京元電子 King Yuan Electronics: The Chokepoint of the AI Chip Supply Chain

I. Introduction & Episode Roadmap

There is a building in Miaoli County, about an hour south of Taipei by high-speed rail, that almost nobody outside the semiconductor industry has ever heard of. It has no showroom. It has no consumer brand. Its lobby is unremarkable. Inside, though, sit rows of machines the size of restaurant refrigerators, humming in filtered air, each one holding a chip at elevated voltage and temperature for hours at a time, deliberately trying to break it.

Almost every high-end AI accelerator shipped out of Taiwan in the past two years — the GPUs that fill hyperscale data centers, the custom silicon that Google, Amazon, Meta and Microsoft have commissioned for their own workloads — passed through a room like that one before a customer ever powered it on. The company that owns those rooms is 京元電子 King Yuan Electronics, listed in Taipei as 2449, and it describes itself as the world's largest professional pure-play testing company.¹

That phrase is worth slowing down on, because it contains the entire strategic identity of the business. KYEC tests chips. It does not package them. In an industry where the dominant model has been to bundle assembly and test into a single "OSAT" — outsourced semiconductor assembly and test — KYEC spent nearly four decades doing only the second half of that acronym. Testing services have averaged roughly 98% of total revenue.[^2] There is no hidden second business, no consumer division, no software arm. It is a company that makes its money by being the last checkpoint before the world's most expensive silicon ships.

For most of its history that was a respectable, deeply cyclical, largely anonymous way to earn a living. And then AI happened.

Consider what this 39-year-old contract tester has done in roughly eighteen months. It walked away from its entire mainland China manufacturing business, selling its Suzhou subsidiary outright for RMB 4.9 billion.² It posted its first NT$10 billion quarter in company history in the first quarter of 2026, then immediately beat it in the second.³⁴ It raised capital spending to a level that, in 2025, consumed roughly 93% of revenue and two and a half times operating cash flow.⁵ It opened its first-ever overseas factory, in Singapore, in May 2026.⁶ It approved up to US$1.4 billion for a first-ever United States plant in July 2026.⁷ And at the annual meeting in between, shareholders replaced the man who had run the board since 1998 with a vice chairman whose day job is running a medical clinic.⁸

Three simultaneous capacity build-outs on three continents, financed substantially with borrowed money, under a chairman in his first year, at a company whose entire operating history is a lesson in what happens when semiconductor demand turns. That is either one of the great asymmetric bets in the AI supply chain or a textbook illustration of what "capex supercycle" means when the cycle ends.

This is the story of how a wafer-grinding shop became a chokepoint, and of the wager it is now making on that chokepoint being permanent. We will start with the contrarian decision to be a test-only house, move through two decades of scaling inside Taiwan's semiconductor complex, sit inside the down-cycle and the China reckoning of 2022–2024, open up the test floor to understand what actually changed in the physics of testing an AI chip, follow the reacceleration through 2025 and 2026, examine the new leadership and the capital program it inherited, war-game the competitive structure, and then lay out — honestly, with the evidence for and against — why this company might win from here and what would break the case.

The place to begin is with a choice made in the early 1990s that looked, at the time, like a limitation.

---

↑ Back to Top

II. Origins: Betting on Test-Only, 1987–2001

The man who built King Yuan Electronics into what it is today did not found it, and did not start out in semiconductors at all.

李金恭 Lee Chin-Kung graduated from National Taiwan Ocean University in 1980 with a degree in shipping management, and took a job as a fire-insurance surveyor in Taipei, riding a company motorcycle to inspect buildings for NT$8,000 a month. He tried real estate next, opening an agency on Zhongxiao East Road and deliberately courting Japanese and American clients because the commissions were larger. In February 1982 he joined 聯華電子 UMC as a night-shift manufacturing engineer, employee number K02, supervising a crew of twenty-odd technicians. Over the next ten years he climbed from group leader to manufacturing department manager — a decade of learning, from the inside, exactly how a fab actually runs and where its yield problems hide.⁹

King Yuan Electronics itself had been established in May 1987 as a modest back-end shop doing wafer grinding and die sawing.¹⁰[^2] In 1991, Lee and a group of co-investors took it over. At the time it employed a few dozen people. He would run it as president from 1991 and as chairman from 1998.¹¹

The formative decision came in the years that followed, and it was a decision about what not to do. Taiwan in the early 1990s was assembling the world's first fully disaggregated semiconductor supply chain: 台積電 TSMC and UMC would make the wafers, fabless designers would design the chips, and a cluster of back-end houses would package and test them. The natural commercial instinct was to offer everything — take the wafer, package the die, test the part, ship it. That is the model 日月光 ASE and its peers built into a global business.

KYEC went the other way. It added burn-in services in 1993 and logic and memory wafer probe in 1996, and it kept stacking capabilities within test rather than expanding sideways into packaging.[^2] Packaging, in the end, is a materials-science and mechanical-assembly business — substrates, wire bonds, molding compounds, thermal interfaces. Test is an information business. It is about automated test equipment, about writing and optimizing test programs, about squeezing seconds out of test time per unit, about knowing which failure modes to hunt for. The two disciplines share a customer but not much else.

By staying narrow, KYEC could point every dollar of capital and every senior engineer at one question: how do you find the defective chip faster and more cheaply than anyone else? That produced a habit that shows up in the company's engineering culture to this day — it designs and manufactures a meaningful portion of its own test equipment, including burn-in ovens, probe cards and handling tools, rather than buying everything from the equipment majors.[^2] It is an unusual trait for a services company, and it is one of the few places where KYEC has something proprietary rather than merely rented.

It is worth spelling out why specialising here compounds, because the economics are not obvious. A test house is essentially renting out extremely expensive machines by the hour, so its profit on any given contract is determined by one variable above all others: how many seconds the machine spends on each chip. Shave the test program from twelve seconds to nine and you have not improved margin by a quarter — you have created a quarter more sellable capacity out of the same building, the same machines and the same depreciation, without spending a dollar. That work is not glamorous. It is engineers arguing over which test patterns are genuinely catching defects and which are inherited habit, over how many devices can be tested in parallel on one machine, over how to keep a handler from becoming the bottleneck. A firm that also runs a packaging operation has to split its best people and its capital budget across two unrelated technical problems. A firm that does only test does not. Over thirty years, that difference accumulates into something a competitor cannot simply buy.

The narrowness had a cost, and the bill arrived almost immediately after the company listed on the Taiwan Stock Exchange in May 2001.¹⁰ The dot-com collapse gutted memory chip demand, which at the time drove roughly 70% of KYEC's business. Lee later described the economics bluntly: the company was losing money on every unit it tested. A test machine that cost US$5 million had gone from earning NT$10,000 an hour to being unsellable at NT$3,000 an hour. The company faced a cash shortfall of around NT$600 million and survived only by raising US$120 million in convertible bonds.⁹

That is the moment worth carrying forward, because it defines the shape of the risk in this business. A test house's cost base is depreciation and clean-room overhead. Its revenue is hours of machine time sold. When demand falls, the machines do not get cheaper and the clean room does not get smaller — only the hourly rate collapses. The operating leverage that makes an up-cycle spectacular is precisely what makes a down-cycle existential.

