kim’s own comments assume Korea is becoming more central to AI infrastructure, not less.

But the multiple discussion now has an extra variable.

The old question was: how large will Samsung and SK Hynix profits be in 2026 and 2027?

The new question is:

How much of the fiscal and social gain from those profits will the Korean state try to formalize as public policy?

That is policy-risk premium. If it never becomes law, the impact can fade. If “AI excess-profit social return” becomes recurring language, foreign investors may assign Korea semiconductors a slightly lower multiple.

6.1 Samsung Electronics and SK Hynix

The core thesis remains intact. There is no confirmed direct corporate-profit levy. The intraday rebound after the clarification suggests the market also moved back toward an excess-tax-revenue interpretation.

But chasing after the rebound requires a higher burden of proof: cleaner policy language, foreign futures and cash-market flow, and closing-price resilience.

6.2 Samsung Electro-Mechanics, AI Substrates and Optical Names

Second-derivative AI infrastructure names are more sensitive to policy headlines and global AI-beta sentiment. They depend on the capex and infrastructure expansion that follows the Samsung-SK Hynix memory supercycle.

The citizen-dividend debate is not a direct EPS hit to these names. The more important variables remain US AI valuation pressure, order confirmation and the timing of actual revenue recognition.

6.3 Banks, Platforms and Telecom

This is not only a semiconductor issue. Once the government begins to speak in terms of social returns from industry-level excess profits, the frame can later apply to banks, platforms and telecom companies — sectors that already carry public-interest narratives.

The broader signal is that Korean sectors with unusually large excess profits may face more policy language around social return.

7. What to Watch Next

Checkpoint	Interpretation
Presidential Office or MOEF clarification	If they emphasize tax revenue, risk falls
Annual tax-code proposal	No semiconductor special levy means limited EPS risk
Budget proposal	Excess-tax-revenue spending is neutral; sector levies are negative
Ruling-party bills	Social funds or profit-sharing schemes would be multiple-negative
Foreign broker notes	Watch for repeated “Korea AI windfall tax risk” language
Foreign futures and cash flow	Distinguish headline short-covering from structural derisking
Samsung and SK Hynix close and next-day gap	Market verdict on whether the risk was one-day noise

Final Note

The May 12 KOSPI selloff and rebound showed the next variable in the Korea AI semiconductor rally. When profits become very large, markets stop asking only how large those profits can be. They start asking who gets to keep them.

Kim Yong-beom’s AI citizen-dividend language currently looks closer to a debate over how to use excess tax revenue than a plan to directly seize Samsung or SK Hynix profits. That is why the direct EPS risk is low, and why the intraday rebound makes sense.

But the communication was poor. If “excess tax revenue,” “excess profits,” “citizen dividend” and “structural return” appear together, foreign investors will read windfall tax. The Korea semiconductor re-rating remains valid, but after May 12 it carries a small additional policy-risk premium.

The debate itself is unavoidable. Once Samsung Electronics and SK Hynix profits become large enough to upgrade the national economy, the question of how those gains are divided among shareholders, workers, the state, regions and non-beneficiaries will grow with them. May 12 was the first day that question entered the market price.