Full Report

What Goldkey Is

Bottom line. Goldkey Technology (TWSE 3135, brand Neo Forza) is a small Taiwanese memory-module house that buys DRAM and NAND flash from the chip makers, assembles and sells modules, and adds an ODM arm. It listed in August 2025 at NT$28 and now trades near NT$167 after touching NT$260. FY2025 was a record — revenue NT$7.7bn, net income NT$446m, a 10% gross margin — but the seven-year record shows those margins swing with the memory price cycle, and the record year consumed cash rather than produced it.

A memory-module house, not a chip designer

Goldkey describes itself as an IC-design memory company, but the economics are those of a module assembler and distributor. In FY2025, DRAM products were 80.9% of sales and flash products 19.0% [1]. The company sells JEDEC DRAM modules and flash SSDs under its own Neo Forza brand and through system-integration and ODM services, buying the underlying chips from the upstream memory manufacturers. The clearest signal of the business model is research spend: FY2025 R&D was NT$29.7m, or 0.39% of revenue [2]. A genuine fabless designer spends multiples of that; a firm at 0.39% is buying its silicon and adding assembly, qualification, and distribution.

That places Goldkey near the bottom of Taiwan's crowded memory-module field. By the company's own FY2025 comparison of local listed peers, it held roughly a 6% share of the group's combined NT$128bn of revenue — ahead of only Silicon Power, and dwarfed by ADATA at 41% [3].

Source: FY2025 Annual Report, peer capital and revenue comparison (Goldkey ~6% of the group; ADATA ~41%) [4].

Seven years of the memory cycle

The FY2025 headline is a genuine record. Revenue rose 39.9% to NT$7,704m, gross profit more than tripled to NT$774m, net income rose to NT$446m, and basic EPS reached NT$6.33 against NT$2.37 a year earlier [5].

Source: FY2025 Annual Report, letter to shareholders [6].

The record reads differently against the longer record. Goldkey's revenue has moved sideways-to-up for seven years, but its gross margin has oscillated between roughly 2% and 10% — and net income with it. In the last DRAM upswing, FY2021, gross margin reached 9.1% and net income NT$212m; the very next year, FY2022, margin fell to 2.0% and net income collapsed to NT$18m [7]. FY2025's 10.0% margin is the highest of the seven years, and it sits at the top of that same range.

Source: FY2019–FY2023 from FY2023 Annual Report five-year summary [8]; FY2024–FY2025 from FY2025 Annual Report [9].

Source: FY2023 Annual Report five-year summary [10]; FY2025 Annual Report [11].

Sources: FY2023 Annual Report five-year summary (FY2019–FY2023) [12]; FY2025 Annual Report (FY2024–FY2025) [13]. EPS spans a rising share count and is less comparable across years than the margin line.

Management attributes the FY2025 step-up to broad-based demand and to an AI industrial-control division stood up in FY2024, and the product roadmap does tilt toward AI-server, edge, and AI-PC memory. Those are real. The open point is how much of the margin gain is that mix shift versus the DRAM price recovery lifting every module maker at once — the same force that lifted, then dropped, the FY2021 peak.

The listing and the re-rating

Goldkey, founded in 1998, listed on the Taiwan Stock Exchange in August 2025, issuing 9.12m new shares at NT$28 to raise about NT$255m [14]. The market's response has been the dominant fact of the past year: the shares climbed from the low-NT$30s to a peak of NT$260 in May 2026 before easing to NT$167 by July 2026 — roughly a five-fold gain from the offer price, and still about 36% below the peak.

Source: exchange price history, as reported (monthly closes, July 2025–July 2026).

With 77.5m shares outstanding [15], NT$167 implies a market value near NT$12.9bn and a trailing P/E of about 26 times FY2025 earnings. That multiple is applied to what the seven-year record marks as a cyclical high in profitability — the tension this report is built to examine.

Earnings that have not been cash

The FY2025 record has a second feature that matters as much as its cyclicality: it was not cash. Operating activities used NT$1,774m of cash in FY2025, against NT$446m of reported net income; the gap was funded by NT$1,839m of financing inflows — the IPO proceeds and net borrowing — and period-end cash still stood at only NT$61m [16].

This is the module maker's structural trait, not a one-off. When revenue and memory prices rise, inventory and receivables absorb cash faster than profit creates it: FY2023, another ramp year, also burned NT$909m of operating cash [17], while FY2022, when revenue fell, released NT$636m [18]. Across FY2023–FY2025 combined, Goldkey reported about NT$665m of net income but consumed roughly NT$2.5bn of operating cash.

Sources: FY2025 Annual Report (FY2025, FY2024) [19]; FY2024 Annual Report (FY2023) [20].

The balance sheet carries the strain. Total assets roughly doubled to NT$5,179m, financed to NT$3,055m of liabilities against NT$2,123m of equity — a debt-to-equity ratio near 1.4 [21]. Even so, the board proposed a cash dividend of NT$4.1 per share on FY2025 earnings — NT$316m returned while the operation was consuming cash [22]. One structural mitigant runs the other way: Goldkey both sells and buys in US dollars, giving it a natural currency hedge and limiting FX translation risk in its NT$ accounts [23].

The through-line

The report is built around one question, and every chapter that follows connects to it: how much of Goldkey's record FY2025 profitability is durable earning power — a genuine shift toward higher-value AI and industrial-control memory — versus a cyclical windfall from the DRAM upswing that a downturn would take back, given that the shares have re-rated roughly five-fold since listing and that the record year produced no operating cash. The seven-year margin record and the FY2025 cash statement are the two facts that make the question live; resolving it is the work of the chapters ahead.

---

Cash Conversion

Bottom line. Goldkey's record FY2025 profit did not become cash. Net income of NT$446m sat against an operating cash outflow of NT$1,774m, a gap funded by borrowing and IPO proceeds. Almost the entire outflow is a working-capital build: inventory more than tripled to NT$2,619m — half the balance sheet — while receivables rose NT$600m. The earnings are real accounting profit, but low quality as cash, and the pattern recurs across the memory cycle rather than being a one-off.