Lee drew the correct lesson and applied it once more. In April 2008, months before Lehman Brothers failed, he pushed the company to build up its cash reserves. When the financial crisis hit, KYEC's capacity utilization fell from around 80% to 30% and monthly revenue collapsed from roughly NT$1.2 billion to NT$390 million — and the company survived on the balance sheet he had pre-loaded. His stated philosophy afterward was almost military: in a war, provisions matter.⁹

Hold that thought. Two decades later, the company he built would enter the largest capital commitment in its history with net financial debt rising rather than falling — and with him no longer in the chair.

---

↑ Back to Top

III. Scaling Inside Taiwan's Silicon Shield, 2000s–2010s

If you plotted King Yuan Electronics' physical footprint against a map of Taiwan's semiconductor build-out, the two lines would be nearly indistinguishable.

The company completed its Hsinchu headquarters around 2000, then built Zhunan Fab 1 in 2002 and Zhunan Fab 2 in 2004, added a Suzhou subsidiary in 2004 to serve mainland customers, opened Tongluo Fab 1 around 2008, Zhunan Fabs 3 and 4 by 2010, Tongluo Fab 2 by 2014, Zhunan Fab 5 by 2016, and Tongluo Fab 3 by 2020. Along the way it planted overseas sales subsidiaries in the United States (1998), Singapore (2012) and Japan (2014), and absorbed 東琳精密 Tunglin Precision around 2018.[^2]

Read that sequence again and notice what it is not. There is not a single transformational acquisition in it. For thirty years, KYEC's capital allocation policy was: build another clean room in Miaoli, fill it with testers, repeat. It is one of the purest organic-compounding stories in Taiwanese manufacturing, and it produced a company that by the mid-2010s had become the largest dedicated test house in the world while remaining, by revenue, a fraction of the size of the bundled OSAT giants.

The business it built during those years was a portfolio of end markets rather than a bet on one. Chips tested at KYEC went into memory modules, smartphones, LCD and OLED driver ICs, CMOS image sensors, RF and wireless front-ends, MEMS microphones and accelerometers, automotive controllers, and general logic and mixed-signal parts.[^2]¹² That breadth was deliberate, and management has consistently described a balanced customer structure as the thing that keeps operating profit growing through a volatile semiconductor market.[^2]

The scale that accumulated is easy to underestimate because none of it is visible from a public road. KYEC's Taiwan operations occupy roughly 947,000 square metres of floor area, of which about 502,000 square metres is clean room — a footprint measured in hundreds of football pitches of filtered, temperature-controlled space. Monthly throughput runs to approximately 350,000 wafers on a 12-inch-equivalent basis at chip probe and around 350 million individual integrated circuits at final test, and the company employed 8,976 people as of March 2026.¹²[^2] These are not the numbers of a boutique.

They are also the numbers that explain the cyclicality, and this is the mechanism worth internalising before the rest of the story makes sense. Clean-room square footage and installed testers are decided years before the demand that fills them. A test house commits capital when its customers' forecasts look strong, and the capacity arrives — after construction, installation, calibration and qualification — several quarters later, by which point the forecast that justified it may have reversed. The revenue side of the business responds to global chip demand almost immediately; the cost side responds to a decision made eighteen months earlier and cannot be reversed. That structural mismatch is why KYEC's margins swing far more violently than its revenue does, and why the single most consequential judgement any leadership team at this company ever makes is when to build.

Two capabilities developed in this era matter enormously for what came later.

The first is burn-in. Burn-in is the deceptively simple idea that if you run a chip hot and hard for a while, the units destined to fail early will fail now, in your oven, rather than in a customer's product. KYEC has been building its own burn-in ovens since 1996, and the roadmap of those machines is effectively a history of semiconductor power consumption. The 1996-era oven handled memory parts drawing under a watt. By 2002 it was logic chipsets, still under a watt. By 2008, 20 watts. By 2012, 50 watts. By 2014, automotive parts at 120 watts. By 2016, GPUs at 300 watts, then 500 watts. By 2019, gaming silicon at 700 watts. By 2024, AI chips at 1,000 watts. In 2025 the company was running 1,500-watt air-cooled ovens, and in 2026 it put a 3,000-watt liquid-cooled rack system into mass production.[^2]

That escalation is the whole story in one chart. Every doubling of power dissipation is a thermal engineering problem that has to be solved before a single chip can be stressed, and solving it in-house rather than waiting for a vendor is a genuine, if unglamorous, source of advantage.

The second capability is system-level test, or SLT. Conventional testing checks a chip against a specification: apply these signals, confirm these outputs. SLT does something different — it drops the chip into a rig that behaves like the machine it will actually live in, boots something resembling real software, and watches whether the whole assembly misbehaves. The analogy is the difference between a doctor checking your bloodwork and a doctor putting you on a treadmill. Bloodwork catches many things. The treadmill catches the things that only appear under load.

SLT is expensive because it is slow. Where a conventional final test might take seconds, an SLT run can take many minutes or longer, which means the throughput per machine collapses and the capital required per unit of output multiplies. For most of the 2010s, that made SLT a niche applied to the highest-value, highest-risk parts. Which was fine — until the industry started building chips where every single unit was a highest-value, highest-risk part.

By the end of the 2010s, then, KYEC occupied a very particular position: it was the largest independent test specialist on earth, it had spent thirty years accumulating thermal and system-test capability that nobody needed at scale yet, and it was a rounding error next to ASE in any table that mixed packaging revenue with test revenue. It had built the right factory for a demand curve that had not arrived.

Before it did, the company had to survive one more down-cycle — and this one came bundled with a geopolitical problem it could not engineer its way out of.

---

↑ Back to Top

IV. The Down-Cycle and the China Reckoning, 2022–2024

The post-COVID hangover in semiconductors was not subtle. After two years in which the world bought every laptop, phone and television it could find, the inventory in the channel had to be worked off, and the back end of the supply chain — the part that only gets paid when units actually move — absorbed the shock directly.

For KYEC the arithmetic was brutal and quick. On a continuing-operations basis, quarterly revenue fell from roughly NT$7.4 billion in the second quarter of 2022 to NT$5.5 billion by the first quarter of 2023, a decline of about a quarter in three quarters. Gross margin, which had been running in the mid-30s, compressed to 32.5%. And capacity utilization for the full year 2023 came in at 54.0% — meaning nearly half the company's installed test capacity sat idle across an entire fiscal year.[^2]

Consolidated reported revenue tells the story with an important complication attached. The headline series shows NT$36.8 billion in 2022, NT$33.0 billion in 2023, and NT$26.9 billion in 2024 — a two-year, roughly 27% decline that reads as an unbroken collapse.¹³

Myth versus reality: what actually shrank

It is worth being precise here, because the popular version of this story is wrong in a way that matters.

The 2022 and 2023 headline figures included the Suzhou operation. The 2024 figure did not. Strip the China business out of every year and the underlying Taiwan test business went from roughly NT$27.6 billion in 2022 to NT$24.0 billion in 2023 — a real but ordinary cyclical decline of about 13% — and then grew to NT$26.9 billion in 2024.[^2] In other words, the core business bottomed in 2023 and was already recovering through 2024. The apparent second down-year in the reported numbers is substantially an artifact of removing a divested subsidiary from the consolidation.

