6531 — AP Memory Technology

Thesis

Verdict: PASS at NT$1,125 / WATCH NT$650–800 / BUY NT$400–550. Medium conviction — high on the business and the secular thesis, low on the entry price. AP Memory is a fabless Taiwanese IC designer that has stapled a genuine AI-packaging growth engine onto a mature, cash-generative specialty-memory base. The base is low-power PSRAM (IoTRAM, ApSRAM) for wearables and IoT; the thesis is stack silicon capacitors (S-SiCap) — passive silicon die that deliver clean, ultra-fast power inside the most advanced AI accelerator packages, in the three places ceramic capacitors physically cannot (above ~100 MHz, ultra-thin, embedded-in-substrate). As of Q1 2026, S-SiCap is no longer a rounding error: NT$570M in the quarter, +767% YoY, ~27% of revenue. AP Memory is the first qualified supplier into Intel's EMIB-T bridge socket, with discrete-capacitor (IPD) shipments starting Q2 2026 into the Google/MediaTek "TPU 8i" inference ASIC.

The price has already eaten it. At NT$1,125 the market cap is NT$183.2B (~$5.86B); against a fortress balance sheet of ~NT$11.5B net cash, EV is ~NT$171.7B (~$5.49B). (yfinance now prints NT$178.8B EV, down from the erroneous NT$189.8B; still slightly high — the company has near-zero debt.) Trailing ~114–147x earnings (yfinance's TTM P/E is 147x on a NT$7.65 trailing EPS base). The forward multiple is the trap: the 43.6x screeners show rests on a yfinance consensus FY26 EPS of NT$25.80 that no primary source supports — Morgan Stanley now models NT$17.65 (raised 21 May from NT$14.01), local Taiwanese sell-side clusters at NT$16–18 (one bear at NT$10.4) — so the true forward multiple is ~63–70x. NT$25.80 quietly assumes earnings rise 3.3x off the FY25 base of ~NT$7.7 in a single year; the Q1 2026 run-rate (NT$4.15/qtr) annualizes to ~NT$16–17, so even MS's own number sits well below the yfinance figure. The "cheap 43.6x" is a data artifact.

The sell-side has caught up to the rally rather than calling the top. The earlier MS-target discrepancy is now resolved (2026-06-02): Morgan Stanley is Overweight at NT$1,555, +38% above spot — raised in steps through the spring (NT$666 on 19 Apr → NT$777→NT$1,000 on 29 Apr, named a Top Pick → NT$1,555 on 21 May, FY26/27/28 EPS NT$17.65 / NT$34.34 / NT$60.74, on the thesis that silicon caps stay short for three years). The NT$666 and NT$777 figures the page carried were real but stale intermediate targets, and the "NT$1,555 bull case" was not a community number — it is Morgan Stanley's house target. The honest reading: the two-analyst mean PT (NT$1,075) sits slightly below spot, while the most-followed name (MS) is +38% above it; the spread is the whole debate. That re-frames but does not erase the asymmetry. MS gets to NT$1,555 only by underwriting EPS toward NT$60+ by FY28 — i.e. by assuming the 1:1:1 ramp lands and the wafer shortage holds. Bake in normal slippage on a single-design, single-foundry ramp and the downside to the NT$650–800 fair-value zone (−29% to −42%) still outweighs the base-case upside. The next gate is the Q2 2026 print (confirmed 30 July). This is a good company you want to own; at NT$1,125 you are paying the analyst, not buying ahead of him. Wait for the parabola to break — or for a print that re-anchors EPS.

Snapshot

One-liner: AP Memory designs the silicon capacitors that deliver clean power inside AI chip packages where ceramic capacitors physically cannot, plus a legacy low-power PSRAM franchise for wearables/IoT — the purest listed silicon-cap play on AI advanced packaging, first-qualified into Intel's EMIB-T socket, but already pricing most of the optionality.