Record profit, negative cash

FY2025 reported net income of NT$446m on revenue of NT$7,704m [1]. The consolidated cash-flow statement tells a different story about the same year: operating activities used NT$1,774m of cash [2]. The NT$2.2bn distance between the two is the subject of this chapter.

Source: FY2025 Annual Report, financial performance and cash-flow analysis [3] [4].

The audited cash-flow reconciliation shows where the money went. Starting from pre-tax profit of NT$561m, the dominant adjustments are not accounting add-backs but movements in working capital: inventory absorbed NT$1,879m of cash and receivables another NT$600m, offset only partly by NT$153m more owed to suppliers. Cash generated from operations was NT$1,703m negative before interest and tax [5].

Source: FY2025 audited cash-flow statement; non-cash items and payables/other grouped for clarity [6].

Management is candid about the cause. The annual report attributes the swing directly to "advance stocking to build inventory in response to anticipated future market demand" [7]. This is not an earnings-recognition problem — the profit is audited and the receivables are collectible (below). It is a capital-intensity problem: the way this business grows is by tying up cash in memory chips.

Nearly two hundred days of working capital

The clearest way to see the strain is the cash conversion cycle — how long a dollar is locked in inventory and receivables before it comes back as cash, net of what suppliers finance. In one year it nearly doubled, from about 106 days to about 199 days, driven overwhelmingly by inventory days rising from 50 to 138.

Source: derived from FY2025 audited balance sheet and income statement — inventory NT$2,619m, receivables NT$1,499m, payables NT$187m against COGS NT$6,930m and revenue NT$7,704m [8].

Two features stand out. Payables are strikingly short — under ten days — because the upstream chip makers extend little credit; Goldkey pays for silicon well before it collects from customers, so almost none of the inventory is supplier-financed. And the receivables, though large, are clean: at year-end effectively all of the NT$1,500m book was current, with only NT$15m overdue and under thirty days, against a loss allowance of less than NT$1m [9]. The cash is not stuck because customers won't pay; it is stuck because the model carries months of chip inventory and settles with suppliers first.

The inventory question

Inventory is now the defining line on the balance sheet: NT$2,619m, or 50% of total assets, up from NT$714m (28%) a year earlier [10]. It is carried at the lower of weighted-average cost and net realisable value, with a valuation allowance for obsolescence and price declines that closed the year at NT$117m, up from NT$92m [11].

That allowance is where cyclicality meets the income statement. When memory prices fall, Goldkey writes inventory down and the charge lands in cost of sales; when prices recover, the earlier write-down is reversed and credited back to cost of sales, lifting reported gross margin. FY2025 gross profit of NT$774m includes a NT$25.5m net reversal of prior write-downs [12]. Excluding it, the gross margin is about 9.7% rather than the reported 10.0% — the reversal is real but modest, roughly a third of a percentage point.

The longer record shows the mechanism working in both directions. In the FY2022 downturn the company took a NT$76m inventory write-down that helped push that year's gross margin to 2.0% [13]; across the FY2023–FY2025 recovery it released reserves back — NT$62m, NT$15m, then NT$26m — each release a tailwind to margin [14].

Source: cost-of-sales notes, FY2020–FY2025 audited statements; a positive bar reduced gross margin, a negative bar lifted it [15] [16].

The disclosed reversal is small; the undisclosed exposure is not. Under weighted-average costing, a rising memory price also flatters margin as chips bought earlier and cheaper are sold at today's higher prices — a holding gain embedded in the 10% margin but not separated out. The same arithmetic runs in reverse if prices roll over. The FY2022 write-down of NT$76m was taken on an inventory base of well under NT$1bn; the NT$2,619m now on the books was accumulated during a price ramp, so a downturn comparable to FY2022 would apply to a base several times larger — the recovery-year margin and the recovery-year inventory bet are the same wager.

A cash consumer across the cycle

Over the full seven years the relationship is close to inverse: operating cash is strongest when profit is weakest, and weakest when profit is strongest. FY2022 earned only NT$18m but generated NT$636m of cash as a falling market unwound working capital; FY2025 earned NT$446m and consumed NT$1,774m as a rising market rebuilt it.

Source: reported net income and operating cash flow, FY2019–FY2025 filings [17] [18].

Summed across the cycle, the two diverge sharply: about NT$1,131m of cumulative reported net income against roughly NT$1,895m of cumulative operating cash outflow over FY2019–FY2025. With capital spending negligible — the manufacturing is outsourced, so FY2025 capex was only NT$27m [19] — this is neither a plant-investment story nor a bad-debt story. It is purely working capital: a low-margin distributor that reinvests each up-cycle's profit, and then some, into more inventory.

There is a genuine two-sided reading here. The bull case is that the cash burn is the cost of growth — revenue rose 84% from NT$4.2bn in FY2023 to NT$7.7bn in FY2025, and a growing distributor must permanently fund a larger book; FY2022 shows the model does release cash when growth stops. The bear case is that the reinvestment happens at the worst possible time — buying the most inventory when prices are highest — so reported profit is chronically trapped in stock bought near the peak. What would move the read is a full cycle in which cumulative operating cash turns positive, or evidence that working-capital intensity per dollar of sales is structurally falling; so far it has risen.

What management projects for FY2026

The most direct evidence on durability is the company's own forward liquidity plan. The annual report projects that FY2026 operating activities will again be cash-negative, at about NT$645m outflow, and that after investing and financing needs the year opens with a projected NT$1,177m shortfall against a starting cash balance of only NT$61m — to be met by issuing a second domestic convertible bond [20].

Source: FY2025 Annual Report, next-year cash-liquidity analysis; remedy stated as a second unsecured convertible bond issue [21].

The strain already shows in three places. Year-end cash of NT$61m stands against NT$1,221m of short-term borrowings, so the company is running on rolled-over bank lines [22]. It has begun factoring receivables with recourse — US$4.5m of trade receivables pledged for NT$141m of bank funding in FY2025, retained on the books because the credit risk is not transferred [23]. And the rising debt is now visible in the income statement: finance costs climbed to NT$43m in FY2025 from NT$28m in FY2024 [24], and will carry a full year of the enlarged borrowings and the convertible bond into FY2026.