This distinction cuts in both directions for an investor. It means the down-cycle was less catastrophic than the headline suggests — but it also means the AI-driven "recovery past the old peak" is being measured against a peak that included a business the company no longer owns. Comparability across 2022–2025 for this company requires care, and any analysis that simply lines up reported revenue year over year is measuring two different companies.

The decision to leave

The China business was not failing. It was becoming unownable.

King Long Technology (Suzhou), established in 2004, sat squarely in the path of the tightening American export-control regime aimed at China's semiconductor industry. In its own filing, KYEC's board stated plainly that it "has made a decision to withdraw from [mainland] China's semiconductor manufacturing business," attributing the move to "the impact of the United States' restrictions on mainland China's semiconductor industry, which has led to changes in the cross-strait chip supply chain, and on intensified market competition."²

In 2024, KYEC Microelectronics agreed to sell its entire 92.16% stake in King Long for RMB 4.9 billion — about US$676 million — to a consortium that included 通富微电 Tongfu Microelectronics, King Legacy Investments, LePower (HK), Anchor Light Holdings, the Suzhou Industrial Park Industrial Investment Fund, and a Shanghai state-linked private equity partnership.² The buyer list is itself a document of the era: a Taiwanese company's mainland assets being absorbed by a mainland competitor and a set of Chinese state-connected funds, precisely because a third country's export rules made the cross-strait arrangement untenable.

Management framed the use of proceeds specifically: reinvestment in "high-end testing technology and equipment to meet the strong demand in artificial intelligence, high-performance computing and related markets."²

What this says about capital allocation

Here is where it is worth being disciplined rather than admiring.

This divestiture is the closest thing King Yuan Electronics has to a major corporate event in the past decade, and it is a sale, not a purchase. That is genuinely unusual. The comparison that frames it is 日月光 ASE, which spent years pursuing 矽品精密 SPIL and consummated the combination through a joint share exchange on April 30, 2018, creating a holding company that consolidated the two largest packaging-and-test operations in Taiwan under one roof.¹⁴ ASE emerged from that process with roughly 44.6% of blended global OSAT revenue in 2024.¹⁵ KYEC has never attempted anything remotely comparable. Its capital-deployment story is essentially all organic capex.

So there is no overpaid-acquisition analysis to run here, and it would be intellectually dishonest to manufacture one. What can be assessed is the character of the exit. KYEC sold a functioning asset, at a negotiated price, to a consortium of willing buyers, at a moment of its own choosing — rather than waiting for an export-control action to strand the asset and force a distressed sale. Proceeds were redeployed into the part of the business with the highest visible demand. On the evidence available, that is a competent piece of capital allocation.

The honest caveat: outside valuation benchmarks for King Long's standalone earnings are not publicly available, so the multiple achieved cannot be independently verified. "Sold at a fair price" is a claim the company has not been asked to prove.

What can be verified is where the money went, and how fast. The cash landed in the first quarter of 2025, and it moved KYEC to a net cash position — net financial debt to equity of negative 17.8% at March 31, 2025, the strongest balance sheet the company had shown in years.[^2]

Within twelve months, that number was positive 29.7%.[^2]

The China exit, in other words, was not a deleveraging event. It was a financing bridge — and the thing it financed is the subject of the rest of this story. But to understand why management believed the spending was justified, you first have to understand what actually happens on a test floor when the chip on the socket is an AI accelerator.

---

↑ Back to Top

V. Inside the Test Floor: What KYEC Actually Does and Why It Suddenly Matters

Picture a freshly finished 300-millimeter wafer coming off a TSMC line. It is a disc of silicon roughly the size of a dinner plate, carrying anywhere from a handful to thousands of individual chips. Some of them work. Some of them do not. Nobody knows which is which.

Finding out is a four-stage process, and KYEC sells all four.

Chip probe (CP) comes first, while the chips are still attached to the wafer. A probe card — a precision fixture with hundreds or thousands of tiny needles — lands on the contact pads of a single die, electrical signals go in, responses come out, and the die is marked good or bad. The point of doing this early is economic: packaging is expensive, and you do not want to spend it on a die that was dead on arrival. In the first quarter of 2026, wafer test was 32.1% of KYEC's revenue.[^2]

Final test (FT) happens after the die has been packaged, and it is the largest single piece of the business at 58.9% of first-quarter 2026 revenue.[^2] Here the packaged part is loaded into a socket by a handler, run through a full functional and parametric sweep at multiple temperatures, and sorted into performance bins. This is where a chip gets its speed grade.

Burn-in is the stress stage described earlier — hours at elevated temperature and voltage to force the weak units to fail before shipment. It contributed 7.3% of revenue in the quarter.[^2] Small in the mix, disproportionately important in the physics.

System-level test (SLT) is the newest and most demanding layer, the treadmill rather than the bloodwork, and it is the layer that AI has made central.

Why the AI accelerator broke the old math

Three things changed at once.

The first is thermal. According to reporting on the company's AI positioning, AI chips have pushed power draw from roughly 3 kilowatts toward as much as 8 kilowatts per device, and that lengthens test cycles.¹ You cannot pump kilowatts of heat into a socket without an engineered solution for getting it back out, which is why KYEC's move to liquid-cooled 3,000-watt burn-in racks in 2026 was not a marketing item but a prerequisite for taking the business at all.[^2]

The second is time. A longer, harder test on a slower, more complex part consumes more machine-hours per chip. In a business where the unit of production is the machine-hour, that is the same thing as saying each AI chip consumes several times the capacity of a mobile processor. Demand for test capacity therefore grows faster than unit volumes of AI chips — a leverage effect that works in KYEC's favor as long as the chips keep getting harder.

The third is stakes. A defective phone SoC costs a manufacturer a phone. A defective accelerator that passes final test and fails in a rack three months later costs a hyperscaler a node, the debugging time to find it, and the trust of an operations team running training jobs across tens of thousands of interconnected devices. When silent data corruption at scale is the failure mode you fear, you buy more test, not less.

Put those together and test capacity became, alongside advanced packaging and high-bandwidth memory, one of the genuine physical bottlenecks in the AI supply chain — a place where the constraint is clean-room square footage and installed equipment, neither of which can be conjured in a quarter.

Why "we don't do packaging" turned into a selling point

For thirty years, KYEC's refusal to offer packaging was a scope limitation. In the AI era it became something closer to a positioning statement.

A fabless AI chip designer, or a hyperscaler commissioning a custom ASIC, has to decide where its most sensitive silicon gets tested. Testing generates extraordinarily detailed data: yields, failure signatures, bin distributions, thermal behavior at the edge of spec. If your test house is owned by the same company that packages chips for your direct competitors, you are handing that data to an organization with commercial relationships across your rivals. KYEC, structurally, has no packaging business to create that conflict. It is a neutral utility.

Whether customers explicitly pay for that neutrality is not disclosed. But the customer roster is consistent with the thesis. In its May 2026 investor presentation, the company disclosed that 48% of the world's top 50 semiconductor companies use its testing services, rising to 63% once mainland Chinese and memory-focused companies — which it does not target — are excluded. The customer logos it marks include Nvidia, Intel, Broadcom, AMD, MediaTek, Analog Devices, STMicroelectronics, Renesas, Marvell, onsemi, Microchip, Bosch and Mobileye.[^2] Trade coverage has identified Nvidia as its principal AI customer, with custom-ASIC programs from AWS, Google, Meta and Microsoft expanding the base.¹

That is a real proof point. Being qualified on the most thermally and electrically demanding accelerators in production is not something a company talks its way into; test programs are qualified device by device, over months, with the customer's engineers on site.