• Ticker / exchange: 6531 on the Taipei Exchange (TPEx, formerly TWO; Yahoo carries it as 6531.TW / "TWSE")
• Legal name: AP Memory Technology Corporation (愛普科技股份有限公司); HQ No.1 Taiyuan 1st Street, Zhubei, Hsinchu Science Park; founded 2011; listed 2017 (TPEx); apmemory.com
• Sector / industry: Information Technology / Semiconductors (specialty memory + silicon capacitors). Sits at the silicon-cap ⨯ AI advanced-packaging intersection.
• Spot price: NT$1,125 (close 2026-06-01; spiked to NT$1,190 on 29 May then pulled back); +145% YTD, +330% off the 52-wk low, −14% off the new intraday high (52-wk range NT$262–1,305 intraday; close range NT$262–1,190)
• Market cap: NT$183.2B (~$5.86B) | EV ~NT$171.7B (net cash ~NT$11.5B; near-zero debt; yfinance now shows NT$178.8B)
• Valuation: Trailing P/E ~114–147x | Forward P/E ~63–70x (NOT the 43.6x screeners show — that uses a corrupted NT$25.80 EPS) | P/B 16.1x | P/S ~27x | EV/EBITDA ~91x TTM | FCF yield ~1.2% | Dividend yield 0.62% (NT$7.00, 91% payout) | beta 1.67
• Shares: 162.85M; insiders 36.6% (founder Chen Wen-Liang + aligned), institutional ~13.4%, float ~50% (~122M)
• Coverage: thin — 2 analysts on the aggregators (mean PT NT$1,075, below spot; high NT$1,150 / low NT$1,000 on the screener tape). The MS-target discrepancy is RESOLVED (2026-06-02): Morgan Stanley is Overweight at NT$1,555 (+38%), raised in steps — NT$666 (19 Apr, FY26 EPS NT$14.01) → NT$777→NT$1,000 (29 Apr, named Top Pick) → NT$1,555 (21 May, FY26/27/28 EPS NT$17.65 / NT$34.34 / NT$60.74, 2028 P/E ~14.4x). The NT$666 and NT$777 the page previously carried were genuine but superseded intermediate targets; the "NT$1,555 bull case" is now confirmed to be MS's own house target, not a community estimate. MA50 765 / MA200 486 (extended); the stock is mid-pullback off the 29 May high.
• Conviction: Medium — PASS at spot; WATCH NT$650–800; BUY NT$400–550

Business

Business model. Fabless. AP Memory owns the IP, runs design, qualification and customer engagement, and outsources wafer manufacturing — primarily to Powerchip (PSMC), with whom it co-markets a "3D AI Foundry" strategy (AMD named as a target customer). Capex ran ~NT$55M in FY2025 against NT$5.67B revenue (sub-1% intensity) because the heavy fab capital sits on Powerchip's balance sheet — the source of 46% gross margin, ~49% FCF margin, and the net-cash position on a business this small. The PSMC relationship is structural, not arm's-length: Powerchip's holding vehicle sits on AP Memory's board via Li Shun Investment (director Hsieh Ming-Lin, President of Powerchip Investment Holding).

Segment mix. No formal segment table is published; the Q1 2026 call (12 May) gave the most precise split to date:

• IoTRAM / PSRAM — still the largest contributor. Low-power pseudo-SRAM for wearables, IoT, displays, connectivity. Within IoTRAM, connectivity modules drove 42% of growth, display 31%. New-gen ApSRAM entered mass production Q4 2025 (10+ design-ins; validates with Arasan xSPI IP).
• S-SiCap — NT$570M in Q1 2026 (+767% YoY, ~27% of revenue). Two sub-lines: IPC (silicon-capacitor interposer for 2.5D packaging) >25% of S-SiCap revenue, stable shipments, four-reticle mass production since end-Q3 2025; IPD (discrete silicon capacitor) validation complete, ramps from Q2 2026.
• VHM (Very High-bandwidth Memory) — stacked custom DRAM for AI/edge; NRE-stage only, mass production not expected until after late 2027.
• Design services / licensing — small.

The single most useful disclosed fact: management's long-term mix target IoTRAM : S-SiCap : VHM = 1 : 1 : 1 — a third each. That implies S-SiCap roughly triples its share, the spine of the bull case and the source of the forward-EPS heroics.

First principles — what S-SiCap is and why it exists. A modern AI accelerator switches billions of transistors at multi-GHz; every switch pulls current, and fast current change (high dI/dt) through any path inductance produces voltage droop (V = L × dI/dt) that crashes timing margins. You need capacitance — a local charge reservoir — as electrically close to the silicon as possible with minimal parasitic inductance. MLCC fails above ~100 MHz: a ceramic cap soldered to the package has 0.3–1 nH equivalent series inductance (ESL), self-resonates, and stops behaving like a capacitor — physics, not laziness. A silicon capacitor is fabricated on a silicon wafer with deep-trench etch (lined with ALD-deposited Al₂O₃/HfO₂, then metallized): an 80–100 µm-thin passive die with ESL below 0.1 nH (~100x lower than MLCC), effective above 1 GHz, and — because it's a silicon die — embeddable inside the package (EMIB bridge, CoWoS interposer, ABF substrate), micrometres from the logic. Silicon caps don't replace MLCC in volume; they complement it in exactly the three places AI advanced packaging needs at once.

