← back to picks

Listing TPEx 5289 HQ Taipei, TW FY Dec Niche #1 industrial SSD · 7 yrs Net cash · 13-yr dividend FCFF · Dark v3

Excellent business, wrong price — the 2026 memory super-cycle is being capitalized as permanent, -72% margin of safety on normalized earnings, and even the windfall-adjusted bull case never reaches today's price

Sell — peak-earnings, expensive at every regime

Innodisk trades at NT$1,905 after a ~9x move, on a 2026 DRAM/NAND super-cycle that drove Q1'26 net margin to 42% versus a 13-17% structural band. Normalized off FY2025 the DCF prints NT$538.77/sh (pre-governance NT$585.62) — about 72% BELOW the market — and Monte Carlo puts the probability that intrinsic sits under today's price at 100%. It is expensive at every rate regime (low bookend NT$655, high bookend NT$376), and even after adding back the discounted 2026-27 windfall (~+NT$160 → ~NT$700) and granting the aggressive permanent-margin bull (NT$1,013), the price is still 47-63% above value. This is a GOOGL-style counter-example: a great company priced for a super-cycle that does not last.

Sector · Computers/Peripherals (industrial SSD / DRAM modules) Mix · TW 26% · CN 22% · US 19% · DE 10% β 1.15 pin (load-bearing) Rf · 2.31% blended (50/50) (load-bearing) Governance · -8% (one-man Chair=Pres=GM) Quality · ROE ~21% mean · net cash · 13-yr dividend

Intrinsic / share

NT$538.77

base · 2.31% blended TWD rf · 7.86% WACC · post 8% gov · pre NT$585.62 · low bookend NT$655 · high bookend NT$376 · windfall-adj ~NT$700 · bull NT$1,013

Market / share

NT$1,905.00

1 Jun 2026 close · TPEx

Margin of safety

-71.7%

vs intrinsic

Enterprise value

NT$56.96B

70.8% terminal

Cost of equity / debt

7.97% / 2.17%

β 1.15 · CRP 0.69%

Terminal ROIC / g

13.00% / 2.00%

spread ~514bp (ROIC 13.00% vs WACC 7.86%)

What it sells, where it sells

Operating segments

Industrial SSD / storage (incl. data-center E1.S/E3.L) · ~45%

DRAM modules & embedded memory (CXL / DDR5) · ~40%

Edge-AI systems / peripherals / sensing (Aetina, Antzer) · ~15%

A fabless value-add module house bundling compute, memory, storage, sensing and software into pre-validated edge-AI building blocks — the "Five-Layer Cake". The bundle raises switching costs; it does not insulate the firm from a cyclical NAND/DRAM commodity upstream, which is what the price is mispricing.

Country mix (revenue-weighted CRP input)

Taiwan26%

China22%

United States19%

Germany10%

Japan8%

Netherlands4%

South Korea4%

France · Vietnam6%

Weighted by end-customer demand, not production site (Taiwan + China). The 3-year trend shows the AI-server pivot — China 15→22, US 14→19, legacy Taiwan 32→26 — but the destinations are all low-CRP, so the country-risk premium stays a thin 0.69%; geography is not the swing factor here, normalization is.

Background — five things to know before the case

Context the share price doesn't carry on its face. Skim this once and the bull / bear bullets below stop reading as inside baseball.

World's #1 industrial-grade SSD provider — 7 straight years (Gartner, since 2018) and the only industrial-focused name in TrendForce's top-10 DRAM module makers. A genuine 20-year niche leader, not a commodity reseller.
Fabless value-add module house. Buys NAND/DRAM die; designs and assembles industrial SSDs, DRAM modules, embedded I/O and Edge-AI systems. No wafer fab — it rides the memory commodity upstream rather than making it.
The 2026 super-cycle. FY2025 revenue +60% to NT$14.26bn; Q1'26 revenue NT$13.18bn — nearly a full FY25 in one quarter — pushing Q1'26 net margin to 42% and TTM operating margin to ~52% versus a 13-17% structural band. The stock is up ~9x off its 52-week low.
Net-cash, dividend-compounding quality. 10-year mean ROE ~21%, 13 consecutive years of dividends, FY2025 record NT$16.9 cash declared. Book equity jumped NT$9.1bn → NT$14.7bn in a single quarter on the banked windfall.
Founder-led, widely held, one-man board. Chairman = President = GM (Randy Chien, ~1.74%); no controlling shareholder (largest holder Ruiding 7.4%); 4 of 9 directors independent. An incoming NT$3.0bn 0%-coupon CB adds latent dilution. An -8% governance haircut covers it.