It is also, viewed from the other side, a concentration risk that this article will return to. Over 90% of KYEC's business comes from fabless firms rather than integrated device manufacturers or foundries,¹² and the fabless world in AI is a very short list of very large names.

One more structural note before moving on. There is no diversification story to unlock at this company and no conglomerate discount to argue about. Test services are roughly 98% of revenue, with the small remainder coming from equipment and tool sales, rental income and miscellaneous items.[^2] All the interesting variation happens in end-market mix within test — and in the first quarter of 2026 that mix was data processing 33.9%, consumer 31.1%, communications 18.9%, automotive 13.3%, and industrial 2.4%.[^2] Note that consumer and communications together were still half the business. The AI story is the growth engine, not yet the whole company.

That mix shift, and the money spent to enable it, is what the past two years have been about.

---

↑ Back to Top

VI. The AI Supercycle Hits, 2025–2026

The turn shows up first in the quarterly revenue line, and once it starts it does not stop.

Beginning from the trough, KYEC's quarterly revenue ran NT$7.3 billion in the first quarter of 2025, then NT$8.4 billion, NT$9.3 billion, NT$10.0 billion, and NT$10.2 billion in the first quarter of 2026 — five consecutive quarters of sequential growth. Gross margin marched in lockstep: 33.5%, 35.5%, 36.0%, 37.7%, 39.7%.[^2] The second quarter of 2026 brought NT$11.14 billion, up 9.3% sequentially and 33.2% from a year earlier, another record.⁴

Margin expanding alongside volume is the analytically interesting part. In a capacity business, revenue growth alone can be bought by adding machines; margin expansion alongside it means either better pricing, better mix, or better throughput per machine. Management has attributed the improvement to AI parts being functionally complex with longer test times, which lifts realized margin.⁴ The mix explanation is the most consistent with the evidence: work is shifting toward higher-value, longer-cycle-time test on parts where the customer's tolerance for failure is near zero.

For the full year 2025, revenue reached NT$34.93 billion and operating income NT$9.00 billion, against NT$26.86 billion and NT$6.17 billion in 2024.¹³ First-half 2025 net profit had nearly doubled to NT$6.47 billion from NT$3.27 billion, on revenue up 25.3%.⁵ Return on equity climbed from 15.0% in 2023 to 17.9% in 2024 to 21.9% in 2025.¹³

Separating the real growth from the accounting

Reported net income for 2025 was NT$11.06 billion. That number requires an asterisk, and investors should apply it.

Net income from continuing operations in 2025 was NT$8.00 billion. The remaining NT$3.05 billion came from discontinued operations — the China divestiture. The same pattern holds in the prior two years: NT$3.33 billion of discontinued-operations income in 2024 and NT$1.73 billion in 2023.¹³ Across three consecutive fiscal years, in other words, a material share of headline earnings came from exiting a business rather than running one.

Strip it out and the underlying trajectory is still good — operating income up roughly 46% in 2025 — but it is a different number than the headline implies, and anyone anchoring on reported EPS growth is partly measuring a one-time gain that will not recur. Management's disclosure of the split is clear and appropriate; the risk is in how the number gets consumed, not how it gets reported.

The market noticed

It noticed thoroughly. Enterprise value to EBITDA for KYEC ran at roughly 3.0x in fiscal 2022, 7.3x in 2023, 10.3x in 2024, and 18.7x in 2025 — a six-fold re-rating in three years.¹³ Over the trailing twelve months into August 2026 the shares have traded between roughly NT$110 and NT$350.[^17]

A three-to-one range in a year is not a valuation; it is an argument. The bull and bear cases about AI test demand are being fought out in the share price in real time, and the honest conclusion from the multiple expansion is that a great deal of the structural thesis is already reflected in the price. An investor buying today is not being paid to discover that AI needs more testing. That is consensus. The open question is whether the duration and terminal capacity assumptions embedded in a high-teens EBITDA multiple hold.

The tension hiding in the utilization number

Here is a detail that deserves more attention than it gets. For a company that describes its expansion as urgent, KYEC's blended capacity utilization was 60.1% for full-year 2025, 63.2% in the fourth quarter of 2025, and 59.2% in the first quarter of 2026.[^2]

Utilization below two-thirds while spending nearly a full year of revenue on new capacity looks contradictory until you decompose it. Two things are happening at once. New capacity is being installed faster than it fills, which mechanically depresses the ratio in any quarter of aggressive build-out. And — more importantly — the installed base is not fungible. The AI-capable test cells with high-power burn-in and SLT capability are the bottleneck; the older cells configured for driver ICs, consumer parts and mature-node logic are not full.

That single number therefore captures both halves of the argument. The bull reads it as headroom. The bear reads it as evidence that the non-AI half of the business — still roughly half of revenue by application mix — is running soft, and that the cushion beneath the AI growth story is thinner than the consolidated numbers suggest.

On the company's November 2025 investor call, management said the binding constraint was clean-room space rather than equipment, and disclosed that it had leased a former Wintek facility in Yangmei expected to reach full capacity in the first half of 2026, describing the need as extremely urgent. It also reported that GPU testing, which began in 2023, remained the primary revenue contributor, while ASIC testing, started in 2025, showed what it called an extremely large growth rate for 2026–2027.¹⁶

That is a specific, falsifiable claim about where growth comes from next, and it is the right one to hold management to. GPU test is the established business. Custom-ASIC test is the incremental bet — and it is the one that depends on hyperscaler silicon programs staying on schedule.

Which brings the story to the question that will decide everything: is this structural, or is it a capex cycle? The company answered by spending as if it were structural — and it did so while changing who sits in the chair.

---

↑ Back to Top

VII. New Chairman, New Capital Program

Leadership transitions at family-influenced Taiwanese industrial companies tend to be gradual and telegraphed. This one was neither.

The sequence begins in 2023 with an operational change. On August 8, 2023, KYEC announced that president 劉安炫 Liu An-hsuan had resigned for family reasons, and that executive vice president 張高薰 Gauss Chang would take the role effective September 1, 2023.¹⁷ Chang was not a parachute hire. He joined KYEC in 2000 after roughly eleven years in semiconductor sales, marketing and engineering management, including a stint as a foundry sales manager at 華邦電子 Winbond. He holds a physics degree from National Cheng Kung University and an MBA from Saginaw Valley State University in Michigan, and before becoming president he ran strategic planning, market development and sales.[^2]¹¹ He now carries the additional title of chief sustainability officer.[^2]

The profile matters. Chang is a commercial and strategic operator, not a fab engineer — someone whose formative years were spent on the customer side of the relationship. In an era when KYEC's job is to persuade a handful of enormous fabless customers to commit their most demanding programs to its floors, that is arguably the right skill set. He has been the public face of the expansion: he was in Singapore in May 2026 explaining why the company chose it, and his framing there was notably unsentimental — the local market is limited, he acknowledged, but Singapore's understanding of global supply chains lets it serve as a strategic node.⁶

Then came the board.