• IPD (discrete): Gen4 (Dec 2025) hits 3.8 µF/mm² (+50% vs Gen3), first to adopt embedded-substrate packaging; qualified into Intel EMIB-T; ramps Q2 2026.
• IPC (interposer): silicon-cap function built into a 2.5D interposer (CoWoS-S-style); stable shipment; four-reticle mass production end-Q3 2025.

The roadmap extends to silicon caps for organic interposers — cheaper than silicon interposers, the direction EMIB-T pushes as the lower-cost alternative to TSMC's full-silicon CoWoS; embedding silicon caps into organic substrates is "where the next socket lives," the bridge from one design win to a platform direction.

Value chain & the bottleneck. AP Memory sits at the design/IP node: PSMC fabricates wafers; TSMC's CoWoS consumes the IPC variant; OSATs assemble; the customer is the package integrator (Intel for EMIB) and ultimately the hyperscaler ASIC owner (Google, via the MediaTek-co-designed TPU 8i). The binding constraint on the whole merchant market is deep-trench silicon-cap wafer capacity at advanced foundries, not end demand — and the asset-light model is double-edged: AP Memory can't fund its own capacity, so it's hostage to PSMC/TSMC allocation while AI silicon caps compete for the same slots as logic and memory. Two Q1-call flags: (1) a PSMC facility transition caused a 2–3 month IPC shipment delay; (2) second-source foundry qualification is underway (more foundries to tape out S-SiCap this year) — de-risking that confirms the dependence was a risk. PSMC also sold a fab to Micron for ~$1.8B (Jan 2026), tightening available wafers — a second-order squeeze management has not addressed.

Customers & concentration. AP Memory does not disclose names. Legacy PSRAM is naturally fragmented (MediaTek, Realtek, Allwinner, Goodix, Qualcomm). The S-SiCap pivot introduces fresh, sharp concentration: the EMIB-T design win is one program — the Google/MediaTek TPU 8i inference ASIC (codenamed Zebrafish; single compute die + I/O die + six HBM3e stacks, TSMC 2nm). Timing clarified (2026-06-02): the 2nm part is a late-2027-class design, but the inference-chip volume ramp now points to H2 2026 — third-party tracking puts mass production starting ~July–August 2026, ~400k units in 2026 rising to 1.5–2.0M in 2027 (TPU v8 splits into Broadcom-built training "Sunfish/8t" and MediaTek-built inference "Zebrafish/8t/8i"). This pulls AP Memory's IPD revenue contribution earlier than the page's prior "late-2027" framing implied and lines up with management's "IPD mass production from Q2 2026" guidance — a modest de-risking of the single-design timing, though not of the single-design concentration. A slip there still slips the entire IPD ramp; the market is not pricing the single-design risk. Structural partners: PSMC (foundry + board-represented via Li Shun); TSMC (CoWoS packaging adjacency, customer/supplier not equity-tied); Intel (EMIB socket); AMD (PSMC 3D AI Foundry target).

Competitive set & moat. Silicon-cap: Murata (#1 merchant, Caen 200mm line, tripling capacity by 2028 but largely internal demand) · SEMCO (#2 emerging, W1.557T 2027–28 contract, scaling fastest, valuation extreme) · TSMC (in-house IPD bundled into CoWoS, competes for the same wafers) · AP Memory (#4, first into EMIB-T, highest relative revenue exposure) · niche (Empower, Vishay). PSRAM: Lyontek (5650.TWO), Etron (5351.TWO), Winbond/Macronix (2337.TW) — moderate moat. The S-SiCap moat is stronger than the small base suggests: 12–18-month qualification cycles, and once embedded in the substrate, swapping requires substrate redesign + re-qualification (high switching cost, multi-year lock-in). The first-qualified EMIB-T position is, for now, effectively uncontested — Murata's wafer line is captive to its own customers and SEMCO's large supply order is for a different end product — though that lead can erode as rivals qualify in the 2027 window. 3-test: (1) 5-yr lock-up — partial pass (real design lock-in, but positions are young and the field is contested by two larger rivals); (2) unique economic engine — pass on quality, unproven on durability (the ~49% FCF-margin asset-light model is real and rare); (3) blank-check disruptor — fail (a component supplier into others' platforms, a beneficiary of an inflection, not its author). Cycle: early-to-mid in the AI-packaging silicon-cap S-curve; the stock is late-cycle in its own re-rating.

