Test Socket on Korea Invest Insights

AI Back-End — 11 Stocks Side by Side. Who's Actually Cheap and Who's Expensive? Real Value on 2027E PER × Operating-Profit Growth

Sat, 16 May 2026 00:00:00 +0000

📚 AI back-end series Previous: Substrates are a “volume beta,” test sockets are a “consumables beta” — same AI tailwind, completely different structures

Yesterday’s post compared substrates and test sockets structurally. Today we go one step further and compare them numerically. Eleven AI back-end names lined up on the same yardstick — 2026 YTD return, operating-profit growth, 2026E and 2027E PER, 2026E operating margin. Even within “AI winners,” how much each has already moved and how much room is left differ wildly. One table tells you most of the story.

Key takeaways

On pure “cheapness”: SK hynix (2027E PER 5.2x), Samsung Electronics (5.6x), Haesung DS (15.4x). But the memory megas are dominated by “cycle-peak” risk.
On “growth-adjusted multiple”: Daeduck Electronics (2027E PER 27.1x vs OP +34%) and Simmtech (20.0x vs OP +59%) are the sharpest setups.
“Already done the work”: Samsung Electro-Mechanics (2026 YTD +296%, 2026E PER 61.9x), Isu Petasys (+346% in 2025 but only +7% YTD 2026), Simmtech (+352% in 2025).
“Quality, but the multiple is rich”: LEENO Industrial (OPM 48.6%, 2026E PER 41.2x), ISC (2026E PER 58.5x).
“Growth at a fair price”: TSE (2026E OP +82%, PER 29.4x).
Bottom line: within AI back-end, attractiveness ranks substrates > test sockets > memory megas. Within substrates, Daeduck and Simmtech are the cleanest.

1. How the comparison is built

1.1 The data set

Field	Source	Meaning
2025 stock return	PyKRX	How much it ran last year
2026 YTD return	PyKRX	How much it has run this year
2025 OP YoY	Company disclosures	Last year’s earnings growth
2026E OP YoY	FnGuide consensus	Market’s view on this year
2027E OP YoY	FnGuide consensus	Market’s view on next year
2026E PER	FnGuide consensus	Multiple on this year’s earnings
2027E PER	FnGuide consensus	Multiple on next year’s earnings
2026E OPM	FnGuide consensus	Business-model margin level

Reference date: May 15, 2026 close.

1.2 Why 2027E PER matters more than 2026E PER

2026E PER:
→ Multiple on this year's expected earnings
→ Already absorbing 1Q26 reality
→ Affected by short-term quarterly noise

2027E PER:
→ Multiple on next year's expected earnings
→ Reflects "will the cycle persist?"
→ Where the market's true view shows up

Examples:
Company A: 2026E PER 30x, 2027E PER 28x
→ Market expects earnings roughly flat
→ "Stable growth" archetype

Company B: 2026E PER 30x, 2027E PER 15x
→ Market expects earnings to nearly double
→ Or the market hasn't bought in yet → undervalued setup

Company C: 2026E PER 7x, 2027E PER 5x
→ Market doubts earnings durability
→ "Cycle-peak fear" — typical of memory megas.

1.3 Why OP growth has to be read with PER

PER alone is misleading.

Same PER 30x:
A: OP YoY +20% → 30 / 20 = 1.5
B: OP YoY +60% → 30 / 60 = 0.5
B is far more attractive.

This is the PEG ratio (PER / OP growth).
≤1.0 attractive, 1.0–2.0 fair, >2.0 expensive.

This post visualizes a simplified version —
a 2027E PER × 2027E OP YoY matrix.

2. Eleven stocks, one table

2.1 The basic comparison

Bucket	Name	2025 return	2026 YTD	2025 OP YoY	2026E OP YoY	2027E OP YoY	2026E PER	2027E PER	2026E OPM
Memory	Samsung Electronics	+124.5%	+125.6%	+33.2%	+688.6%	+26.7%	7.0x	5.6x	51.1%
Memory	SK hynix	+280.3%	+179.4%	+101.2%	+433.1%	+36.0%	6.8x	5.2x	75.9%
Memory	Jeju Semiconductor	+186.4%	+212.3%	+274.0%	n/a	n/a	n/a	n/a	n/a
Test socket	LEENO Industrial	+52.1%	+67.5%	+42.5%	+23.2%	+24.1%	41.2x	33.4x	48.6%
Test socket	ISC	+58.7%	+102.1%	+34.2%	+65.9%	+41.4%	58.5x	43.2x	31.8%
Test socket	TSE	+39.5%	+215.1%	+23.3%	+81.9%	+21.5%	29.4x	24.7x	17.0%
Substrate	Samsung Electro-Mechanics	+108.3%	+296.1%	+24.3%	+72.0%	+50.0%	61.9x	41.9x	11.8%
Substrate	Daeduck Electronics	+206.0%	+188.3%	+334.5%	+375.4%	+34.4%	36.1x	27.1x	15.5%
Substrate	Simmtech	+351.9%	+103.2%	loss→profit	+1,234.5%	+58.7%	31.0x	20.0x	8.6%
Substrate	Haesung DS	+140.6%	+57.3%	-18.3%	+99.6%	+36.6%	20.3x	15.4x	11.4%
Substrate	Isu Petasys	+346.5%	+7.4%	+100.9%	+57.1%	+33.0%	35.6x	26.7x	21.1%

2.2 How to read it

Down the columns:
- High 2025 return = already run a lot
- High 2026 YTD = still strong this year
- High 2026E OP YoY = earnings explosion expected
- Low 2027E PER = market doubts earnings persistence

Across the rows:
- Was the name strong last year too (+200%+)?
- Still strong this year (+100%+)?
- Yet not punitive on the multiple (2027E PER ≤ 30)?
- Decent operating margin (2026E OPM ≥ 20%)?