At the annual general meeting on May 29, 2026, shareholders elected 謝其俊 Hsieh Chi-Chun as chairman, succeeding Lee Chin-Kung after Lee's roughly 28 years in the role. Lee stayed on as an ordinary director. Hsieh had been a director since the late 1990s and vice chairman for over a decade. His day job is running a joint clinic in Toufen, Miaoli County — he is a physician by training. He holds about 5.55 million shares, roughly 0.45% of the company. The meeting also elected six general directors, including Lee, Hsieh and Chang, plus three independent directors.⁸¹⁸

Taiwanese business press described Lee's departure from the chairmanship as unexpected.⁸

The credibility question, stated plainly

There is no professional insult in observing that a physician running a regional clinic is an unusual profile for the chairman of a company about to undertake the largest capital program in its history. It is simply a fact an investor has to underwrite.

The case for continuity is real: Hsieh has sat on this board since the 1990s and served as vice chairman since 2014, so he has been present for every strategic decision of the past decade including the China exit. The chairman's role at a Taiwanese listed company is also more supervisory than executive, and day-to-day operating authority rests with Chang.

The case for scrutiny is equally real. A chairman with a modest personal stake and no deep technical or capital-markets operating record is, structurally, less likely to be the person who tells an enthusiastic management team that a NT$50 billion capital plan should be NT$35 billion. Independent trackers have flagged that KYEC's board independence is comparatively low. And the transition is happening at the exact moment when the board's most valuable function would be saying no.

This is not a prediction that governance will fail. It is an observation that the safeguard is untested, and it will be tested by the next downturn rather than by the next good quarter.

The capital program: what was planned, what was withdrawn, what got spent

The sequence here has been widely mis-told, and the correction is genuinely important.

In August 2025, KYEC's board raised the 2025 capex budget by 37% to NT$37 billion, up from the NT$26.97 billion approved in May and against just NT$14.86 billion actually spent in 2024. Alongside it, the board approved a cash capital increase of 80 million new shares to fund equipment purchases, with 80% offered to existing shareholders, 10% to employees and 10% to public subscription.⁵

The rights issue never happened. On October 9, 2025, the company announced it was withdrawing the offering, citing the impact of recent US tariff policy, sharp changes in domestic and international economic conditions, and high short-term uncertainty in capital markets, saying the withdrawal was to protect the interests of all shareholders and that it would revisit timing later.¹⁹

So the 2025 build-out was not funded by equity. Actual 2025 capital expenditure came in at NT$32.36 billion — below the NT$37 billion plan but more than double the prior year — against operating cash flow of NT$13.15 billion.[^2] Free cash flow was deeply negative. The gap was bridged by the China sale proceeds and by roughly NT$9.9 billion of net new long-term borrowing.¹³

You can watch the balance sheet absorb it quarter by quarter. Net financial debt to equity moved from negative 17.8% in March 2025 to 4.0% in June, 18.5% in September, 24.4% in December, and 29.7% by March 2026.[^2] On the November 2025 call, management put the debt-to-total-assets ratio at about 46% and characterized it as reasonable and controllable.¹⁶

That characterization is defensible on today's numbers — gearing near 30% is not aggressive for an asset-heavy business with an EBITDA line above NT$16 billion. What makes it worth watching is the trajectory rather than the level: the company consumed an entire divestiture windfall plus meaningful new debt in four quarters, and the spending is accelerating rather than moderating.

2026: bigger, and on two new continents

The 2026 program is larger than 2025's. The board initially approved NT$39.37 billion, then raised it to NT$50 billion — roughly a 27% increase and a company record, with the incremental spending directed at fab and clean-room construction plus equipment.²⁰²¹ For context, Taiwan's three largest OSAT players combined were projected to spend around NT$370 billion in 2026, a third consecutive record year for the industry.²⁰

And for the first time, the money is going overseas.

The Singapore facility opened in Ang Mo Kio on May 22, 2026, backed by a SGD100 million capital injection — roughly NT$2.47 billion — and equipped for chip probe, final test, burn-in and system-level test, targeting automotive, high-performance AI and consumer semiconductor testing across Southeast Asia.²²⁶ Executives from Nvidia and MediaTek attended the opening, which is itself a signal about who the plant is being built for.⁶

Then, on July 10, 2026, the board approved an investment of up to US$1.4 billion — roughly NT$44.9 billion — for a chip-testing plant in the United States, explicitly in the context of continued American pressure to expand domestic chip manufacturing.⁷ Reuters characterized the company as an Nvidia supplier in reporting the decision.²³ As of this writing, KYEC has not disclosed a state, city or site, and no construction timetable or completion date has been announced. The US$1.4 billion figure is an authorization ceiling, not a committed spend.

That last point deserves emphasis, because it is easy to read a board authorization as a construction budget. It is not. It is permission.

Capital returns: a company that has never really done them

For the 2026 distribution, shareholders approved NT$1.50 per share — NT$1.00 in cash and NT$0.50 in stock — with the press explicitly linking the mixed structure to the company's substantial capital expenditure requirements.¹⁸⁸ A stock dividend, of course, distributes nothing; it retains cash and issues paper.

There have been no meaningful buybacks. This has always been a reinvest-in-capacity company rather than a cash-return one, and the pattern is intensifying under the new regime, not changing. Investors should treat KYEC's dividend as a gesture rather than a policy.

The activist's stress test

What would a skeptical investor actually challenge here?

They would start with the funding stack. The 2025 rights issue was announced, then withdrawn under market-condition pressure, meaning the equity leg of the original financing plan simply did not happen — and the shortfall was met with debt and divestiture proceeds. Analyst work in April 2026 suggested KYEC had substantial undrawn bank credit lines and was not expected to need a near-term cash capital increase.²¹ But "we have credit lines" and "we can fund this from operations" are different statements, and only the first is currently true.

They would challenge the simultaneity. Three build-outs — Taiwan expansion, a first-ever Singapore plant ramping toward mass production, and a first-ever US plant — is a lot of concurrent construction, hiring and qualification for a management team whose entire institutional experience is building clean rooms within a 90-minute drive of Hsinchu.

They would note the governance timing: an unexpected chairman change, a comparatively less independent board, and a chairman with a 0.45% stake, coinciding with the largest commitment in company history.

And they would ask the uncomfortable question about the dividend structure. Shifting toward stock dividends while capex runs near revenue is a reasonable liquidity decision — but it is also the behavior of a company that would rather not admit it cannot afford a full cash payout.

Management has been visible with investors throughout — appearances at UBS's Asian Investment Conference in May 2026, HSBC's conference on June 1, and Nomura's Investment Forum Asia on June 4–5, 2026, among others.²⁴²⁵ Visibility is not the same as accountability, but a management team avoiding hard questions does not typically book that calendar.

To judge whether the spending is rational, you have to understand who else could take this work.

---

↑ Back to Top

VIII. Industry Structure: Five Forces and the Sources of KYEC's Power

Line up the outsourced assembly and test industry by 2024 revenue and King Yuan Electronics looks almost irrelevant.

ASE Technology sat at roughly 44.6% of blended global OSAT revenue on about US$18.5 billion. Amkor followed at around 15.2%, then 长电科技 JCET at about 12%, 通富微电 Tongfu at 8%, 力成科技 Powertech at 5.5%, 华天科技 HT-Tech at 4.8%, WiseRoad at 3.7%, and then a cluster in the low single digits: 하나마이크론 Hana Micron at 2.2%, KYEC at 2.2%, and 南茂科技 ChipMOS at 1.7% on roughly US$710 million.¹⁵

Ninth place. Two percent. Hardly a chokepoint.