Financials

Core Four. Growth: revenue FY22 NT$5.10B → FY23 NT$4.23B (−17%) → FY24 NT$4.19B (−1%) → FY25 NT$5.67B (+35%); TTM NT$6.79B; Q1 2026 NT$2.1B (+115% YoY, +12% QoQ) on S-SiCap +767% and the IoTRAM ramp — real acceleration, durability past the first design win is the open question. Margins: gross 46.4% TTM (guided ~45% steady-state), operating 28.0%, net 23.6% (TTM net depressed by a one-time Q2 2025 charge — NT$684M EBIT / NT$546M net loss, likely qualification NRE / inventory). Incremental margins are high: operating income +236% YoY in Q1 2026 on +115% revenue — the legitimate engine behind the forward-margin-expansion thesis. Capital intensity: sub-1% of revenue. Deployment: 91% payout, no buyback, ~NT$11.5B net cash; a NT$7.00/share dividend (~NT$1.14B) goes ex 2026-06-22.

Valuation (current, 2026-06-01). Trailing P/E ~114–121x (the 145x yfinance prints uses a calendar-year EPS base); forward ~63–70x on a realistic NT$16–18 EPS (the 43.6x is a data artifact — see Valuation/DCF); P/B 16.1x; P/S ~27x; EV ~NT$171.7B. ROE 13.3% (deflated by the net-cash pile; underlying ROIC much higher on the small operating-capital base, comfortably above a ~9–10% WACC).

Income statement & margins (calendar year, NT$M):

Metric	FY2022	FY2023	FY2024	FY2025	FY2026E (consensus-implied)
Revenue	5,095	4,227	4,192	5,666	~8,400–9,000
YoY growth	n/m	−17%	−1%	+35%	+48% to +59%
Gross profit	2,221	1,755	2,146	2,634	n/d
Gross margin	43.6%	41.5%	51.2%	46.5%	n/d
Operating income	1,500	882	1,064	1,398	n/d
Operating margin	29.4%	20.9%	25.4%	24.7%	n/d
Net income	1,942	1,445	1,578	1,258	~2,700–2,900 (realistic)
Net margin	38.1%	34.2%	37.6%	22.2%	~32–34% (realistic)
EPS (basic, NT$)	11.99	8.92	9.74	~7.7	NT$16–18 realistic (MS NT$17.65); 25.80E on yfinance is corrupted — see stress test
R&D / Revenue	9.1%	13.8%	18.6%	15.5%	n/d

(FY25 net-income compression despite +35% revenue is the one-time Q2 2025 charge; clean operating margin is structurally 25–28%. EPS shows as NT$7.67 in the profile and NT$7.73 in the deep-dive — both retained; ~NT$7.7.)

Quarterly cadence (NT$M):

Quarter	Revenue	YoY	GM%	Net income	EPS (NT$)
Q4 2024	1,225	—	53.2%	502	n/d
Q1 2025	975	+30%	46.5%	332	2.03
Q2 2025	1,329	—	42.2%	−546	−3.36
Q3 2025	~1,495	+17%	n/d	n/d	4.32
Q4 2025	1,868	+52%	49.9%	766	4.71
Q1 2026	2,099	+115%	46%	676	4.15 basic (+103% YoY; NT$3.65 adj ex-GDR-FX; NT$4.00 on yfinance, +27% vs NT$3.14 est.)

Cash flow & balance sheet (NT$M):

Metric	FY2022	FY2023	FY2024	FY2025
Operating cash flow	1,575	2,043	1,512	2,853
Capex	(45)	(23)	(57)	(55)
Free cash flow	1,529	2,021	1,454	2,797
FCF margin	30.0%	47.8%	34.7%	49.4%
Net cash	8,118	8,528	4,065	7,031
Dividends paid	(968)	(1,132)	(1,135)	(1,137)
ROIC (rough)	~28%	~15%	~16%	~13%

Fortress balance sheet: near-zero debt, ~3.97x current ratio. (FY25 year-end net cash NT$7.0B; the current ~NT$11.5B figure reflects 2026 build, trimmed ~NT$1.14B by the 22 Jun dividend.)

The heroic forward EPS, stress-tested. yfinance's "consensus" FY26 EPS NT$25.80 vs FY25 ~NT$7.7 = 3.3x. To get there off a ~NT$5.67B base you need revenue toward ~NT$10–11B and net margin expanding to ~47–50%. The Q1 2026 run-rate (NT$2.1B/qtr, NT$4.15 EPS) annualizes to ~NT$8.4B and ~NT$16–17 EPS — well short of NT$25.80. The credible cluster is NT$16–18: Morgan Stanley raised its FY26 number to NT$17.65 (21 May, up from NT$14.01), local sell-side sits NT$16–18, one bear at NT$10.4. That puts the real forward multiple at ~63–70x, not 43.6x. The "cheap" 43.6x is an illusion resting on a single yfinance estimate no primary source corroborates. Note the convergence: MS's own number (NT$17.65) now sits inside the local cluster, so the FY26 disagreement has narrowed — the genuine debate has moved out to FY27–28, where MS underwrites NT$34.34 then NT$60.74 (the 1:1:1 ramp delivering in full).