None of this makes the FY2025 profit fictitious. The measured read is narrower: Goldkey's earnings are real but structurally cash-absorbing, and the FY2025 record was accompanied by the largest inventory bet in its history, funded by debt and equity rather than by the business. Whether that bet reads as durable earning power or a cyclical windfall is most sensitive to memory prices, which will decide both the value of the inventory and the margin on what it sells.

---

Competitive Standing

Bottom line. Goldkey is the smallest of the seven listed Taiwan memory-module makers, at roughly 6% of the domestic module market against ADATA's 41%. It sells a differentiation story — a shift toward high-margin industrial, AI-edge and automotive memory that is "decoupled from the price cycle" — but the numbers do not yet show it. In the same FY2025 up-cycle, genuine industrial-memory specialists earned 28%–47% gross margins; Goldkey earned 10.0%. On R&D of 0.39% of sales and 77 employees, the edge is a direction, not yet a moat.

The smallest in the room

Goldkey names its own peer group in its filings: the listed Taiwan makers of memory and flash modules — Innodisk, ADATA, Apacer, Transcend, Team Group and Silicon Power. On the company's own scale comparison, Goldkey sits second-from-last, with NT$7,704m of net revenue and an estimated 6% share of the domestic module market [1]. ADATA, the leader, turns over NT$53,087m — roughly seven times Goldkey — and holds 41% [2]. The company concedes the point in the same disclosure: peers operate at "different production scales and different areas of expertise," and no clean sales ranking exists [3].

Source: FY2025 Annual Report, market-share peer comparison table [4].

Sub-scale in a commodity assembly business is a disadvantage, not a neutral fact. The traditional module model — buy DRAM and NAND chips, mount them on a board, sell the standard part — competes on procurement cost and scale, and Goldkey buys fewer chips than any peer but Silicon Power. The company describes this model plainly and unfavourably: "low margin, highly dependent on the pricing cycle," with a moat that rests on "economies of scale and procurement-cost control" [5]. On that model's own terms, Goldkey is the wrong end of the scale ladder.

The margin gap

A memory moat shows up most clearly in what a company earns in an up-cycle. FY2025 was a strong one for the whole group as DRAM prices ramped, so it strips out the excuse that "the cycle was bad." In that shared tailwind, the gap is stark.

Sources: Goldkey FY2025 audited statements, gross profit NT$773,691k on revenue NT$7,704,142k [6]; ADATA FY2025 gross margin 27.81% [7]; Team Group Q1 2026 gross profit 3,438,206 on revenue 9,044,806 [8]; Transcend FY2025 gross margin 46.8% [9]. ADATA and Transcend are full-year FY2025; Team Group is the latest reported quarter.

Goldkey's 10.0% gross margin — its own cyclical high — is under half of ADATA's 27.8% and under a quarter of Transcend's 46.8% [10][11]. The comparison is not cherry-picked at the top: Transcend's own investor deck puts the industry gross-margin range at 20.7%–31.1% for 2025, and Goldkey sits below the bottom of that range [12].

What separates the high-margin peers is precisely the thing Goldkey says it is building. Transcend attributes its margins to a "successful transformation into industrial-control and embedded applications," lifting gross margin through high-value-added products and a brand that competes "beyond price" [13]. That is the shape of a real industrial-memory moat: it shows up as a 40%-plus gross margin that holds. Goldkey's does not, at least not yet.

The differentiation it sells

Goldkey's own account of its edge is explicit and, in places, self-aware. It contrasts the "traditional module model" with what it calls its core value: a "technology-intensive service business" delivering "stable technology and system solutions," whose moat is "software-hardware integration and path-engineering capability," and whose aim is to "raise the share of non-consumer revenue, resist the cycle, and stabilise cash flow" [14]. Concretely, it points to a decade of moving up the application stack — its own Neo Forza overclocking brand from 2017, ECC and wide-temperature (-55 to 105°C) industrial-grade parts and server-memory certification from 2020–2022, and a stated push into AI-server, edge, automotive, robotics and medical memory [15]. The goal it names — lifting industrial-control, AI-edge, automotive, robotics and medical revenue so that pricing "decouples from the cyclical cycle" — is exactly the right ambition for this business [16].

Three facts sit against the claim that the ambition has become a moat.

Sources: FY2025 Annual Report — R&D NT$29,702k, 0.39% of sales [17]; 77 employees [18]; share and margin per [19] and [20].

First, the research budget. A "technology-intensive" firm building proprietary integration would show it in R&D; Goldkey spent NT$29.7m — 0.39% of sales — in FY2025, and 0.35% the year before [21]. Second, the headcount: 77 employees, none with a doctorate and around one in ten holding a master's degree [22]. That is a distribution-and-assembly organisation, not an engineering one. Third, and most telling, the revenue mix moved the wrong way in the record year: DRAM rose from 68.4% of sales in 2024 to 80.9% in 2025, as the company leaned into the commodity chips whose prices were rising, not away from them [23]. The FY2025 profit was made on commodity DRAM riding a price ramp, which is the opposite of counter-cyclical.

Reinforcing those three facts, Goldkey's own gross margin still tracks the memory cycle almost perfectly — the behaviour a genuine non-consumer mix is supposed to dampen.

Sources: FY2022–FY2024 per company investor deck [24]; FY2025 audited statements [25].

Margin fell to 1.7% at the 2022 trough and recovered to 10.0% at the 2025 peak — a near-sixfold swing over four years, on top of the ~9% peak reached in the prior 2021 up-cycle covered in What Goldkey Is [26]. A business whose earnings had genuinely decoupled from memory pricing would not print a 1.7% margin one year and 10.0% three years later. Transcend's did decouple, and its margin held in the 40s; Goldkey's has not.

What would settle it

The honest read is that Goldkey has no established moat today — the scale is sub-critical, the margin is the lowest in its own peer group, and the returns that a real advantage produces are absent. The direction, though, is not imaginary: the industrial and AI-edge certifications are real, the Neo Forza brand exists, and Transcend is living proof that a sub-scale Taiwan module maker can transform into a 40%-margin industrial-memory specialist over several years. Goldkey may be early on the same path.