Myth versus reality: the market share number is measuring the wrong thing

That table adds packaging revenue to test revenue and ranks the total. ASE's US$18.5 billion is overwhelmingly packaging. KYEC does essentially no packaging. Ranking a test-only company inside a packaging-weighted table is like ranking a specialist cardiac hospital by total beds against general hospitals — the number is accurate and the conclusion drawn from it is wrong.

KYEC's actual competitive weight is within test specifically, where its position is far stronger. The company's own framing is that it ranks second globally by testing revenue and first among pure-play testers.¹²¹

The necessary caveat: no independently verified test-only market share figure is publicly available. "World's largest pure-play tester" is a company claim repeated in trade press, not an audited statistic. It is probably directionally right — the customer roster and the AI qualifications support it — but an investor should hold it as a well-supported assertion rather than a fact.

With that framing, run the forces.

Supplier power: high, and it doubles as a moat. Automated test equipment is dominated by アドバンテスト Advantest and Teradyne, with Cohu and SPEA in supporting roles. KYEC's own disclosure of its installed platforms reads like a catalogue of both majors' product lines — Advantest's T2000 and V93000 families, Teradyne's J750, UltraFlex and ETS families — across logic, mixed-signal, image sensor, RF and SoC applications.[^2] These are expensive, long-lead-time machines, and in a tight cycle allocation itself becomes a constraint.

That gives suppliers real pricing power over KYEC. But it cuts the other way at the industry level: the same equipment cost that squeezes KYEC's margins is what stops a new entrant from appearing. This is Hamilton Helmer's scale economies power in its most literal form. With more than 5,500 test systems installed, KYEC can amortize equipment, spread engineering across many programs, and — critically — keep a diverse enough platform inventory to serve low-volume, high-mix work that a smaller house cannot economically staff.[^2]

There is also a partial escape hatch. KYEC designs and builds a meaningful portion of its own burn-in ovens, testers and probe cards, which reduces dependence on the majors for exactly the equipment classes where AI demand is most acute and lead times are longest.[^2] That is closer to Helmer's process power — accumulated, hard-to-copy internal capability — than to anything a competitor can purchase.

Customer power: high, and concentrated. A short list of enormous fabless designers and hyperscaler ASIC programs represents a large share of incremental demand. These are sophisticated buyers with procurement organizations designed to extract price. Over 90% of KYEC's customers are fabless.¹²

The counterweight is switching cost, and it is genuine. A test program is not a commodity service; it is a device-specific package of hardware fixtures, probe cards, load boards, thermal solutions and test software, qualified over months with the customer's engineers involved. Moving a program mid-life means requalifying, re-correlating results between sites, and accepting risk on a part already in production. That is real switching power on existing programs.

It is materially weaker on new programs. Every new device generation is a fresh competition. KYEC's advantage is incumbency and qualification history, not lock-in.

Threat of new entrants: low at the high end, real at the low end. Nobody is building a competing high-power AI burn-in and SLT operation from a standing start; the capital, the clean-room lead time and the thermal know-how make that a multi-year project. But Chinese state-backed testers — JCET, Tongfu, HT-Tech — have been scaling aggressively, and in 2024 the Chinese cohort collectively posted the strongest growth in the industry.¹⁵ They compete on legacy and mature-node test, where the technical barrier is lowest, and they compete on price.

This matters more than it appears. The consumer, communications and driver-IC work that still constitutes roughly half of KYEC's revenue mix is precisely the work Chinese competitors are targeting. That business is the shock absorber under the AI growth story, and it is under structural price pressure.

Threat of substitution: slow-moving but not zero. Large integrated manufacturers — Samsung, Intel, Texas Instruments — retain significant test capability in-house. A hyperscaler could in principle build its own SLT capacity for its custom silicon. The reason this has not happened is not that it is impossible; it is that test is a utilization business, and a single customer's own volume cannot fill a facility efficiently across the peaks and troughs of a product cycle. Outsourcing exists because pooling demand across customers is what keeps expensive machines busy. That logic holds unless one customer's AI volume alone becomes large enough to fill a dedicated facility — which is, notably, no longer unimaginable.

Rivalry: the disintermediation risk that hasn't happened. The clearest long-run threat to KYEC's model is the bundled one-stop shop. ASE and Amkor can offer packaging and test as a single integrated flow, with one supplier relationship, one quality system and one throat to choke. In advanced packaging especially — where the package itself has become an active engineering component — the case for keeping test physically and organizationally close to packaging gets stronger, not weaker.

That risk has not materialized in the AI era so far, and the reason appears to be the neutrality argument combined with capacity scarcity: when everyone is short of test capacity, customers dual-source. If capacity ever becomes abundant, the integrated bundle becomes a more compelling pitch, and KYEC's independence converts from an asset back into a scope limitation.

Summing the powers honestly: KYEC has scale economies within test, real process power in self-built equipment and accumulated thermal engineering, and switching costs on qualified programs. It does not have network effects, brand power in any meaningful sense, or cornered resources. It is a capital-intensive specialist with a genuine but bounded edge — and its edge is strongest exactly where demand is currently hottest, which is a nice place to be and a dangerous place to extrapolate from.

---

↑ Back to Top

IX. Bull vs. Bear

Strip away the narrative and the investment question reduces to a single disagreement about time.

Why this company wins from here

The bull case rests on a physical argument rather than a financial one. Each generation of AI accelerator consumes more test time than the last, because it dissipates more power, contains more heterogeneous components in a single package, and serves an application where undetected failure is catastrophically expensive. That means test demand grows faster than AI chip units. If AI silicon volumes merely hold flat while complexity keeps rising, KYEC's addressable machine-hours still grow.

The evidence supporting this is concrete rather than rhetorical: gross margin has expanded through five consecutive quarters of revenue growth,[^2] which is what mix shift toward harder work looks like in a financial statement. Management's own account of the constraint — clean-room space, not equipment or demand — is consistent with a business turning away work.¹⁶

The independence argument has evidentiary support too. A pure-play tester with no packaging business creates no conflict for a fabless designer or a hyperscaler, and KYEC's presence on the most demanding accelerator programs, with 63% of the world's top 50 non-Chinese, non-memory semiconductor companies as customers, indicates the qualification bar has been cleared repeatedly.[^2]

Geographic diversification, whatever its execution risk, reduces two exposures at once. The Singapore plant and a prospective US plant lower the share of the company's earnings sitting inside a single geopolitical fault line — a fault line KYEC has already been burned by once. And the separation of the chairman and president roles, with an experienced commercial operator running the company day to day, is a modest governance improvement over a single-figure structure.

What breaks the case

The bear case does not require AI to fail. It only requires AI capex growth to decelerate.

Start with valuation. A six-fold expansion in EV/EBITDA over three years¹³ means the structural thesis is not a discovery — it is the price. The share price range of roughly NT$110 to NT$350 over the trailing year[^17] tells you the market's own confidence interval is enormous.