Industry landscape

A silicon-cap component play riding the AI advanced-packaging inflection and the broader passives content explosion. Industry-wide detail lives on the sector pages — see advanced-packaging (the package-architecture shift to 2.5D/3D — CoWoS-S/L, EMIB, Foveros — that creates the silicon-cap socket) and passives-mlcc (the MLCC/passives peer cluster: Murata, SEMCO, TDK). Sector primer: silicon-capacitor-primer-2026-05-28. Merchant silicon-cap TAM ~$4.0B in 2026 → ~$5.9B by 2031 (Mordor, ~8% CAGR); the AI advanced-packaging slice is smaller but far faster-growing, gated by foundry capacity not demand — Citi's SEMCO model implies AI silicon-cap external TAM ~$1.5–2.5B by 2027–28, transformative against AP Memory's ~$185M FY25 base.

Management

Founder Chen Wen-Liang (Ph.D. Applied Physics, Yale) is both Chairman and CEO. President/GM Hung Chih-Hsun (Ph.D. NYCU) runs operations and holds a board seat. No separately named CFO or CTO on the public governance page — a disclosure-quality flag (typical for Taiwan small-mid caps). Governance officer Hong Shu-Ling since May 2021.

Board: 8 members, 4 independent (50%, above Taiwan small-cap norm). Deep supply-chain DNA — a feature for technical credibility, a flag for related-party transparency:

Name	Role	Independent?	Background / affiliation
Chen, Wen-Liang	Chair	No	Ph.D. Applied Physics, Yale — CEO (executive chair)
Hung, Chih-Hsun	Director	No	Ph.D. NYCU — President of AP Memory
Yeh, Jui-Pin	Director (Shanyi Investment)	No	M.S. EE, NCU; director at Mytek
Hsieh, Ming-Lin	Director (Li Shun Investment)	No	MBA NTU — President of Powerchip Investment Holding (direct PSMC linkage)
Wang, Hsuan	Indep	Yes	Ph.D. Accounting, NTU; CPA + CIA
Sun, Elizabeth	Indep	Yes	Ph.D., U. Cincinnati; consultant at MediaTek
Chen, Ken	Indep	Yes	Ph.D. Materials Science, Stanford; CBO Brillnics, consultant at Unimicron
Liu, Frank	Indep	Yes	MIT MS EECS — Chairman, Chelmsford International

Alignment: insider ownership 36.6%, founder-aligned — real skin in the game; no dual-class shares. Capital allocation — grade B−: the pointed critique is that AP Memory pays out 91% of earnings as dividends (yield 0.59%) with no buyback, while sitting on the steepest growth ramp of its life. It works only because the model is asset-light and PSMC carries the capex; forward NI of ~NT$4.2B against a ~NT$1.14B dividend run-rate leaves headroom that would be better spent on a second foundry relationship or a buyback at a sane price than out the door at 91% north of 60x a realistic forward number. Share count crept <1% over three years — not a serial diluter. Key-person: founder-as-chair-and-CEO, no disclosed succession plan — typical Taiwan founder-led profile; flag, not dealbreaker.

Catalysts & risks

Catalysts — near-term: Q2 2026 print (confirmed 30 July) is the gate — watch IPD shipment commencement, S-SiCap trajectory (does NT$570M hold/grow despite the 2–3-month IPC delay?), ~45% GM defense, forward-mix color. A clean beat-and-raise validates the rally; a wobble on the PSMC-transition delay catalyses 30–40% downside on positioning alone. Monthly Taiwan revenue disclosures give a real-time IPD-ramp read; second-source foundry tape-out confirmation guided "this year." The TPU 8i inference volume ramp now points to H2 2026 (mass production ~Jul–Aug 2026, ~400k units 2026 → 1.5–2.0M 2027), so the IPD contribution should show up in 2H26 monthly revenues rather than waiting on the late-2027 silicon — a near-term read that did not exist when the page was first written. Medium-term: IPD volume scaling on the Google/MediaTek TPU 8i program through 2026–2027; Intel EMIB silicon-cap mass adoption in 2027 (per TrendForce) — the EMIB platform's broader demand spans Microsoft Maia, Amazon Trainium, and Meta MTIA, though AP Memory's confirmed content today is the single TPU 8i (Zebrafish) inference program; VHM mass production "late 2027 to 2028" (the third 1:1:1 leg, per the Q1 call); sell-side coverage expansion beyond two analysts.