Two things would change the read, and both are checkable in future filings. The first is disclosure and growth of a non-consumer revenue share — the company markets the ambition but has not yet put a number on how much of sales is industrial, automotive or AI-edge, so the mix cannot be verified. The second is margin behaviour in the next downturn: if gross margin holds well above the low-single-digits when DRAM prices fall — the way Transcend's now does — the differentiation will have become real. Until then, the FY2025 record is best read as a sub-scale price-taker's good turn in a strong cycle, not the earnings power of a defended niche.

---

A Shared Windfall

Bottom line. Goldkey's Q1 FY2026 gross margin of 30.4% is triple any full year it has ever reported, and the same quarter shows every Taiwan DRAM-module peer posting a record too — ADATA at 55.7%, Team Group at 38.0%, Transcend above 60% — as DRAM contract prices roughly quintupled. That margin is a group-wide inventory holding gain, not a Goldkey-specific one: Goldkey captured the least of it, and the peers that captured the most carry the industrial franchise it has yet to build.

The quarter every module maker printed a record

In the three months to March 2026 Goldkey booked NT$3,958m of revenue at a 30.4% gross margin — up from a 5.0% margin in the same quarter a year earlier — and NT$856m of net income, more than its entire FY2025 profit [1]. Read alone, that quarter looks like a step-change in earning power. Read alongside the peers who filed the same quarter, it looks like the tide.

Every Taiwan memory-module maker that reported Q1 2026 posted a record gross margin at the same moment. ADATA reached 55.7%, Team Group 38.0%, and Transcend — in its most recent disclosed quarter, Q4 2025 — 61.6%.

Sources: Goldkey Q1 FY2026 report [2]; Team Group Q1 2026 conference [3]; ADATA Q1 2026 conference [4]; Transcend 2Q 2026 conference [5].

The lift was not a single-quarter surprise either. ADATA's margin climbed every quarter through the up-cycle — 13.8%, 19.0%, 22.7%, 48.4%, then 55.7% — and Team Group's did the same on the quarters it disclosed [6][7]. Goldkey's two data points — 5.0% a year ago, 30.4% now — sit on the same slope, at the bottom of it.

Sources: ADATA Q1 2026 conference [8]; Goldkey Q1 FY2026 report [9]; Team Group Q1 2026 conference [10]. Goldkey line drawn between its two disclosed quarterly margins.

What moved the margin

The common cause is the price of the chips these makers resell. DRAM contract prices, quoted in US dollars industry-wide, roughly quintupled over the year: ADATA's own price sheet shows the DDR4 8GB module going from about US$16.5 in Q1 2025 to US$85 in Q1 2026, and DDR5 from US$23 to US$75 [11].

Source: ADATA Q1 2026 conference, DRAM contract price chart (TrendForce data) [12].

A module maker buys chips from Samsung, SK Hynix or Micron, assembles them, and sells the module on. When the chip price climbs faster than the inventory turns, the modules sold this quarter were built from stock bought at last quarter's lower price — the difference falls straight to gross profit. That is the pass-through, holding-gain mechanism the report identified in Goldkey's own numbers (Cash Conversion). The peer filings show it is not proprietary to Goldkey: it is what happens to any inventory-heavy reseller when the commodity it holds reprices upward. ADATA said as much, attributing a 42-point year-on-year margin jump to the price environment rather than to any change in what it sells [13].

The gap inside the spike

That every peer's margin rose together is the cyclical read. What separates them at the top is the more useful signal — and the order is the same one the down-cycle produced. Goldkey's 30.4% is the lowest of the group even at the peak, exactly as its FY2025 margin was the lowest (Competitive Standing). The windfall lifted everyone; it did not close the gap.

Latest-quarter GM: Goldkey, ADATA and Team Group are Q1 2026; Transcend is Q4 2025 (Q1 2026 gross margin not disclosed). Sources: Goldkey [14][15]; Team Group [16][17][18]; ADATA [19][20]; Transcend [21][22][23].

The right-hand column is where the peers diverge from Goldkey. Transcend, the highest margin of the four, sells 62.8% of its output into industrial and embedded applications [24] — an engineered mix that holds a margin floor through the cycle rather than only at the peak, benchmarked against these peers in Competitive Standing. Team Group is moving the same way, with its business-to-business share reaching 47.6% of sales this quarter, up from 16.8% a year earlier [25]. Goldkey leaves the equivalent line blank: it markets an industrial and AI direction but has never quantified a non-consumer revenue share (What to Watch). The company nearest the top of the margin spike is the one with the least commodity exposure; the company at the bottom is the one that has published no evidence of any.

The symmetry on the way down

A holding gain is symmetric. The stock that inflates margins while prices rise becomes the stock that gets written down when they fall, and the whole group carries a lot of it: inventory is 46% of Goldkey's assets, 38% of ADATA's, and 62% of Team Group's at the most recent quarter [26][27][28]. The last time the cycle turned, in 2022, Goldkey's gross margin fell to about 2% (Competitive Standing). The mechanism that produced 30.4% this quarter is the same one that produced 2% then, run in reverse.

Transcend's lighter inventory — 20% of assets, against 46% to 62% for the others — is the balance-sheet signature of a franchise that sells engineered products to design-in customers rather than reselling commodity chips [29]. It is what durability looks like in this industry, and it is what Goldkey's filings do not yet show. The point is not that Goldkey's peers are safe — they too are levered to the same cycle in the same direction — but that Goldkey rides it with the least franchise and one of the fuller inventory positions in the group.

What would change the read

The peer read-through does not decide the durable-versus-cyclical question; it sharpens what the answer will look like. If Goldkey's margin holds well above its historical high-single digits as DRAM prices fall — while the group's margins compress toward their own through-cycle norms — that would be evidence of a franchise forming, and it would most likely arrive with a disclosed, rising non-consumer revenue share. If instead Goldkey's margin falls with the group, or faster from its lower base, and an inventory write-down replaces this year's reserve releases, the Q1 record was the cyclical spike it most resembles. The next two or three prints, set against the same peers filing the same quarters, will show which.