Where we diverge from the sector

Each row asks: if we'd used the sector median for this one assumption instead of our override, how much would the share price change? Negative = our override is more conservative (worth less); positive = more aggressive (worth more). Skim down the rightmost column to see which bets are doing the work.

Revenue CAGR (Y1-5)+11.0%≈ demonstrated 2017→2025 peak-to-peak VOLUME CAGR; strips the price spike

Operating margin (Y10)17.0%top of the durable 13-17% band; NOT the 42-52% spike

Sales-to-capital (Y1-5)2.0Yilan Plant 2/3 build + peak-price working-capital balloon

Levered β (pin)1.15vs 0.85 AI-distorted regression / 1.49 re-levered sector

Risk-free regime · WACC2.31% → 7.86%blended 50/50; low bookend 1.13%→6.68%, high 4.30%→9.84%

Base intrinsic NT$538.77 (-71.7%) · low bookend NT$655 (-65.6%) · high bookend NT$376 (-80.3%) · bull NT$1,013 (-46.8%) · windfall-adj ~NT$700 (-63%) · market NT$1,905

The operating rungs are the disciplined extrapolation of Innodisk's own 10-year volume record with the 2026-27 memory-price spike stripped out. Unlike a normal Taiwan name, the value is NOT decided by the rate regime — it is expensive at all three (low NT$655, base NT$539, high NT$376). The headline applies the defensible blended 2.31% TWD risk-free; the bookends apply the pure 1.13% local-bond floor and a normalized 4.30% rate.

The story & the five claims

The 10-year story this DCF is built on, plus the five anchor claims that translate the story into model inputs.

The 10-year story

Innodisk is a genuinely excellent business: a net-cash, ~21%-mean-ROE industrial-memory niche monopolist, #1 in industrial SSD for seven straight years, paying a dividend for thirteen. The valuation question is not quality — it is price. The stock has run ~9x off its low to NT$1,905 on the 2026 DRAM/NAND super-cycle, which pushed Q1'26 net margin to 42% and TTM operating margin to ~52% against a 13-17% structural band. The decisive modeling fact, from ten years of filings, is that annual gross margin is mix-driven, not spot-driven — it held a tight 30-34% band the whole decade, and the 2023 trough margin was actually higher than the 2025 boom margin. So in a memory cycle it is revenue (price × volume) that balloons, not annual margin %, and revenue is what reverts. We therefore base off FY2025 (revenue NT$14.26bn, operating margin 16.8% — already in-band), grow revenue 11% a year for five years (the firm's demonstrated peak-to-peak volume CAGR, well below the NT$68-82bn analyst price-chasing peak), hold a normalized 17% operating margin, and fade terminal ROC to 13%. On those sober inputs the DCF prints NT$538.77/sh (pre-governance NT$585.62) — about 72% below the market — and the Monte Carlo puts a 100% probability that intrinsic sits under today's price. Crucially, this is not a rate-regime artifact: the name is expensive at the suppressed 1.13% local-bond floor (NT$655), at the defensible 2.31% blend (NT$539), and at a normalized 4.30% rate (NT$376) alike. We then add back the discounted 2026-27 super-cycle windfall the model deliberately omits (~+NT$160/sh → ~NT$700, still -63%), and even grant the aggressive permanent-margin bull (22% margin, β 0.95 → NT$1,013, still -47%). Every defensible path lands far below the price. The market is capitalizing the 2026 peak as permanent; history says it will not hold. SELL — a GOOGL-style counter-example: excellent company, wrong price.

Two debates worth pressure-testing

Isn't basing off FY2025 and ignoring the 2026-27 super-cycle profit a hand-wave?

It was — the council flagged exactly that. So the windfall is added back explicitly, not absorbed into book equity. Gross excess after-tax operating profit above the normalized path is ~NT$24bn (2026 ~NT$18bn, 2027 ~NT$6bn); after a ~30% haircut for working-capital cash-conversion timing, capacity reinvestment and 2027 uncertainty, ~1yr discount and the 8% governance haircut, that is ~NT$15-16bn ≈ ~NT$160/sh. Windfall-adjusted intrinsic ≈ NT$700, still 63% below price. The SELL survives the add-back decisively — the 4-of-5 council prediction.

What would it take for NT$1,905 to be right?

Reverse-solving: market cap ~NT$183bn. Even crediting the full NT$24bn windfall, the operating business must be worth ~NT$159bn — ~2.8x the base-case EV (NT$57bn) and ~1.5x the bull-case EV (NT$106bn). You need simultaneously a permanent 22%+ margin AND ~NT$50bn+ year-10 revenue AND a sub-1.0 β AND the windfall — i.e. you must capitalize the 2026 peak (NT$68-82bn revenue at 40%+ margins) close to permanently. That is "peak-forever", outside any defensible normalized envelope.