Then the balance sheet. Capital expenditure ran at approximately 93% of revenue and 2.5 times operating cash flow in 2025, producing deeply negative free cash flow, funded by divestiture proceeds and new borrowing after the planned equity raise was pulled.¹³¹⁹ The 2026 plan is larger still.²⁰ A company that spends a year of revenue on fixed assets is making a bet on utilization it cannot easily unwind — clean rooms and testers cannot be resold into a soft market at anything like cost, as the company itself learned in 2001 when US$5 million test machines could not be booked at NT$3,000 an hour.⁹

That is the operating-leverage trap in a sentence. The same fixed-cost base that turns a 33% revenue increase into a 46% operating income increase turns a revenue decline into something much worse. And with blended utilization already below two-thirds while capacity is being added,[^2] the buffer is thinner than the growth rate suggests.

Customer concentration compounds it. Growth is disproportionately driven by a small number of AI accelerator programs. Specific customer revenue percentages are not disclosed, which is itself a disclosure gap worth noting — but the qualitative concentration is not in dispute. A single hyperscaler pausing or re-timing a custom silicon program does not reduce KYEC's revenue proportionally; it strands purpose-built capacity.

Execution risk is now stacked three deep. Taiwan expansion, a Singapore plant ramping toward mass production, and a US plant with no announced site or schedule, all running concurrently under a management team with no prior overseas manufacturing experience. Overseas fabs are historically where Asian manufacturers discover that labor cost, labor availability, construction timelines and regulatory processes do not behave the way they do at home.

Chinese competitive pricing erodes the base. And governance provides less of a brake than it might: a first-year chairman, a comparatively less independent board, and an owner-operator founder who has stepped back from the chair.

The honest synthesis

King Yuan Electronics has a strong, evidence-backed answer to the first question an investor should ask: why does a chip need to go through this company? The answer is qualified test programs on the hardest devices in production, thermal capability built over three decades, more than 5,500 installed systems, and a structural neutrality that packaging-owning rivals cannot replicate.

It has a much thinner answer to the second: why should this capital program not be the thing that breaks the company if AI capex pauses? The available answer is that management believes demand is structural, that gearing near 30% is manageable, and that the credit lines are there. Those are reasonable statements. None of them has been tested by a downturn, and the one management team that had been tested — the one that raised cash before Lehman — is no longer the one in the chair.

That is the trade. Everything else is detail.

---

↑ Back to Top

X. Risk Radar

AI demand cyclicality. This is the dominant risk and it does not require a crash to hurt. KYEC has roughly tripled its annual capital spending in two years on the assumption that AI test demand keeps compounding.[^2] Fixed costs in a test house are depreciation and clean-room overhead; revenue is machine-hours sold. If hyperscaler capex growth merely decelerates, newly commissioned capacity arrives into a market that no longer needs it, and the operating leverage that produced 39.7% gross margins in the first quarter of 2026[^2] reverses with equal force. The 2023 experience is the calibration point: 54% utilization for a full year and gross margin down to 32.5%.[^2]

Customer concentration. A small number of named AI accelerator customers drive an outsized share of recent growth, with Nvidia identified in trade coverage as the principal one and hyperscaler ASIC programs as the expansion vector.¹ KYEC does not disclose revenue by customer. The mechanism of harm is specific: AI test capacity is purpose-configured, so a paused program does not free up capacity that can be redeployed to consumer parts at similar economics.

Execution risk across three concurrent builds. Taiwan capacity expansion, the Singapore ramp toward mass production, and a US plant with no disclosed site or timeline are all live simultaneously.⁶⁷ Each requires clean-room construction, equipment installation, and — hardest of all — trained test engineers. Semiconductor test talent is scarce in Taiwan and scarcer in Singapore and the United States. The most likely failure mode is not cancellation but delay and cost overrun, which in a business earning returns on installed capital shows up directly as depressed returns.

Leverage and future dilution. Net financial debt to equity moved from negative 17.8% to 29.7% in twelve months,[^2] with debt-to-assets around 46% by management's own account.¹⁶ The equity leg of the 2025 funding plan was withdrawn rather than executed,¹⁹ which means the shareholder-dilution question was deferred, not resolved. If the US project proceeds toward the full US$1.4 billion authorization while free cash flow remains negative, a return to the equity market becomes a live possibility that investors should price rather than assume away.

Geopolitical and export-control exposure. Exiting mainland China removed one specific exposure and did not remove the category. KYEC remains a Taiwan-headquartered supplier to American chip designers, dependent on Japanese and American test equipment, now building in the United States partly in response to reshoring pressure. Tariff policy has already had a concrete effect on this company — it was the stated reason the 2025 rights issue was withdrawn.¹⁹ Further shifts could affect equipment sourcing, customer routing decisions, or the economics of the planned US facility.

Chinese competitive pricing. JCET, Tongfu and HT-Tech are scaling in legacy and mature-node test with state backing and posted the industry's strongest growth in 2024.¹⁵ This does not threaten AI accelerator test in the near term. It does threaten the consumer, communications and driver-IC revenue that still made up a substantial share of KYEC's application mix in early 2026[^2] — the cushion beneath the AI story.

Governance and succession. A chairman in his first year, without a deep semiconductor or capital-markets operating record and holding roughly 0.45% of shares,⁸ inherits the largest capital program in company history alongside a board flagged by independent trackers as comparatively less independent. The founder remains on the board, which provides institutional memory but not formal authority.

Accounting and disclosure notes. Two items merit ongoing attention rather than alarm. First, discontinued-operations gains materially inflated reported net income in fiscal 2023, 2024 and 2025,¹³ making headline earnings growth a poor proxy for operating momentum across that window. Second, KYEC does not disclose revenue by customer, capacity utilization split between AI-capable and legacy cells, or SLT revenue separately — all of which are the metrics that would most directly test the bull case. Neither is a red flag on its own. Both mean an investor is working with less resolution than the thesis requires.

---

↑ Back to Top

XI. Playbook: Business & Investing Lessons

Deliberate narrowness can be a strategy, not an accident — until customers stop valuing it. KYEC spent nearly four decades declining to do the thing that its largest competitors built their businesses on. For most of that time, the decision looked like a ceiling on ambition. What changed was not the strategy but the customer: when the buyer became a fabless AI designer or a hyperscaler with proprietary silicon and acute concerns about who sees its yield data, "we only do test" turned from a limitation into a reason to be chosen. The lesson generalizes badly if you take it as "specialization wins." It generalizes well if you take it as: a durable strategic identity accumulates capability that becomes valuable only when the market's requirements move toward it — and the operator's job is to keep the identity intact long enough for that to happen, which requires surviving the years when it doesn't pay.

A decisive divestiture from a position of strength is a better capital-allocation signal than a large acquisition. KYEC did not sell its Suzhou business because it was losing money. It sold because the regulatory environment made the asset's future ownership untenable, and it sold while there were still competing buyers willing to pay. Compare that to the far more common corporate pattern of holding a structurally impaired asset until the writedown is forced. The signal value is in the timing and the cleanliness — a full exit, not a partial one, with proceeds explicitly redirected to the highest-return part of the business. The caveat is equally instructive: because comparable valuation data is not public, shareholders have to take the price on faith. Good process does not always come with proof.

Expanding capacity ahead of demand only rewards operators when the demand is structural — and nobody knows which it was until the downturn. This is the central unresolved question about KYEC's 2025–2026 build-out, and it is unresolvable in advance. What an investor can assess is whether the operator is behaving as though the answer is uncertain. The signals here are mixed: the company withdrew an equity raise when market conditions turned, which suggests discipline, but it then funded the same spending with debt, which suggests the spending itself was never really optional. Watch what a company does with its financing plan when conditions change; it reveals whether the capex was a considered choice or a competitive necessity.