---

Priced for Peak

Bottom line. At about NT$167 Goldkey trades near 27 times trailing earnings and 7 times book — a richer multiple than Transcend, the proven 48%-margin industrial-memory franchise, which trades at roughly 19 times and 4.7 times. The premium sits on peak-cycle profit: the first quarter of FY2026 alone earned NT$856m at a 30% gross margin, triple the full-cycle norm. On mid-cycle economics the same shares cost 50–70 times, and the count they divide keeps growing through serial convertible-bond conversion.

What the market is paying

Goldkey's re-rating turned a NT$28 offer into a stock that touched NT$260 and settled near NT$167 by July 2026 — a valuation that only makes sense against the earnings you assume are durable. On the headline numbers, the market is not treating this as a cheap cyclical.

Source: market data as of 1 July 2026 (exchange quotes and reported statistics); Goldkey fundamentals cross-checked to audited filings [1].

The instructive comparison is against the peers named in Competitive Standing. Transcend — the sub-scale Taiwan module maker that genuinely transformed into a 40%-plus-margin industrial-memory specialist, earning a 48% net margin over the trailing year — trades at about 19 times earnings and 4.7 times book. Team Group, whose DRAM-and-flash product mix is almost identical to Goldkey's, trades near 20 times. Goldkey, the smallest of the group and the lowest-margin, carries the highest earnings multiple and by far the highest multiple of book value.

Source: market data as of 1 July 2026 (exchange quotes; trailing multiples and margins as reported). Transcend and Team Group net margins are their own reported trailing figures; Goldkey's 12.4% trailing margin is itself a cycle-peak reading.

A commodity assembler trading above a defended franchise is not, on its own, proof of mispricing: the multiple capitalizes whichever earnings prove durable, and the rest of this chapter tests which those are.

The earnings under the multiple

The 27-times multiple divides a share price by profit that has just gone vertical. Goldkey's gross margin ran between 1.7% and 10.0% across the seven-year cycle traced in What Goldkey Is. In the three months to March 2026 it reached 30.4% — roughly triple the FY2025 full-year figure and eighteen times the FY2022 trough. Quarterly revenue of NT$3,958m was more than half of the entire prior fiscal year, and single-quarter net income of NT$856m already exceeded all of FY2025's NT$446m [2].

Sources: FY2022–FY2024 per company investor deck and five-year summary [3]; FY2025 audited statements [4]; Q1 FY2026 audited statements [5].

That 30% margin is an inventory holding gain, not a franchise: at a firm spending 0.39% of sales on R&D, it reflects DRAM bought cheaply before the ramp and sold into a spiking market, and it reverses when memory prices stop rising. A Shared Windfall decomposes that group-wide mechanism in full.

What the price implies at each point of the cycle

For a business this cyclical, a single P/E carries little information: the figure swings widely with the quarter chosen as the denominator. The table below runs Goldkey's profit through the range its own seven-year record supports, on the roughly 90m shares now outstanding.

Source: derived from reported financials — FY2019–FY2025 net margins applied to the FY2025 revenue base and divided by ~90m shares; Q1 FY2026 net income annualized [6][7]. Illustrative; the revenue base is itself cyclical.

The valuation is two-sided. Annualized, the March quarter puts the stock at about 4x forward earnings, the bull's case, and not a trivial one while the cycle runs. On the mid-cycle margin the record averages, the same shares cost 60-plus times; on a normal down-year, close to 100 times. A collapsing forward multiple at a cyclical top is what a value trap looks like: the denominator is peaking, not the price. What settles it is whether 30% gross margins are the new normal or a memory-price artifact, the durability question carried through Competitive Standing.

Where the cycle stands

The near-term evidence says the peak is not over. Industry price trackers reported DRAM contract prices still rising through mid-2026 on tight supply and low inventories, with the PC-DRAM uptrend projected to extend into the third and fourth quarters — and DDR5 profitability expected to surpass even high-bandwidth memory. For Goldkey that means the holding-gain engine likely runs a few quarters longer, and FY2026 profit could exceed FY2025 by a wide margin. It also means the setup that eventually reverses is getting larger: the higher prices climb, the bigger the inventory position exposed when they turn, and the further a 30% margin has to fall. The reversal is not imminent on the mid-2026 data; that it will come is what the seven-year record shows unambiguously.

A moving share count

The per-share math has a second complication: Goldkey is funding its structural cash burn — the negative operating cash flow documented in Cash Conversion — by issuing convertible bonds that convert into stock well below the current price. The first, a NT$1bn zero-coupon bond issued in December 2025 at a NT$50.85 conversion price, was already about 88% converted by April 2026, adding 17.4m shares; only NT$117m of face remained [8]. A second NT$1.5bn zero-coupon bond followed in May 2026 at a NT$129.9 conversion price, none yet converted [9].

Both are struck below NT$167, so both dilute rather than raise cash at market value. The effect is already visible in the accounts: Q1 FY2026 diluted EPS was NT$9.81 against basic EPS of NT$10.76 — a 9% gap — and the second bond's full conversion would add roughly another 11m shares [10]. Share count has risen from 77.5m at the end of FY2025 [11] toward roughly 90m, and heads higher — so the earnings the multiple capitalizes are spread across a growing base even as they rise.

What would change the read

The valuation and the fundamentals point the same way: at 7 times book and a higher earnings multiple than a proven 40%-margin peer, the price capitalizes the current DRAM-cycle profit as if it were durable earning power. The strongest fact against that read is the live cycle — with contract prices still climbing, FY2026 earnings should be very large, and a stock at a few times forward earnings can rise further before the turn. What is being paid for, though, is a memory-price spread, not yet a franchise: the 30% margin has no counterpart in Goldkey's own history outside a price spike, and the balance sheet is being funded by dilution, not cash generation.

Two things would change it, both checkable in the filings. If gross margin holds well above the low-single digits when DRAM prices stop rising — the durability test set in Competitive Standing — then peak earnings would be closer to real earning power and the multiple less demanding than it looks. And if the company began converting its cyclical profit into operating cash rather than inventory and new convertibles, the per-share arithmetic would stop working against holders. Until one of those turns, the multiple is a bet on the memory cycle, priced as if it were a bet on the company.