CLAIM 01Revenue

Revenue compounds ~11%/yr for five years off the FY2025 NT$14.26bn base — essentially the demonstrated 2017→2025 peak-to-peak VOLUME CAGR — reaching ~NT$31.5bn (~2.2x) by year 10. Well above the NT$14bn base, deliberately well below the NT$68-82bn analyst price-chasing peak.

Y1-5 CAGR +11.0% → ~NT$31.5bn by Y10

CLAIM 02Margin

Operating margin holds a normalized 17% — the top of the durable 13-17% band, a modest mix-driven step-up for the Edge-AI/CXL/data-center shift. We refuse the 42-52% spike outright: it is a transient inventory-holding gain in the quarterly P&L that reverts.

Y10 operating margin 17.0%

CLAIM 03Capital

Sales-to-capital held down to 2.0 in Y1-5 for the Yilan Plant 2/3 build and a peak-price working-capital balloon (inventory + AR ballooning at spiking memory prices), recovering to 2.8 in Y6-10 as plants fill and the WC unwinds.

S/C 2.0 (Y1-5) → 2.8 (Y6-10)

CLAIM 04Risk

Unusually, the rate regime does NOT decide this call — the name is expensive at all three. Defensible blended 2.31% TWD risk-free → 7.86% WACC → NT$538.77. Low bookend (1.13%) NT$655; high bookend (4.30%) NT$376. β pinned at 1.15, above the AI-distorted 0.85 regression, a discount to the 1.49 re-levered sector.

WACC 7.86% base · 6.68% low · 9.84% high

CLAIM 05Terminal

Terminal growth pinned at TWD inflation 2.0% (≤ the 2.31% risk-free ceiling). Terminal ROC faded to 13% versus the model's mechanical ~30%, refusing an absurd ~22pt perpetual excess return for a value-add house riding a cyclical commodity upstream. No failure risk: net-cash, 13-year dividend record.

Terminal g 2.0% · ROC 13% · TV ~71% of EV

Where we diverge from sell-side

We normalize in the flows, not with a bolt-on discount. Revenue (not annual margin %) is what balloons in a memory cycle, so we model the durable revenue LEVEL off FY2025 and strip the 2026-27 price spike — exactly the opposite of capitalizing the peak.
We refuse the analyst peak as an anchor. 2026E ~NT$68bn / 2027E ~NT$82bn is 2-3 analysts chasing spot prices, revised up violently. Even the sell-side models 2027E -34% with a 2027-low NT$38 EPS — full normalization.
We add the windfall back explicitly. Rather than hand-wave it as "already banked", we discount the 2026-27 excess (~NT$24bn gross → ~NT$160/sh) as a discrete line — and the SELL still survives at ~NT$700.
We fade the terminal ROC. 13% rather than the model's mechanical ~30%, refusing a ~22pt indefinite perpetual excess return for a commodity-exposed module house.
We pin β above the regression. 1.15 (not the 0.85 AI-parabola-distorted regression) for a hyper-cyclical memory-exposed name — the conservative-against-the-buy choice.

Two-sided case — bear anchors

Expensive at every rate regime. Low bookend NT$655 (-66%), base NT$539 (-72%), high bookend NT$376 (-80%). This is a robust SELL, not a rate-regime artifact — the inverse of a robust buy.
Monte Carlo is unanimous. The base band runs NT$400 (p5) to NT$651 (p95), median ~NT$516; across 1,000 correlated stress draws not one lands above the market. P(intrinsic < market) = 100%.
The bull doesn't reach the price. Even granting permanent 22% margin and a sub-1.0 β, the expansionist bull prints NT$1,013 — still 47% below market; +windfall ≈ NT$1,190, still 38% below.
Peak-earnings mean reversion is near-certain. Memory is the most cyclical business in tech; management itself admits "overbooking… definitely exists." The 42-52% margins are a spot-price artifact that reverts on timing, not on whether.
Earnings quality is inflated during the ramp. 26-64% accrual ratio, a dividend paid despite negative FCF, a working-capital build at peak prices — a price reversal turns that into inventory write-downs and AR risk.

Risks to thesis (tail, not bear case)

Peak-earnings reversion

The 42-52% Q1'26 margins are a memory-spot-price spike against a 13-17% structural band. Reversion is near-certain; only the timing is open. The price capitalizes the peak as permanent.

Overvalued at every regime

Low rf NT$655 (-66%), base NT$539 (-72%), high rf NT$376 (-80%). Even the windfall-adjusted ~NT$700 (-63%) and the bull NT$1,013 (-47%) sit far below price. P(intrinsic < market) = 100%.