Leadership transitions that coincide with a company's largest-ever commitments deserve more scrutiny, not less. New chairmen inherit strategy instantly and credibility never. The founder who navigated 2001 and pre-positioned cash before Lehman had a track record built from surviving two near-death experiences — precisely the kind of record that cannot be handed over. An investor evaluating KYEC today is not evaluating a management team with a demonstrated capital-cycle discipline; they are evaluating a management team that inherited one team's reputation while making a different team's bet.

---

↑ Back to Top

XII. What to Watch: KPIs That Matter Most

Three metrics carry almost all of the information about whether this thesis is working. Everything else is commentary.

1. Gross margin trend alongside revenue growth. This is the single cleanest proxy for whether the mix is genuinely shifting toward high-complexity AI and system-level test work, or whether KYEC is simply renting out more machine-hours at similar economics. The reason it works is structural: AI parts take longer to test and command better realized pricing per unit, so a shift toward them shows up as margin expansion even as volumes grow. The historical range gives the calibration — 32.5% at the 2023 trough, 39.7% in the first quarter of 2026.[^2] Margin expanding with revenue means the mix story is real. Margin flat or compressing while revenue grows means the company is buying growth with capacity rather than earning it with capability, and the entire premium multiple rests on the first interpretation.

2. Capacity utilization, read alongside capex-to-revenue. These two have to be looked at together or neither tells you anything. Utilization was 60.1% for 2025 and 59.2% in the first quarter of 2026[^2] while capital spending consumed roughly 93% of revenue.¹³ The healthy pattern from here is utilization rising while capex intensity moderates — evidence that the Taiwan, Singapore and US capacity is filling as it comes online. The warning pattern is utilization flat or falling while capex stays elevated, which would mean capacity is arriving faster than demand and the company is building into its own overhang. Given that management identified clean-room space as the binding constraint,¹⁶ this pairing is the most direct test of whether that constraint was real.

3. Free cash flow and net gearing. Free cash flow was deeply negative in 2025¹³ and net financial debt to equity climbed from net cash to 29.7% in four quarters.[^2] The question this metric answers is not "is the company solvent" — it plainly is — but "when does this capital program start paying for itself?" The Singapore plant targets meaningful production from 2027, and a US plant would be later still. If free cash flow is still negative and gearing still climbing when those facilities are supposed to be contributing, the structural-demand thesis is failing in the only place it ultimately has to show up: cash. Conversely, a return to positive free cash flow with gearing stabilizing would be the strongest available evidence that the build-out was correctly sized.

---

↑ Back to Top

XIII. Epilogue

The calendar ahead is unusually legible for a company this obscure.

The Singapore plant has to ramp from an inaugurated building into a qualified, revenue-generating operation, with mass production targeted from 2027 — the first real test of whether KYEC's test-floor discipline travels outside Miaoli.²²⁶ The US project has to convert from a board authorization into a site, a schedule and a construction budget; until it does, the US$1.4 billion figure remains permission rather than a plan.⁷ The new chairman will preside over his first full cycle of capital-allocation decisions, including whether the withdrawn equity raise comes back. And somewhere in the next several quarters, the market will find out whether the record revenue of early 2026 was the beginning of a plateau or the top of a capex cycle.

What makes King Yuan Electronics worth understanding is not that it is a good investment or a bad one — that depends on assumptions about AI infrastructure spending that nobody, including its management, can currently verify. It is worth understanding because it is a clean read on the physical reality underneath the AI boom.

The narrative version of artificial intelligence is about models and capabilities and the companies whose logos appear on the chips. The material version is about clean-room square footage, thermal engineering, machine-hours, and the unglamorous question of whether a given piece of silicon actually works before it goes into a rack. A company that spent thirty-nine years perfecting the answer to that question now finds itself, briefly, indispensable — and has responded by making the largest and most leveraged bet in its history on the assumption that "briefly" is the wrong word.

Whether it is right will be settled the way it always is in this industry: not by the next record quarter, but by the first bad one.

---

References

1. AI Chip Testing Demand Transforms KYEC — TVBS World Taiwan ↩↩↩↩↩

2. Taiwan semiconductor firm KYEC to exit mainland China amid geopolitical tensions — South China Morning Post, 2024-04-30 ↩↩↩↩

3. Taiwan AI testing boom lifts KYEC, MPI, and WinWay to record 2026 revenue — DigiTimes, 2026-05-12 ↩

4. 京元電Q2獲利估寫同期高 Q3營運續看升 — Yahoo奇摩股市 ↩↩↩

5. King Yuan hikes capital expenditure by 37% — Taipei Times, 2025-08-09 ↩↩↩

6. Taiwan's King Yuan Electronics opens Singapore plant — Taiwan News, 2026-05-22 ↩↩↩↩↩↩

7. KYEC approves US$1.4 billion US plant to expand chip testing capacity — DigiTimes, 2026-07-13 ↩↩↩↩

8. 京元電董事長換謝其俊，李金恭意外卸任 — 自由財經 Liberty Times, 2026-05-29 ↩↩↩↩↩

9. 《時報周刊》從勘險員到封裝巨擘 李金恭熬過兩大危機 打造京元電成龍頭 — 中國時報, 2016-05-06 ↩↩↩↩

10. 京元電子 2449 公司簡介 — 鉅亨網 Anue ↩↩

11. 京元電子 — 經營團隊 (Executive Team) ↩↩

12. King Yuan Electronics — About / Company Overview ↩↩↩↩↩

13. King Yuan Electronics — Investor Relations, Quarterly Financials ↩↩↩↩↩↩↩↩↩↩↩

14. SPIL and ASE execute Joint Share Swap Agreement to establish industrial holding company — Jones Day ↩

15. ASE, Amkor Top OSAT Rankings But China Gains Ground — Mark LaPedus ↩↩↩↩

16. 【京元電子法說會重點內容備忘錄】未來展望趨勢 20251128 — Fugle 富果 ↩↩↩↩↩

17. 【公告】京元電子總經理異動 — Yahoo奇摩股市, 2023-08-08 ↩

18. 京元電股東會／通過配發每股1元現金及0.5元股票，謝其俊接任董事長 — 聯合新聞網 UDN, 2026-05-29 ↩↩

19. 京元電宣布撤回現金增資發行新股案，公司揭露主要原因 — 風傳媒 Storm Media, 2025-10-09 ↩↩↩↩

20. ASE, Powertech, KYEC CapEx May Hit NT$370B This Year as AI Drives Record OSAT Investment — TrendForce, 2026-05-05 ↩↩↩

21. 京元電大幅提高資本支出因應需求 — TechNews 科技新報, 2026-04-13 ↩↩

22. KYEC inaugurates Singapore facility, marking a new milestone in global expansion — Singapore EDB, 2026-05-22 ↩↩

23. Nvidia supplier King Yuan Electronics to invest up to $1.4 billion in US facility — Reuters via Yahoo Finance, 2026-07-10 ↩

24. 京元電子 本公司受邀參加瑞銀證券（UBS）舉辦的 Asian Investment Conference — Yahoo奇摩股市 ↩

25. 京元電子6/4~5參加野村證券舉辦之海外法說會 — Yahoo奇摩股市 ↩