---

Concentration

Bottom line. Goldkey's revenue rests on a narrow base. Two customers supplied 41% of FY2025 sales, and the company buys its raw material from a three-firm memory-chip oligopoly, its two largest suppliers around 42% of purchases. The auditor's sole key audit matter for FY2025 is the genuineness of revenue from a few large, fast-growing customers — not inventory. Concentration eased as Taiwan sales tripled, but the base a 6%-share price-taker depends on stays thin at both ends.

The customer base

Two customers each cleared 10% of Goldkey's FY2025 revenue and together accounted for NT$3,174m of the NT$7,704m top line — 41.2% [1]. Customer A alone was NT$1,826m (23.7%); Customer B NT$1,348m (17.5%) [2]. Neither is named in the filings.

Source: FY2025 Audited Financial Report, Note 28 (major customers) [3] and Note 34 (revenue by customer location) [4].

The direction of travel matters as much as the level. In FY2024, Customer A supplied NT$2,279m — 41.4% of that year's revenue on its own — and the top two together were 60.1% [5]. The FY2025 surge did not come from those two accounts; Customer A's revenue actually fell in absolute terms. It came from elsewhere, so the base widened even as it stayed concentrated.

Source: FY2025 Audited Financial Report, Note 28 — customer revenue as a share of the annual total [6].

Where that widening came from is visible in the geographic split. Revenue by customer location shifted decisively toward Taiwan — from NT$2,416m (43.9%) in FY2024 to NT$4,630m (60.1%) in FY2025 — while US-located revenue fell from NT$2,312m to NT$1,894m even as the total grew [7].

Source: FY2025 Audited Financial Report, Note 34 — revenue disaggregated by customer location [8].

One inference is worth drawing. Customer A's revenue tracks US-located revenue almost exactly in both years — NT$1,826m against NT$1,894m in FY2025, and NT$2,279m against NT$2,312m in FY2024 — which points to the largest single account being the US channel. That leaves the biggest customer relationship exposed to US trade and inventory policy on top of the ordinary risk of a concentrated account. The filings do not confirm the identity, so this is a read from the numbers, not a disclosed fact.

The supplier side

Concentration runs through the supply side as well. Goldkey's raw material — DRAM and flash ICs — comes from what its own prospectus calls an oligopolistic ("寡佔") market supplied by a handful of large wafer makers, principally SK Hynix, Samsung and Micron [9]. Those three set memory prices; Goldkey, a NT$7.7bn module assembler, takes them.

Purchases are as concentrated as sales. In FY2024 the two largest suppliers were 42.1% of total purchases; in FY2023, 48.6% [10]. The company reports no single supplier above 50% of purchases and describes a multi-source strategy [11], and the top supplier does rotate — Supplier A led in FY2023, a different Supplier C in FY2024, with Supplier B second in both years [12].

Source: Listing Application Prospectus 2025 — suppliers exceeding 10% of annual purchases [13].

The picture is a firm with ~6% domestic share (Competitive Standing) sitting between a supply side it cannot influence and a demand side of a few large accounts. Its gross margin is the spread between two concentrated ends it does not control. That is why the FY2025 record reads as a good pass-through year on a rising memory price rather than pricing power. The structure affords Goldkey little of the latter.

The auditor's key audit matter

For FY2025 the auditor, Deloitte, named a single key audit matter: the genuineness ("真實性") of revenue from specific customers. Its stated reason is that certain customers' annual sales were material and rose sharply against the prior year — FY2025 revenue grew NT$2,196m — so it treated the occurrence of that revenue as the year's principal audit risk and sampled those accounts against credit-limit approvals, original orders, shipping documents and cash receipt [14].

A key audit matter is a focus disclosure, not a finding of misstatement; the accounts passed. But the auditor and the concentration data point the same way. When 41% of revenue runs through two unnamed accounts and one of them appears to be the US channel, the quality of that revenue — that it occurred, on arm's-length terms, and converts to cash — is exactly where a skeptic looks first. That it did not convert to cash in FY2025 is the subject of Cash Conversion.

Receivables

The concentration carries into the balance sheet. At year-end the five largest trade-receivable balances were NT$1,180m of the NT$1,500m book — 79% — with the single largest at NT$332m [15]. Credit quality itself looks clean — almost all receivables are current and the loss allowance is negligible — but the exposure is lumpy: a single large account going bad would matter. To fund the working-capital build, Goldkey also factored US$4.5m of receivables with recourse, retaining the credit risk and keeping NT$224m of factored receivables on the balance sheet as collateralized borrowing [16].

The read

Concentration on both sides is real and, for a sub-scale price-taker, structural — it is the shape of the business, not a passing feature. The strongest fact against reading it as an acute risk is direction: the FY2025 growth came from outside the two largest accounts, the top customer's share fell from 41% to 24%, and the top-two supplier share eased from 49% to 42%, so the base is widening as the company grows. The company also markets an AI industrial-control division, set up in FY2024, as the source of newer, more diversified demand [17].

What would change the read is disclosure the corpus does not yet contain: who Customers A and B are, and whether the Q1 FY2026 revenue tripling (Priced for Peak) was broad or ran back through the same one or two accounts. A concentrated surge would sharpen the risk the auditor flagged; a broad one would confirm the base is genuinely widening. Until then, the honest statement is that Goldkey earns a thin, cyclical spread between an oligopoly it buys from and a short list of customers it sells to — and that its auditor named the genuineness of those sales the sole key audit matter for FY2025.

---

Control and Guarantees

Bottom line. Goldkey is a founder-controlled company that runs on its chair's personal credit. Tseng Chen (曾珍) chairs the board, serves as president, and — with a fellow director — personally guarantees NT$1.87bn of Goldkey's bank lines, which is essentially all of them. A corporate director sits on both sides of the company's trade. Board and executive pay is modest at 6.3% of a peak-year profit. The FY2025 cash dividend is funded by a convertible bond, not by cash the business produced.

Who controls Goldkey

Goldkey is run by its founding family through a cluster of investment holding companies. Tseng Chen is chairman and concurrently president (董事長暨總經理), holding 18.52% of the shares through Sheng-Yun Investment [1]. Her brother, Tseng Wan-Chuan (曾萬全), is executive vice-president and, through his own vehicle, the tenth-largest shareholder [2]. Two more directors, Wu Hui-Jung (吳惠容) and Wu Hui-Ling (吳惠玲), are sisters, and each represents or holds a further block [3].