Names that hit all four = most attractive.

3. Reading each bucket

3.1 Memory megas — “cheap, but cycle-peak fight”

The numbers look stunning:
SK hynix: 2027E PER 5.2x, OPM 75.9%
Samsung Electronics: 2027E PER 5.6x, OPM 51.1%
→ In normal industries, PER 5x screams "bankruptcy risk"
→ Memory is different. When the cycle turns, earnings fall by half+.

Why the market still applies 5x:
"We'll grant you 2026 earnings.
 We are not convinced 2027–28 earnings will stay here."

This is the heart of memory investing:
- The stock looks cheapest right before the cycle peaks
- The moment the cycle turns, PER spikes (because EPS collapses)
- Buying "trough PER" can mean selling at the actual top.

Read:
SK hynix — best case if HBM supercycle runs into 2027.
 Large drawdown if the cycle turns.
Samsung — strike risk + HBM share doubts = discount vs. SK hynix.
Jeju Semi — no consensus available; excluded from direct compare.

3.2 Test sockets — “quality intact, multiple full”

LEENO Industrial:
- OPM 48.6% (top tier of Korean manufacturing)
- 2027E OP YoY +24.1% (steady growth)
- 2027E PER 33.4x
→ Read: highest quality, but multiple is full.

ISC:
- OPM 31.8% (very good)
- 2027E OP YoY +41.4% (strong growth)
- 2027E PER 43.2x
→ Read: cleanest AI data-center beta, but tightest multiple.

TSE:
- OPM 17.0% (lowest among the three)
- 2027E OP YoY +21.5% (decent)
- 2027E PER 24.7x (cheapest in test sockets)
- 2026 YTD +215% (already running hard)
→ Read: best price-for-growth balance, but chase risk after the YTD move.

Read:
Don't treat LEENO and ISC as the same "test-socket stock."
- LEENO = quality compounder (hold-and-own)
- ISC = direct AI data-center beta (momentum)
- TSE = the balanced "value-for-growth" option.

If you must pick one new entry:
- 1-year hold: LEENO (stability)
- 6-month momentum: TSE (value-for-growth)
- Direct AI exposure: ISC (speed).

3.3 Substrates — the most interesting bucket

Samsung Electro-Mechanics:
- 2026 YTD +296% (among the top movers)
- 2026E PER 61.9x → 2027E PER 41.9x
- OPM 11.8% (blended across MLCC + FC-BGA)
→ Great company, but chasing is inefficient.

Daeduck Electronics:
- 2025 +206%, 2026 YTD +188% (persistent strength)
- 2026E PER 36.1x → 2027E PER 27.1x
- OPM 15.5% (top of substrate group)
- 2026E OP YoY +375%, 2027E +34%
→ The numbers justify the leadership premium.

Simmtech:
- 2025 +352% (biggest mover)
- 2026 YTD +103% (still strong)
- 2026E PER 31.0x → 2027E PER 20.0x (lowest)
- 2026E OP YoY +1,234% (loss → profit)
- 2027E OP YoY +58.7%
- OPM 8.6% (margin needs to expand further)
→ Turnaround + multiple appeal — but margin durability needs proof.

Haesung DS:
- 2026 YTD +57% (lagger of the group)
- 2026E PER 20.3x → 2027E PER 15.4x (cheapest of the group)
- 2026E OP YoY +99.6% (very strong)
- OPM 11.4%
→ The cheapest candidate. Weaker leadership profile,
 but the price-for-growth is good.

Isu Petasys:
- 2025 +346% (a parabolic year)
- 2026 YTD +7% (resting this year)
- 2026E PER 35.6x → 2027E PER 26.7x
- OPM 21.1% (highest in substrates)
→ All the move came pre-emptively last year; resting.
 Wait for relative-strength recovery.

4. The growth-vs-multiple matrix — the most intuitive view

4.1 2027E PER × 2027E OP YoY

X-axis: 2027E operating-profit growth
Y-axis: 2027E PER

low ← PER → high
+5x --10x---20x---30x---40x---50x PER

+60% │ Simmtech
 │
+50% │ Samsung Electro-Mechanics
 │
+40% │ ISC
 │
+35% │ SK hynix
 │
+34% │ Daeduck Electronics
 │
+36% │ Haesung DS
 │
+33% │ Isu Petasys
 │
+27% │ Samsung Electronics
 │
+24% │ LEENO Industrial
 │
+22% │ TSE
 │
 └──────────────────────────────

How to read:
- Upper right (high PER + low growth): expensive
- Lower left (low PER + high growth): best value
- Upper left (low PER + low growth): trap
- Lower right (high PER + high growth): momentum

Most attractive zone:
Left (low PER) + top (high growth) = Simmtech, Daeduck

Most punitive zone:
Right (high PER) + mid growth = LEENO

Special zone:
Far left (very low PER) = SK hynix, Samsung Electronics
→ Looks cheap, but cycle-peak risk lives here.