A material ambiguity at the heart of the thesis: the 28 May TrendForce piece on Intel adopting silicon caps for Google v8e in 2027 names only Murata and SEMCO — not AP Memory — while Chinese-language coverage and AP Memory's own positioning put it first-qualified into EMIB-T for the Google/MediaTek ASIC. Most likely two different EMIB programs (v8e training, where Murata/SEMCO are named; TPU 8i/Zebrafish inference, MediaTek-co-designed, where AP Memory is named). Flagged as inference, not fact.

Risks (structured):

1. Valuation overshoot (High) — +145% YTD, above the 2-analyst mean PT; the single most likely source of loss is multiple compression, independent of the business.
2. The screen multiple is fictional, not the EPS (High → Medium) — the NT$25.80 yfinance "consensus" is the artifact, not a real estimate to be cut; the credible FY26 number is NT$16–18 (MS NT$17.65), which the stock already trades on at ~63–70x forward. The risk is no longer "consensus gets cut" (the real estimates are already low) but "the market keeps paying ~65x and that multiple compresses." Re-rating risk, not estimate risk.
3. Single-design concentration (Medium-High) — the IPD ramp hinges on one program (TPU 8i).
4. PSMC capacity / transition (Medium) — already a 2–3-month IPC delay; the Micron fab sale tightens wafers; single-foundry dependence not yet resolved.
5. Competitive erosion from Murata/SEMCO (Medium-High) — both larger and better-capitalized; TrendForce names them (not AP Memory) for v8e.
6. Disclosure opacity (Medium) — no segment table, no named CFO, 2-analyst coverage; the market is valuing a black box on press-release fragments.
7. IoT/wearable cyclicality (Medium) — base business is cyclical; durability rests on LTA prepayment agreements.

Dilution risk: Low (share count <1% creep in three years; no shelf/ATM; 91% payout signals distribution-over-issuance).

Bear case. Q2 disappoints on the PSMC delay; the market stops paying ~65x forward on a NT$16–18 base and de-rates to ~30x = ~NT$540, overshooting toward the pre-breakout NT$400–550 base. Downside target NT$450–550 (−51% to −60%). Bull-invalidation: Q2/Q3 monthly revenues stall, IPD slips past 2026 despite the H2-26 ramp signal, or AP Memory is confirmed out of the major EMIB sockets. Bear-invalidation (forces a re-rate higher): clean Q2 beat-and-raise that lifts the realistic FY26 EPS toward NT$20+ and validates the FY27 step (MS NT$34.34), plus confirmation of multi-program EMIB content beyond the single TPU 8i design. Morgan Stanley at NT$1,555 (+38%) is the standing bull-invalidation marker for the bears.

Valuation / DCF

Valuation grade: expensive on any honest near-term number; defensible only on a 2028 horizon. Two clocks:

• On 2026 — expensive. ~63–70x forward on realistic FY26 EPS (the 43.6x screens show rests on the corrupted NT$25.80 yfinance estimate; MS NT$17.65, local NT$16–18). Trailing ~114–147x; P/B 16x. Price sits above the 2-analyst aggregator mean PT (NT$1,075) but ~38% below Morgan Stanley's house target of NT$1,555 — the wide gap between the screener tape and the lead-bank target is itself a tell that the aggregators have not refreshed. No margin of safety on near-term numbers regardless of which target you anchor to.
• On 2028 — defensible, if the ramp delivers — and now with a named bank behind it. Morgan Stanley's NT$1,555 OW is explicitly a 2028-horizon call: FY26/27/28 EPS NT$17.65 / NT$34.34 / NT$60.74 (silicon-cap shortage holding three years, 1:1:1 mix delivering), putting today's NT$1,125 at ~18x MS's 2028 number — reasonable-to-cheap for that growth if you believe the ladder. This supersedes the earlier reliance on the Collyer Bridge "Humufish/EMIB-T" community math (10k wafers/month × 12 × $7k ≈ ~NT$26bn IPD revenue by 2028) — the community estimate and MS's model now point to the same NT$50–70 EPS zone by 2028, which raises confidence in the shape of the bull case without resolving its three load-bearing assumptions: (1) PSMC allocates the wafers, (2) ASP holds as Murata/SEMCO add capacity, (3) the EMIB programs convert to volume. The negative case is concrete and unchanged: PSMC just sold a fab to Micron, tightening the very capacity the thesis needs.