Source: FY2025 Annual Report, major-shareholder list [4].

The named family entities — Sheng-Yun, Rui-Rui, Wu Hui-Ling and Yu-Le — hold roughly a third of the register between them, and the family fills every non-independent board seat. The board has eight directors, four of them independent, so the independent ratio is 50%. But the chair and chief executive are the same person, and the four insider seats trace back to the two founding families [5]. This is ordinary for a small Taiwan issuer; it is worth stating plainly because the decisions that follow — how the company funds itself, what it pays out, whom it trades with — are made by an owner-operator, not by a dispersed shareholder base.

The business runs on the chair's signature

One of Goldkey's larger financial exposures sits off the income statement, in the financing notes. At the end of FY2025 the chair, Tseng Chen — jointly with fellow director Wu Hui-Ling on part of it — had personally guaranteed NT$1,867.8m of the company's bank facilities: NT$524.0m under a joint guarantee and a further NT$1,343.8m in her own name [6]. Against those guarantees the company had actually drawn NT$1,673.3m [7].

That drawn figure is close to Goldkey's entire bank debt. Goldkey's total bank borrowing at year-end — short-term loans of NT$1,221.0m, short-term notes of NT$49.9m, and long-term loans of NT$402.4m — comes to NT$1,673.3m [8]. The two numbers are identical: essentially every dollar of Goldkey's bank debt is personally guaranteed by its chair.

Sources: FY2025 Audited Financial Report, Note 29 (guarantees) [9] and Balance Sheet (borrowings) [10].

The guarantees grew with the borrowing — from NT$1,598m a year earlier to NT$1,868m — as the working-capital build documented in Cash Conversion pulled the company deeper into debt [11]. This cuts two ways. It is genuine alignment: the chair is personally on the hook for a sum that dwarfs the company's NT$446m of net income and its NT$61m of year-end cash, so her incentive to keep Goldkey solvent is real and concentrated. It is also a statement about the business — a company whose banks require the owner's personal signature for every facility is one the market does not yet judge creditworthy on its own balance sheet. For an investor, the guarantee is a support that is only as durable as the chair's willingness and capacity to keep providing it.

A director on both sides of the trade

One of Goldkey's four insider board seats is held by Tait Technology (泰特科技), a corporate director represented by Tsai Ling-Ling (蔡玲玲) and itself a 5.60% shareholder [12]. Tait is also a trading counterparty on both sides. In FY2025 Goldkey sold NT$30.4m of goods to Tait (down from NT$47.4m in FY2024), and through the first nine months of 2025 it bought NT$29.4m of goods from Tait, up sharply from NT$7.4m a year earlier [13][14].

The amounts are small against a NT$7.7bn top line, and both prices are disclosed as market-negotiated [15]. But two-way trade with a company that holds a board seat is the kind of arrangement a professional investor prices in as a standing question rather than a settled fact: it is where a controlled company's interests and its counterparty's can quietly diverge. The volumes are worth tracking precisely because they moved — purchases from Tait quadrupled year-on-year even as sales to it fell.

A financed dividend

For FY2025 the board proposed a cash dividend of NT$316.3m — NT$4.1 per share — plus a stock dividend of NT$0.6 per share [16]. The dividend policy commits the company to distributing at least 30% of each year's earnings, with at least 10% of the distribution in cash [17].

The problem is where the cash comes from. FY2025 operating cash flow was negative NT$1,774m [18]. And in its own FY2026 liquidity plan, management projects opening cash of NT$61m, another year of operating outflow of roughly NT$645m, and a cumulative cash shortfall of NT$1,177m — with the stated remedy being a second unsecured convertible bond, and the financing outflows it must cover explicitly including "repaying bank loans and paying cash dividends" [19].

Sources: FY2025 Audited Financial Report, cash-flow statement [20]; FY2025 Annual Report, dividend proposal [21] and FY2026 cash forecast [22].

Put together, the sequence is a company paying out cash it did not generate, in a year when it consumed NT$1.8bn of it, and funding the gap by selling convertible bonds that dilute existing holders, the same convertible-bond dilution covered in Priced for Peak. The distribution is real, and the family, as the largest holder, receives the largest share of it. But calling it a return of capital overstates the case: on a consolidated view, the dividend is financed.

Executive and director pay, by contrast, is modest. The board and its executive officers were paid NT$28.0m in total for FY2025 — 6.28% of net income — of which director fees were NT$15.1m, or 3.38%, and the balance was executive salary [23]. Key-management short-term benefits roughly doubled, to NT$49.6m from NT$24.6m, as profit surged [24]. The 6.28% ratio flatters the picture because it is measured against a cycle-peak profit; the same absolute pay would be a far larger share of earnings in a trough year like FY2022. The company's only buyback was a small one — 350,000 shares for NT$11.7m in 2023, held as treasury stock for employee transfer — so this is not a business returning capital through repurchases [25].

How to read the stewardship

The capital-allocation signature here is that of an owner-operator managing a thin, cyclical balance sheet: distribute at the top of the cycle, dilute to fund the distribution, and keep the credit lines open with the chair's personal guarantee. That is a coherent way to run a family-controlled cyclical, and it is not the profile of a franchise compounding durable cash for outside holders. The strongest facts against a harsher read are genuine: the personal guarantees are real alignment rather than extraction, independent directors are half the board, and the related-party trade is small and disclosed at market prices. What would change the read in the owner's favour is specific and checkable — the dividend covered by operating cash rather than by a convertible bond, the personal guarantees shrinking as Goldkey earns its own bank credit, and related-party volumes with Tait staying immaterial rather than climbing.

---

What to Watch

Bottom line. This chapter reconciles the seven before it. On the evidence the central tension is real and unresolved: Goldkey's Q1 FY2026 gross margin of 30.4% and one-quarter profit larger than all of FY2025 sit on a DRAM price spike that industry trackers still had climbing into late 2026 — while the same franchise produced no operating cash, earns through two customers, and funded its dividend by issuing shares. What separates a durable inflection from a cyclical windfall is observable, dated, and public.