4.2 Reading “real value” from the matrix

Simplified PEG (PER / growth):

Simmtech: 20.0 / 58.7 = 0.34 ★★★★★
SK hynix: 5.2 / 36.0 = 0.14 (memory exception)
Samsung Elec: 5.6 / 26.7 = 0.21 (memory exception)
Haesung DS: 15.4 / 36.6 = 0.42 ★★★★
Daeduck: 27.1 / 34.4 = 0.79 ★★★★
Isu Petasys: 26.7 / 33.0 = 0.81 ★★★
Samsung E-M: 41.9 / 50.0 = 0.84 ★★★
TSE: 24.7 / 21.5 = 1.15 ★★★
ISC: 43.2 / 41.4 = 1.04 ★★★
LEENO: 33.4 / 24.1 = 1.39 ★★

Bucketing:
- ≤ 0.5: very attractive
- 0.5–1.0: attractive
- 1.0–1.5: fair
- ≥ 1.5: expensive

Memory megas are a special class — cycle peak.
On normal terms, Simmtech > Haesung DS > Daeduck.

LEENO looks the worst on PEG, but its
earnings stability and predictability are unmatched.
It deserves a "premium multiple" — PEG alone underrates it.

5. Cross-bucket — “where should fresh capital go?”

5.1 Substrates vs test sockets vs memory

Scenario 1: "AI back-end shortage cycle continues"
→ Substrates most attractive
 shortage → ASP hike → margin expansion
 Daeduck Electronics, Simmtech are top picks

Scenario 2: "AI chip diversification has further to run"
→ Test sockets most attractive
 new chip = new socket
 LEENO Industrial, ISC are top picks

Scenario 3: "HBM supercycle extends into 2028"
→ Memory megas most attractive
 SK hynix is top (OPM 75.9%)
 But you have to accept cycle-peak risk

Safest framing:
Diversify across the three buckets — 4:3:3 or 5:3:2.

5.2 Illustrative allocation across the 11

Just an example. Not real money advice.

Aggressive (short-term momentum tilt):
Daeduck 25% / Simmtech 20% / Samsung E-M 15%
ISC 15% / TSE 10% / SK hynix 15%

Balanced (growth + stability):
Daeduck 20% / LEENO 20% / SK hynix 15%
Samsung E-M 15% / Haesung DS 10% / Simmtech 10%
TSE 10%

Defensive (quality tilt):
LEENO 30% / Samsung E-M 20% / SK hynix 20%
Daeduck 15% / Haesung DS 15%

Allocation rules:
1. Higher OPM = larger weight
2. Heavier multiple burden = smaller weight
3. Don't concentrate in a single bucket
4. Cap cycle-risk names.

6. One-line read by name

6.1 Author’s attractiveness ranking

Rank	Name	One-liner
1	Daeduck Electronics	Best balance of growth, leadership, and multiple. Direct AI MLB / FC-BGA beta.
2	Simmtech	Best growth-vs-PER setup on 2027E. Needs further 2026 margin expansion to confirm.
3	Haesung DS	Cheapest substrate candidate. Less leadership profile, but strong value-for-growth.
4	SK hynix	The memory leader. Cheap on screens; cycle-peak risk is the swing variable.
5	TSE	Best value-for-growth within the test-interface group. But 2026 YTD +215% — already in motion.
6	LEENO Industrial	Top quality, full multiple. Best as a core 1–2 year holding.
7	Samsung Electro-Mechanics	A great company, but +296% YTD and 2026E PER 61.9x make chasing inefficient.
8	ISC	Real growth, but tightest multiple. Cleanest AI data-center beta.
9	Isu Petasys	Strong long-term structure, but +7% YTD — wait for relative-strength recovery.
10	Samsung Electronics	Cheap on numbers, but discount vs. SK hynix on the HBM narrative is real.
11	Jeju Semiconductor	Strong stock and earnings, but no 2026–27 consensus → excluded from direct compare.

6.2 Pre-entry checklist

Before buying any of these:

1. Macro gate (see the prior post)
 - US 10-year below 4.45%
 - Brent below 105 USD
 - USD/KRW below 1,480
 - VIX below 18
 → 2–3 of 4 stable before new buying.

2. Stock-level catalysts
 - Quarterly print dates
 - New customer LTA announcements
 - New-fab ramp schedule (LEENO)
 - 2Q guidance

3. Scaling rules
 - No one-shot buys
 - Adjust size as macro gates flip
 - Prefer pullback entries over chasing.

4. Bucket diversification
 - Avoid single-bucket concentration
 - Sensible balance across memory / substrate / socket.

7. Frequently asked

7.1 “Why not put memory megas at #1 if PER is single-digit?”

PER 5x looks like "trough PER,"
but a multiple built on "trough EPS" is dangerous.