Multiples vs peers (forward): SEMCO ~77x (richer, half the margins, negative FCF); Murata ~41x (genuinely cheaper for the better business); Yageo/Macronix 12–20x (lower growth). On the illusory 43.6x AP Memory looks most defensible of the trio; on the realistic ~63–70x it is the most expensive on forward earnings, above SEMCO only because SEMCO's earnings base is even more depressed. AP Memory's premium is pure AI optionality, fully extended.

Entry zones (multiples on the realistic FY26 EPS of ~NT$17–18, not the corrupted NT$25.80):

• WATCH NT$650–800 (−29% to −42%) — re-engage if a Q2/Q3 print validates the IPD ramp (FY26 EPS holding NT$17–18 with the PSMC delay cleared and the H2-26 TPU ramp confirmed in monthlies); ~36–47x forward, the upper bound of fair for a high-margin specialty franchise with proven AI optionality, and ~9–14x MS's FY28 number if you underwrite the ladder.
• BUY NT$400–550 (−51% to −64%) — the pre-breakout base; ~22–32x forward on NT$17–18 EPS. Where you are paid for the disclosure opacity and single-design risk.

Decision log

2026-06-02 — Deep-dive refresh (live data + MS-target resolution). Verdict UNCHANGED: PASS at NT$1,125 / WATCH NT$650–800 / BUY NT$400–550, Medium conviction. The business thesis and the entry discipline hold; what changed is the sell-side context and three data corrections:

• MS price-target discrepancy RESOLVED. The NT$777-vs-NT$666 flag was stale on both counts. Morgan Stanley walked the target up through the spring — NT$666 (19 Apr) → NT$777→NT$1,000 (29 Apr, named Top Pick) → NT$1,555 OW (21 May), FY26/27/28 EPS NT$17.65 / NT$34.34 / NT$60.74. Current live MS target = NT$1,555 (+38% above spot), not below it. The "NT$1,555 bull case" the page treated as a community number is in fact MS's house target. This does not flip the verdict — MS gets there only by underwriting NT$60+ FY28 EPS, and the +38% upside is still less than the −29%-to−42% downside to fair value on a single-design, single-foundry ramp. But it sharpens the bull-invalidation marker for the bears.
• MS FY26 EPS is NT$17.65, not ~NT$14. Stale on the page. The FY26 estimate debate has narrowed (MS now inside the local NT$16–18 cluster); the real disagreement has moved to FY27–28. yfinance's NT$25.80 remains a corrupted outlier — the 43.6x forward multiple is still fictional; real forward ~63–70x.
• Q1 2026 EPS is NT$4.15 basic (NT$3.65 adj ex-GDR-FX; NT$4.00 on yfinance), not NT$4.00. Minor.
• TPU 8i (Zebrafish) volume ramp pulled into H2 2026 (mass production ~Jul–Aug 2026, 400k units 2026 → 1.5–2.0M 2027), vs the page's prior "late-2027" framing — modest de-risking of the single-design timing (IPD revenue should appear in 2H26 monthlies), not the single-design concentration. Confirmed next print 30 July.
• Live (close 2026-06-01): NT$1,125; mcap NT$183.2B; EV ~NT$171.7B (yfinance NT$178.8B); trailing P/E ~114–147x; forward ~63–70x; 52-wk intraday range NT$262–1,305; NT$7.00 dividend ex 22 Jun.

Net: refresh confirms the prior call. A good company you want to own, at a price you should not pay. The sell-side caught up to the rally rather than leading it; wait for the Q2 print (30 Jul) or the parabola to break.

2026-06-01 — Consolidated verdict (deep-dive + pre-buy checklist, both 2026-06-01; reaffirms the 2026-05-28 profile):

• Deep-dive — PASS at NT$1,125 / WATCH NT$650–800 / BUY NT$400–550, Medium conviction. High on the business and secular thesis, low on entry. Adversarial fact-check caught the headline "43.6x forward" as a corrupted yfinance EPS (NT$25.80); real forward ~63–70x; MS PT NT$666 ≈ −41%. Bear target NT$450–550. Next gate Q2 (~30 Jul); NT$7 div ex 22 Jun.
• Pre-buy checklist — WATCH at NT$1,125 (starter-only for a high-conviction 2–3yr horizon). Scorecard: thesis clear ✓, business understood ✓, FundamentEdge gates clear (valuation flagged), incentives partially aligned (36.6% ownership vs 91% payout), financials healthy ✓, valuation No on 2026 / conditional on 2028, 3 of 4 behavioral-trap flags lit (FOMO + narrative seduction + authority bias — evaluating after a +145% move, prompted by a bull chat), technicals extended (not a clean entry). The Collyer Humufish/EMIB-T thread is a genuine, thesis-broadening positive (reduces the single-design risk; on a 2028 horizon makes today ~16–23x) but rests on unverified community estimates + three load-bearing assumptions. Do not chase full size at NT$1,125.
• Profile (2026-05-28) — PASS at NT$1,080. Established that the silicon-cap thesis is real, disclosed, and ramping (not speculative TSMC-adjacency framing): S-SiCap is a distinct disclosed product line, Intel EMIB qualification is independent of CoWoS adjacency, and AP Memory is a credible top-4 merchant with the highest % revenue concentration to the AI thesis.