The bull pivot (what flips it)

The entire BUY case is one claim: does the AI/edge mix-shift permanently ratchet through-cycle margin toward the low-20s AND lift durable revenue toward NT$50bn+? The 10-year filings argue against it — stable mix-driven margins on a cyclical upstream.

Earnings quality / WC balloon

26-64% accrual ratio, dividend despite no FCF during the ramp, inventory + AR ballooning at peak prices. A price reversal converts the windfall into write-down and receivable risk.

Governance · dilution

One-man Chairman=President=GM, minority-independent board (4/9), consolidated Aetina/Antzer lattice, and an incoming NT$3.0bn 0%-coupon CB. A second-order -8% haircut — the first-order risk is the price, not the structure.

β classification swing

Computers/Peripherals (β_u 1.47) vs Semiconductor (β_u 1.87) is a load-bearing call; the 0.85 regression is distorted low by the AI parabola. We pin 1.15 and sample 0.95/1.40 — none of which rescues the price.

10-year forecast

Bars are revenue and FCFF; the line is operating margin. The explicit window is an 11% Y1-5 volume recovery off the FY2025 NT$14.26bn normalized base — deliberately stripping the 2026-27 memory-price spike — with operating margin held to a normalized 17%, not the transient 42-52% peak.

Monte Carlo distribution

Monte Carlo (base, 2.31% blended rf, 4 axes incl. a widened 0.13 through-cycle margin low-tail) spans NT$400 (p5) to NT$651 (p95), median ~NT$516 — and across all 1,000 correlated stress draws not one lands above the NT$1,905 price. P(intrinsic < market) = 100%: the simulation, not just the point estimate, reads the stock as priced for a permanent super-cycle.

Mean NT$520.58 ± NT$77.02/sh, 1000 iterations (0 failed). P(intrinsic < market NT$1905.00) = 100.0%.

Cost of capital build

Risk-free rate	2.31%
Mature-market ERP	4.23%
Levered β	1.15
Weighted CRP	0.69%
Cost of equity	7.97%
Pre-tax cost of debt (synth Aaa/AAA)	2.71%
D / V	~2%
WACC	7.86%

Full year-by-year DCF

Year	Revenue	Op mgn	EBIT	EBIT(1−t)	Reinvest	FCFF	PV
1	NT$15.83B	16.80%	NT$2.66B	NT$2.19B	NT$784M	NT$1.40B	NT$1.30B
2	NT$17.57B	16.85%	NT$2.96B	NT$2.44B	NT$871M	NT$1.57B	NT$1.35B
3	NT$19.50B	16.90%	NT$3.30B	NT$2.71B	NT$966M	NT$1.75B	NT$1.39B
4	NT$21.65B	16.95%	NT$3.67B	NT$3.02B	NT$1.07B	NT$1.95B	NT$1.44B
5	NT$24.03B	17.00%	NT$4.09B	NT$3.36B	NT$1.19B	NT$2.17B	NT$1.49B
6	NT$26.24B	17.00%	NT$4.46B	NT$3.65B	NT$790M	NT$2.86B	NT$1.82B
7	NT$28.18B	17.00%	NT$4.79B	NT$3.90B	NT$694M	NT$3.21B	NT$1.90B
8	NT$29.76B	17.00%	NT$5.06B	NT$4.09B	NT$564M	NT$3.53B	NT$1.96B
9	NT$30.89B	17.00%	NT$5.25B	NT$4.23B	NT$404M	NT$3.82B	NT$1.98B
10	NT$31.51B	17.00%	NT$5.36B	NT$4.29B	NT$221M	NT$4.06B	NT$1.98B

Methodology & flags

Damodaran FCFF DCF, 10y explicit + perpetuity, in TWD. Risk-free 2.31% (local-currency government bond). Synthetic credit Aaa/AAA. CRP 0.69% from revenue-weighted country mix × Damodaran 2026 CRPs. Terminal ROIC faded to 13.00%; 8% governance haircut applied post-DCF. Monte Carlo: 1000 iterations. Engine v1.0.0 · result: /Users/valentin/Documents/notes_jiliac_labs/Finance/Damodaran/valuations/innodisk/output/2026-06-02-result.json

R&D cap: OFF · Lease cap: OFF · Failure: OFF · ESO: OFF
Governance haircut: 8% applied post-DCF (NT$585.62 > NT$538.77)
Sensitivity tornado: not run

FCFF DCF · dark theme v3. Generated 2026-06-02T10:49:17 from /Users/valentin/Documents/notes_jiliac_labs/Finance/Damodaran/valuations/innodisk/output/2026-06-02-result.json.