The memory-cycle trap:
Cycle peak: earnings spike → PER 5x
Cycle decline: earnings cut in half → same price → PER 10x
Cycle trough: earnings cut to 1/3 → same price → PER 15x+

So buying at PER 5x means:
- More cycle ahead → big gains
- Cycle turns → suddenly PER 15x at the same price → big losses

That's why memory investing is not
"how cheap does it look?"
but
"when does the cycle turn?"

To put memory at #1 here, you'd need separate proof on
HBM4 / HBM4E demand visibility, China demand recovery,
and the durability of AI capex.

This post does a clean side-by-side only —
so memory megas are placed mid-pack to reflect cycle-peak risk.

7.2 “Is Simmtech’s 20x PER real?”

Simmtech's 2027E PER 20.0x is consensus-based.

Conditional signals:
1. 2026 OP turns positive (from loss),
 producing a +1,234% optical YoY (it is base-effect, not the run rate)

2. 2027 OPM must expand beyond 8.6% for the 20x PER to be "real"

3. Need a higher share of AI memory module / SSD-controller
 substrate revenue → confirm margin mix improvement.

Risks:
- Margin recovery may lag consensus
- 8.6% OPM is among the lowest in substrates (vs Daeduck 15.5%)
- 1Q26 results could trigger consensus revisions.

Read:
20x PER is "potential" value, not yet realized value.
Margin recovery is the real test.
Wait for 1Q26 / 2Q26 confirmation before adding.

7.3 “Samsung E-M is a great company. Why is it #7?”

Samsung Electro-Mechanics fundamentals are among the best:
- Combined AI FC-BGA + MLCC + camera exposure
- 1Q26 OP +40% YoY
- Core node in the global AI parts supply chain.

But on price:
- 2026 YTD +296% (one of the biggest movers of the 11)
- 2026E PER 61.9x (heavy vs. bucket average)
- OPM 11.8% (FC-BGA alone is higher; cameras drag the blend).

The ranking is "attractiveness at current price,"
not "best long-term holding."

For a long-term core position, Samsung E-M ranks high.
For best value-for-growth at today's price, Daeduck and Simmtech lead.

8. How this links to other posts

Prior post (structure compare):
→ Substrate = "volume beta," test socket = "consumables beta"
→ Structural differences between the buckets.

This post (data compare):
→ Who is most attractive WITHIN each bucket
→ Not raw PER — "growth-adjusted multiple."

Macro-cycle synthesis post:
→ "Cycle before the stock"
→ Scale into these names after the macro gate clears.

KOSPI crash + macro-gate post:
→ After the May 15 -6.12% drop
→ Watch relative strength AND the macro gate
→ The framework used to select names here.

9. The one-line bottom line

Even among “AI back-end winners,” how much price has already done the work differs wildly. Low PER alone ≠ attractive. Low PER can mean “the market fears the cycle is topping” (memory) or “the market hasn’t bought in yet” (turnaround names).

Lined up in one table, the picture is clear. On a PER × OP-growth matrix, the sharpest setups are Simmtech, Daeduck Electronics, and Haesung DS. All three are substrate names — that is not an accident. Substrates are in the middle of a “shortage → ASP hike → margin expansion” cycle.

Within test sockets, LEENO Industrial is the quality leader but the multiple is full; ISC is the cleanest AI data-center beta but carries the tightest multiple. For best price-for-growth on a new entry, TSE is the value option.

Memory megas show stunning numbers but a cycle-peak debate at their core. SK hynix’s 2027E PER 5.2x would imply bankruptcy risk in a normal industry; in memory it means “the market doubts earnings durability.” If the cycle persists, the upside is huge; the moment it turns, the multiple expands automatically.

At today’s prices, the rational order for new capital is substrates > test sockets > memory megas. Within substrates, Daeduck Electronics (leadership premium) and Simmtech (turnaround value) are the cleanest. Whatever you choose, scale in only after the macro gate clears. Buying a great company at the right price is the actual job.

This article is research and commentary only and is not investment advice. Stock returns are per PyKRX, with 2025 returns from 2024 year-end close to 2025 year-end close, and 2026 YTD from the 2025 year-end close to the May 15 close. Operating-profit growth, PER, and operating margin are per the FnGuide CompanyGuide consensus (queried on May 15, 2026) and may differ from actual reported figures. The consensus reflects the market average and can differ from individual broker estimates, and may be revised. The PEG metric (PER ÷ growth) is a simplified valuation aid and is not sufficient by itself. The cycle-peak concern for memory megas is the author’s view; actual cycles can be longer or shorter than expected. The attractiveness ranking reflects “best at current price” and may differ from a “best long-term holding” ranking. Jeju Semiconductor is excluded from the direct comparison due to a lack of 2026–2027 consensus and warrants its own analysis. Global macro variables (US rates, oil, FX, VIX) can independently move the stocks. The analysis may be wrong. Data cut-off: May 15, 2026 close, 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.