Position plan. PASS at spot — watchlist slot, not capital. WATCH NT$650–800 (bulk; ~25–31x forward, real margin of safety); BUY NT$400–550 (deeper tranche). Size below a normal slot even in the buy zone — 2-analyst coverage, founder-as-chair-and-CEO, single-design concentration, TPEx liquidity. Scale the second half only after a second EMIB program or a second foundry is confirmed. Scale-in triggers: Q2 (30 Jul) confirming the IPD ramp live; second-source foundry tape-out; multi-program EMIB confirmation. Hard re-evaluate / max loss −25% from any entry; disconfirmed PSMC allocation is a hard exit. Theme construction: pair AP Memory (thesis-pure, asymmetric) with Murata (6981) (quality, scale, pricing leader) on a joint pullback; avoid stacking SEMCO (009150) on top — it is the short leg of this theme, not a third long.

Net stance: a good company you want to own, at a price you should not pay. Wait for the parabola to break.

Sources

Fragments folded into this canonical page (consolidated 2026-06-01; originals archived to _migration-archive/2026-06-01/6531/): 6531-tw-deep-dive.md (spine, 2026-06-01) · 6531-tw-profile.md (2026-05-28) · 6531-tw-buy-checklist.md (2026-06-01).

Appears in / comparisons:

• 6531-vs-4092-showdown — AI-passives supply-chain: silicon-cap COMPONENT (6531) vs MLCC dielectric MATERIAL (Nippon Chemical 4092); both PASS for opposite reasons (composite 3.53 / 2.83)
• 6531-vs-6981-vs-009150-showdown — silicon-cap showdown: AP Memory vs Murata vs SEMCO; risk-adjusted rank AP Memory > Murata > SEMCO (3.55 / 3.50 / 2.43), all PASS at parabolic prices

Related vault pages: advanced-packaging · passives-mlcc · silicon-capacitor-primer-2026-05-28 · Murata (6981) · SEMCO (009150) · mlcc-sector-2026-05-23 · mlcc-peer-swarm-2026-05-20

Filings / earnings history: 6531-filings — chronological earnings-call and filing log (Q1 2026 print, MS target chain, dividend).

Key external sources: yfinance live (close 2026-06-01, pulled 2026-06-02 via python3.12); AP Memory Q1 2026 earnings call (Vocus + Fugle/富果 12 May memo + Sinotrade 豐雲學堂) + PR Newswire S-SiCap Gen4 (18 Dec 2025) + IR governance page; Morgan Stanley target chain via China Times / 旺得富 (19 Apr NT$666 FY26 EPS NT$14.01; 29 Apr NT$777→NT$1,000 Top Pick; 21 May NT$1,555 OW, FY26/27/28 EPS NT$17.65 / NT$34.34 / NT$60.74) — corroborated by two independent Chinese-language searches; China Times article bodies blocked WebFetch (403), figures cross-confirmed via search excerpts and cnyes/愛玩股 aggregators; Digitimes; TrendForce (28 May — Intel/v8e; 12 May — MediaTek dual-packaging); cmoney 6531 EMIB-T analysis; wccftech / TechPowerUp / TheNextWeb / Tom's Hardware / SemiAnalysis X (Google TPU v8 Sunfish/Zebrafish split, H2-26 inference ramp, 400k→2M units); Mordor Intelligence (silicon-cap TAM); Citi SEMCO silicon-cap model; Collyer Bridge chat (Humufish/EMIB-T 2028 community estimates — now corroborated by MS's NT$60.74 FY28 number, still load-bearing-assumption-dependent). SemiAnalysis cross-check (2026-06-02): no dedicated 6531/AP-Memory/silicon-cap coverage in the local SA mirror (only a 2022 NAND piece mentions "deep trench" in an unrelated context); SA validates silicon-cap-for-AI and the TPU v8 architecture at the chip level but names no merchant silicon-cap supplier — no contradiction (SA covers demand/architecture side, AP Memory is supply side).