Where the seven chapters land

Read in sequence, the report establishes what Goldkey is and leaves open whether the record result is durable earning power or a cycle windfall. What Goldkey Is placed it as a sub-scale Taiwan DRAM and flash module house — roughly 6% of the listed module group, R&D at 0.39% of sales — whose gross margin has oscillated between 2% and 10% for seven years. Cash Conversion showed the record FY2025 profit of NT$446m arriving alongside an operating cash outflow of NT$1,774m, the gap driven by a tripling of inventory. Competitive Standing found no margin, share, or return signature of a moat. A Shared Windfall set the same-quarter Q1 margin against peers and found it a group-wide DRAM inventory holding gain in which Goldkey captured the least, earned without the industrial franchise the top-margin names carry. Priced for Peak put the shares at roughly 27x trailing earnings and 7x book — richer than the proven 40%-plus-margin franchise it trades against. Concentration traced 41% of revenue to two unnamed customers and the auditor's sole key audit matter to the genuineness of that revenue. Control and Guarantees showed a founder-family that distributes at the peak, dilutes to pay for it, and personally guarantees every dollar of bank debt.

None of that decides whether the FY2025–Q1 FY2026 result is durable earning power or a cycle windfall a downturn takes back. It sharpens the question and identifies exactly which facts would answer it.

The margin, through the cycle

Source: derived from reported income statements — five-year condensed table [1]; FY2024–FY2025 operating results [2]; Q1 FY2026 income statement [3].

The report's central finding is visible in the chart. Goldkey's gross margin has no trend — only a cycle, swinging from 2.0% in FY2022 to 9.1% in FY2021 and back, with the FY2025 high of 10.0% already the top of that historical band [4]. The Q1 FY2026 print of 30.4% is not a new plateau on that chart; it is three times anything the company has earned in a full year, on revenue that nearly tripled year-over-year to NT$3,958m [5]. A company that has never held a double-digit annual margin printed 30% in a single quarter — the signature of NT$2.6bn of cheaply-bought DRAM sold into a price surge, not of pricing power [6].

The tension, as shared facts

Each row below is a fact both sides accept; they disagree only on what it means, and each disagreement has a specific, checkable resolver.

Sources: Q1 FY2026 income statement [7]; FY2025 cash flow statement [8]; major-customer note [9]; auditor key audit matter [10]; bank guarantee facility [11]. Cycle pricing per industry tracker TrendForce, as of mid-2026.

The rows do not net to a verdict. Rows 1, 3, and 4 favour the cyclical read; the bull answer to each is the same single bet — that Q1 FY2026 is a new base rather than a peak — and the report has found no structural reason (moat, mix, pricing power) for it to be. Rows 2, 5, and 6 are genuinely open on today's evidence: prices were still rising, the base did widen in FY2025, and the chair's guarantee is real capital at risk. That is why the finish is a watch-list, not a call.

Three cycle paths

The signals to watch fall out of three ways the cycle can go from here, each with a different mechanical effect on a pure pass-through model carrying inventory equal to half its balance sheet [12].

Sources: inventory position [13]; the FY2022 low-margin, write-down precedent [14]. Cycle path per TrendForce, as of mid-2026.

The near-term reading favours the first path. As of the report date, TrendForce had conventional DRAM contract prices up 90–95% quarter-over-quarter in Q1 2026 and still rising into the third and fourth quarters — the steepest run on record. For Goldkey that supports strong FY2026 headline prints, and it is the concrete evidence behind the bull's run-rate. It also enlarges the inventory exposed when the move ends: the same mechanism that produced a 30% margin on the way up produced a 2.0% margin and a write-down charge on the way down in FY2022 [15]. A higher, more vertical peak does not lower that risk; it raises it.

What to watch

The items below are ranked by decision value — how much each would move the durable-versus-cyclical read — and every one is falsifiable against a named filing line.

Sources: monthly revenue is a TWSE disclosure requirement; margin and write-down lines per the FY2025 cost-of-sales and inventory notes [16]; customer mix per Note 28 [17]; dilution per shares-outstanding disclosure [18]; guarantee per the financing facility note [19].

The two highest-value items are the fastest. Taiwan requires monthly revenue disclosure by the tenth of the following month, so the first sign of the cycle turning arrives well before any quarterly margin — a single monthly print that falls year-over-year, after the price surge has run, is the earliest falsification of the run-rate bull case. The inventory write-down line is the tell that the down-leg has begun in earnest: through FY2023–FY2025 that line was a reversal adding to margin; its return as a charge would confirm prices have fallen through the carrying value of the NT$2.6bn position [20].

The number that isn't disclosed

A specific disclosure would speak to the durable half of the thesis more directly than any near-term price movement. Goldkey established an "AI Industrial-Control Business Division" in FY2024 and says it has "shown results," and the FY2025 strategy statement is built around a shift toward higher-value industrial-control and AI-edge memory [21]. Nowhere in the corpus is that business quantified — the company states plainly that it publishes no financial forecasts, and no filing breaks out a non-consumer or industrial-AI revenue share [22]. The valuation premium over proven peers rests on a transformation the numbers have not yet been asked to show. A first disclosed figure — even a rough share of revenue that is genuinely non-commodity and sticky — would be among the most informative things management could publish, in either direction.

What would change the read

On today's evidence the weight of the report sits with the cyclical interpretation: a sub-scale price-taker with no established moat, a margin that has only ever cycled, a record year that produced no cash, and a distribution funded by dilution. The strongest fact against that read is genuine and current — contract prices were still rising into late 2026, the customer base did widen in FY2025, and the chair has real capital guaranteeing the balance sheet.

Two developments would materially shift the read toward durability: a quantified, growing non-consumer revenue share that holds while commodity prices fall, and a gross margin that steps down to a new floor above the historical high-single digits rather than back into it. Two would confirm the cyclical read: a monthly revenue print that turns down year-over-year once prices peak, and the reappearance of an inventory write-down charge. All four are observable in filings this company must publish, on the cadence in the table above. The case does not need to be settled by argument; it will be settled by the next few disclosures.