Two Betas of AI Back-End — Substrates Are a 'Volume Beta,' Test Sockets Are a 'Consumables Beta.' Same AI Tailwind, Completely Different Structures

Fri, 15 May 2026 00:00:00 +0000

📚 AI back-end series Previously: Samsung Electro-Mechanics MLCC and FC-BGA deep dive, Jeju Semiconductor legacy-memory thesis Read next: AI back-end — 11 stocks side by side on 2027E PER × OP growth

Who actually makes money when AI silicon sells? Nvidia (GPU) and SK hynix (HBM) are the names that come to mind first. But the “back end” of the supply chain has major winners too. We call it AI back-end. Two regions matter most: substrates and test sockets. Both are parts that an AI chip passes through before completion — but the investment structures are completely different. Substrates are a bet on “AI server units.” Test sockets are a bet on “AI chip complexity.” Margins are roughly 3x higher in test sockets, and momentum is faster in substrates. Which is the better investment depends on how long you plan to hold.

Key takeaways

Two regions in AI back-end: substrates and test sockets.
Substrate essence: a “direct CAPEX beta” on AI-server build-out. More units → more revenue; larger packages → higher ASP.
Test-socket essence: a “high-margin consumables beta” on chip complexity. More complex chips → harder testing → more / harder sockets.
Profitability gap: on 1Q26, Daeduck Electronics 14.8% vs ISC 35% vs LEENO Industrial 47.4%. Test sockets dominate.
Momentum gap: substrate earnings revisions are faster — ASP hikes, new big-tech LTAs, capacity tightness are all near-term catalysts.
Cycle risk: substrates are a CAPEX cycle (shortage → CAPEX → glut). Test sockets, being consumables, run with lower cyclicality.
Bottom line: short-term momentum trades → substrates (Daeduck Electronics, Samsung Electro-Mechanics). 1–2 year compounding → test sockets (LEENO Industrial, ISC). These are not the same “AI theme.”

1. Setup — what substrates and test sockets actually are

1.1 How AI silicon gets made

Production flow of AI silicon:

1. Design (Nvidia, AMD, Google, Meta, etc.)
2. Wafer fab (TSMC, Samsung Foundry)
3. Dicing
4. Packaging (HBM, interposer, package substrate join)
 ← "substrates" used here
5. Test (performance / reliability / burn-in)
 ← "test sockets" used here
6. Shipment

Substrate = "the pedestal the chip sits on"
Test socket = "the slot a chip plugs into briefly for testing"

Both are parts AI silicon passes through before completion,
but their roles and usage patterns are completely different.

1.2 Substrate — what is it

A substrate connects the microscopic circuitry inside a chip
(nanometer scale) to the outside world (millimeter scale).

What it does:
1. Electrical connection (thousands of pins → board)
2. Mechanical support (stable mounting of large packages)
3. Thermal management (heat-flow path away from the chip)
4. Signal integrity (high-speed signals without loss)

Why AI matters:
- AI chips are much larger than typical silicon
 (Nvidia H100 = 814 mm², 2–3x a typical CPU)
- Thousands to tens of thousands of pins
- High-speed data movement
- Heavy heat budget (700W+)

→ Specialized substrates (FC-BGA) are required
→ Harder to make and more expensive than ordinary substrates
→ Korean players: Samsung Electro-Mechanics, Daeduck Electronics,
 Isu Petasys.

1.3 Test socket — what is it

A test socket is a slot used briefly to test a finished chip.

What it does:
1. Electrically connects every chip pin to the tester
2. Verifies the chip works under signal / power / temperature stress
3. Filters out failing units
4. Reliability validation (long-duration testing)

Analogy:
Chip = car
Test socket = the diagnostic port at the inspection bay
→ Connected briefly during inspection, then disconnected
→ The same port is used for many different cars
→ Ports wear out over time and have to be replaced

Why AI matters:
- AI chips are expensive (\~USD 30k for an H100)
- A single bad unit costs a lot → test intensity rises
- High-current / high-speed / high-temperature environments
 → tougher socket spec
- Each chip generation needs new sockets → recurring revenue

Korean players: LEENO Industrial, ISC, TSE.

1.4 Why they are not the same “AI theme”

Substrate:
- Made once, shipped with the chip (capital-good character)
- Roughly one substrate per AI server
- Revenue = volume × ASP
- Adding capacity = CAPEX burden

Test socket:
- Reused many times, then replaced (consumable character)
- Roughly 50–200 sockets to test 10,000 chips
- Revenue = chip output × test intensity × socket replacement
- Capacity additions are smaller

Margin structure:
Substrate: OPM \~10–15% (CAPEX, depreciation drag)
Test socket: OPM \~30–50% (custom design, consumable economics)

→ Both benefit from AI,
→ but revenue mix, margin profile, and cycle sensitivity differ.

2. The substrate side — “AI server volume beta”

2.1 What 1Q26 showed

Samsung Electro-Mechanics (009150) — Package Solution segment
(FC-BGA-led):
1Q26 revenue: KRW 725B
YoY: +45%
QoQ: +12%

Daeduck Electronics (353200):
1Q26 revenue: KRW 346.3B (YoY +61%)
1Q26 OP: KRW 51.3B (turn-around to black)
OPM: 14.8%

Product mix:
- FCCSP (mobile-grade package substrate): 39%
- FCBGA (PC / server-grade): 23%
- CSP (memory): 22%
- MLB (multilayer board, AI server): 16%

YoY growth:
- Package substrates: +65%
- MLB: +43%

→ AI-server package substrates + network boards growing together
→ A direct beta on AI infrastructure.

2.2 Why substrates are short

Characteristics of AI-server FC-BGA:
- Surface area 2–3x a PC FC-BGA
- Multilayer (20–30 layers)
- High-speed signaling (PCIe 5.0 / 6.0)
- Strict thermal management

Companies that can make these:
- Korea: Samsung Electro-Mechanics, Daeduck, LG Innotek, Isu Petasys
- Taiwan: Unimicron, Nan Ya PCB
- Japan: Ibiden, Shinko Denki

The problem: capacity (CAPA) can't keep up with demand
- Big-tech AI-server orders have exploded
- Greenfield substrate fabs take 2–3 years to build
- Short supply → ASP hikes

Press cited:
"Samsung Electro-Mechanics FC-BGA demand exceeds capacity by \~50%;
 price-hike negotiations under way."

→ This is why substrate stocks have been strong.

2.3 Substrate cycle risk

Substrate is a classic CAPEX-cycle industry:

Up-cycle (today):
shortage → ASP hike → margin expansion → stock strength
→ companies announce sizable CAPEX

Build phase (2026–2027):
fabs under construction → revenue rising, CAPEX burden rising
→ stocks still trending strong

Build complete (2027–2028):
new fabs come online → supply surges
→ if demand doesn't keep up, prices fall
→ margins squeeze → stocks weaken

Historical patterns:
2017–2018 memory cycle: same shape
2021–2022 MLCC cycle: same shape

→ For substrate investing, "where in the CAPEX cycle" is the key
→ Today: early- to mid-stage of an up-cycle
→ Until build-out completes, upside is intact.

3. The test-socket side — “chip-complexity beta”

3.1 What 1Q26 showed

ISC (095340):
1Q26 revenue: KRW 68.3B
1Q26 OP: KRW 23.6B
OPM: 35%
Math: 23.6 / 68.3 = 34.55%

Revenue mix:
- AI revenue: KRW 55.3B (81% of total)
- Data-center revenue: KRW 54.2B (79% of total)
→ Already an "AI data-center company."

LEENO Industrial (058470):
1Q26 revenue: KRW 99.77B
1Q26 OP: KRW 47.30B
OPM: 47.4%
Math: 47.30 / 99.77 = 47.41%

Revenue mix:
- IC test sockets: 64.10%
- LEENO pins (pogo): 24.65%
- Medical components: 10.46%

→ Custom-designed sockets + vertically integrated pin manufacturing
→ 47% OPM is at the top of Korean manufacturing.

3.2 Why margins are this high

Structural features of the test-socket industry:

1. Customer-specific designs
 → pin layout, size, signal conditions vary by chip
 → hard to standardize → pricing power

2. High qualification cost
 → each new chip requires socket design / test / qualification
 → once qualified, locked in until end-of-life of that chip
 → reliability competition, not price competition

3. Consumable nature
 → sockets wear out / degrade and need periodic replacement
 → recurring revenue

4. Refresh with every chip generation
 → new chip = new socket
 → faster chip cadence = faster revenue cadence

5. Small TAM (a few tens of USD billion)
 → too small for mega-cap entrants
 → oligopoly of specialist players

All five conditions together enable 30–50% OPM.

3.3 Why AI made test sockets more important

AI chip characteristics:

1. Expensive (USD 10k–30k per unit)
 → cost of a single bad unit is high
 → test intensity ↑

2. High pin count (thousands–tens of thousands)
 → more complex sockets
 → ASP ↑

3. High current / temperature / speed
 → tougher socket durability requirements
 → shorter replacement cycle

4. SLT (System-Level Test) share rising
 → from simple functional test → full system test
 → longer test times
 → more socket usage

5. New modules: HBM, SOCAMM2, etc.
 → new socket categories
 → fresh revenue lines

→ In the AI era, test-socket demand grows
 faster than chip output.

3.4 ISC vs LEENO Industrial

Both are "test-socket companies," but their structures differ.

ISC (rubber-socket strength):
- Core tech: silicone rubber sockets
- Strength: AI data-center mass-production testing, SLT
- Customers: global GPU / ASIC majors
- Exposure: AI data center 81%
- Profile: direct AI beta (volatility higher)
- 1Q26 OPM: 35%

LEENO Industrial (pogo-pin strength):
- Core tech: pogo pins, in-house pin manufacturing
- Strength: R&D, mobile AP, RF, ASIC development
- Customers: diversified (hundreds of accounts)
- Exposure: many chip categories (lower AI share than ISC)
- Profile: quality compounder (more stable)
- 1Q26 OPM: 47.4%

→ Don't lump them as "the same test-socket stock."
→ ISC = AI data-center beta
→ LEENO = diversified high-margin platform

They are complements, not substitutes.

4. Head-to-head

4.1 1Q26 operating margins

Same "AI back-end beneficiaries," different margin profiles:

Daeduck Electronics (substrate): 14.8% ████
Samsung E-M Package Solutions: \~12% ███
ISC (test socket): 35.0% █████████
LEENO Industrial (socket): 47.4% ████████████

KRW 100 of revenue → OP:
Daeduck: KRW 14.8
ISC: KRW 35.0
LEENO: KRW 47.4

→ Roughly 3x spread on the same revenue
→ The result of structural industry differences.

4.2 Growth pace vs margin stability

Metric	Substrate (Daeduck)	Test socket (ISC)	Test socket (LEENO)
1Q26 revenue YoY	+61%	Strong YoY	+18%
1Q26 OP YoY	turnaround	Strong YoY	+35%
OPM	14.8%	35.0%	47.4%
AI exposure	Mid–high	Very high (81%)	Mid
Volatility	High	Mid	Low
Cycle sensitivity	High	Mid	Low

Summary:
- Revenue growth: substrate (Daeduck) > test socket
- Margin level: test socket > substrate
- Margin stability: LEENO > ISC > substrate
- Direct AI beta: ISC > Daeduck > LEENO.

4.3 Cycle risk

Substrate cycle risk (high):
- Shortage → CAPEX → glut, in repeat
- 2–3 year build lead time
- When the cycle rolls, utilization and ASP fall together
- Depreciation pressures margin

Test-socket cycle risk (low):
- Consumable economics dampens revenue volatility
- Customer-specific design keeps pricing stable
- New chips keep being released
- Quarterly volatility still exists

Long-hold perspective:
→ Test sockets are far more stable
→ avoids the substrate CAPEX cycle

Short-momentum perspective:
→ Substrate momentum is stronger
→ ASP-hike / capacity-tight headlines are more frequent.

5. Investment priority — different answers for different horizons

5.1 Short-term momentum (3–6 months) — substrates win

Why:
- Earnings-revision pace is faster
- Steady stream of ASP-hike headlines
- Probability of new big-tech LTA
- AI-server shipments accelerating

Order of preference:
1. Daeduck Electronics: turnaround + AI MLB / FC-BGA beta
2. Samsung Electro-Mechanics: AI FC-BGA + MLCC combo
3. Isu Petasys (complement): ultra-high-layer MLB strength

Caveats:
- Substrates have already moved a lot
- Check the CAPEX-cycle position
- Wait for the macro gate to clear (see earlier post).

5.2 Hold 1–2 years (quality growth) — test sockets win

Why:
- Margin structure is structurally superior
- Smaller cycle risk
- AI chip diversification = more socket categories = revenue diversification
- Active capacity expansion (new fabs)

Order of preference:
1. LEENO Industrial: highest quality, but the multiple is rich
2. ISC: direct AI data-center beta
3. TSE (complement): growth at a more reasonable price

Caveats:
- LEENO faces margin-impact risk during the new-fab move
- ISC has quarterly volatility (1Q26 QoQ -6%)
- Both trade at 30–45x PER.

5.3 A combined AI back-end portfolio

Aggressive (momentum tilt):
- Daeduck Electronics 40%
- ISC 30%
- Samsung Electro-Mechanics 20%
- LEENO Industrial 10%

Balanced (growth + stability):
- LEENO Industrial 30%
- Samsung Electro-Mechanics 25%
- ISC 25%
- Daeduck Electronics 20%

Defensive (quality tilt):
- LEENO Industrial 40%
- Samsung Electro-Mechanics 30%
- ISC 20%
- Daeduck Electronics 10%

→ If you must pick one, by horizon and volatility tolerance:
→ 1y+ hold: LEENO Industrial
→ 3–6 months: Daeduck Electronics
→ Direct AI data-center beta: ISC.

6. Four common misconceptions

6.1 #1: “Both are AI plays, so they’re the same”

As shown:
- OPM differs by 3x
- Cycle sensitivity differs
- Revenue mix differs

Grouping them as one bucket weakens diversification.
→ Volatility moves together
→ Pairing with another sector is more effective.

6.2 #2: “Pogo sockets are being replaced by rubber sockets”

Only partially true:

Where the shift happens:
- AI GPU / ASIC large-die chips
- High-current / high-speed signal testing
- SLT (System-Level Test)
→ Rubber-socket penetration rises (ISC strength)

Where it doesn't:
- High-precision R&D testing
- Mobile APs, RF chips
- Small-batch / many-SKU production
→ Pogo retains share (LEENO strength)

→ Not "the whole market converts" — it is "market segmentation"
→ Therefore ISC and LEENO are complements
→ It is rare for only one of them to do well.

6.3 #3: “Substrate stocks have run too much”

The right question is not the absolute level,
it is "where in the CAPEX cycle are we."

Now:
- Shortage still in progress
- ASP-hike negotiations under way
- LTAs with big-tech under discussion
- CAPEX announcements have started, but ramps are 2–3 years away

→ Early- to mid-stage of an up-cycle
→ Price moved, but the cycle has not topped.

That said, from these levels:
- Scaled entries recommended
- Wait for the macro gate to clear
- Chasing is inefficient.

6.4 #4: “The test-socket TAM is too small”

True on the headline number, but the read is wrong:

Test-socket TAM:
- \~USD 3–4 billion
- Less than 2% of the memory market (\~USD 200B)
- Small in absolute terms

Why that's actually a feature:
- Too small for mega-cap entrants
- Sustains the specialist oligopoly
- Price competition stays mild
- 30–50% OPM achievable

Compare:
A USD 200B business at 5% OPM vs
a USD 4B business at 45% OPM
→ Comparable absolute OPDollars
→ The second is more stable and more predictable.

This is the structure LEENO and ISC benefit from.

7. The next six months — checklist

7.1 Substrates (cycle-intact confirmation)

Positive signals:
□ FC-BGA ASP hikes continue
□ Samsung E-M / Daeduck add new big-tech customers
□ Large-die / high-layer mix lifts in revenue
□ MLB demand persists at 800G / 1.6T networking
□ Utilization stays above 90% even with CAPEX rising

Negative signals:
□ ASP hike negotiations delayed / fail
□ Big-tech AI server orders decelerate
□ Pace of new CAPEX announcements accelerates (glut concern)
□ Utilization drops below 80%

Frequency: quarterly earnings + mid-quarter IR commentary.

7.2 Test sockets (growth-intact confirmation)

ISC:
□ 3Q26 revenue re-acceleration (2Q may be a ramp-prep quarter)
□ New data-center infrastructure customer first-volume ramp
□ SOCAMM2 mass-production testing revenue begins
□ SLT mix stays above 70%
□ AI revenue mix stays above 80%

LEENO Industrial:
□ Progress on the new-fab move
□ No margin damage during the move (OPM 45%+ holds)
□ Customer diversification (Apple / TI / HPC / ASIC expansions)
□ Strong R&D socket demand
□ No overhang / governance issues

Frequency: quarterly earnings.

7.3 The macro backdrop

The macro gate from the earlier post:
- US 10-year below 4.45%
- Brent below 105
- USD/KRW below 1,480
- VIX below 18

These must clear for:
- Broad risk-asset recovery
- Back-end stocks to ride the trend
- Good earnings to translate into good price.

If the gate doesn't clear:
- Multiples stay pressured independent of fundamentals
- Wait for confirming close / turnover before scaling in.

8. How this connects to other posts

Samsung Electro-Mechanics piece:
→ MLCC + FC-BGA 1Q26 dual beneficiary
→ The most detailed stock covered in this article's substrate side
→ Scenario PERs for how far the multiple can stretch

Jeju Semiconductor piece:
→ "Commodity memory in shortage because of AI"
→ Another AI back-end shape (memory shortage beta)

Samsung Electronics strike piece:
→ Key variable for the memory supercycle
→ AI silicon → AI server → memory / substrate / test socket
→ Disruption in one ripples into the back end

KOSPI crash + macro-gate piece:
→ "Cycle before the stock"
→ Names in this post also rational only after the macro gate clears.

9. The one-line bottom line

When AI silicon sells, the first beneficiaries are the GPU and HBM makers. But the back end also has two big winners — substrates and test sockets.

Both are grouped as “AI back-end,” but the structures are completely different. Substrates are an “AI server volume beta.” Volumes rise, packages get bigger, ASPs lift — direct upside. Momentum is fast and operating margins expand from a ~12–15% base in 1Q26. But this is a CAPEX-cycle industry, and the build-out-complete date is the risk.

Test sockets are a “chip-complexity beta.” As chips get more complex, testing gets harder, and new sockets are needed each time. OPM sits at 35% for ISC and 47% for LEENO — roughly 3x substrate margins. Consumable economics dampen cyclicality. The downsides: the TAM is small and multiples are already rich.

Don’t lump them as one “AI theme.” For short-term momentum, substrates (Daeduck Electronics, Samsung Electro-Mechanics) lead. For 1–2 year holding, test sockets (LEENO Industrial, ISC) make more sense. The best play is to hold both as complements — when substrates shake at a cycle peak, test sockets defend; when test sockets feel multiple pressure, substrates’ momentum carries.

Same AI tailwind, completely different structures — understanding just that lifts the quality of back-end investment decisions by a full step.

This article is research and commentary only and is not investment advice. Samsung Electro-Mechanics 1Q26 figures are per the company’s official IR release. Daeduck Electronics 1Q26 (revenue KRW 346.3B, OP KRW 51.3B, OPM 14.8%) are per its IR materials. ISC 1Q26 (revenue KRW 68.3B, OP KRW 23.6B, OPM 35%, AI revenue mix 81%, data-center mix 79%) are per its IR materials. LEENO Industrial 1Q26 (revenue KRW 99.77B, OP KRW 47.30B, OPM 47.4%) is per preliminary disclosure reporting; product-mix figures (IC test sockets 64.10%, LEENO pins 24.65%, medical components 10.46%) are per the quarterly report. OPM uses operating profit divided by revenue, rounded to one decimal place. AI / data-center revenue shares are derived from company disclosures. The operating-margin comparison reflects a single quarter (1Q26); annual averages may differ. CAPEX-cycle risk, new-fab transition margin risk, and AI demand volatility are author judgments and not certainties. Global macro variables (US rates, oil, FX, VIX) can independently move the stocks. The analysis may be wrong. Data cut-off: May 15, 2026 KST.