# "Our Technical Moat Is Pretty Thin": A Toolmaker's CEO on Advanced Packaging Inspection
> Statementdog EP542 interviews the CEO of 倍利科技, one of the few Taiwanese firms making its own optical metrology (AOM) tools. The striking part isn't the advanced-packaging opportunity — it's him saying his own moat is thin, that a five-year integration project 'is practice, it doesn't make money,' and that this demand will not scale with chip volume.
Published: 2026-08-04
Locale: en
Tags: statementdog, podcast-notes, semiconductors, advanced-packaging, optical-inspection, moats
> *What is difficult in the world must begin with what is easy;*
> *what is great in the world must begin with what is small.*
> —— *Tao Te Ching*, ch. 63
## What this episode is about
*Statementdog* (財報狗) is a Taiwanese investing podcast; its "industry expert" format brings in operators rather than analysts. This one features the **CEO of 倍利科技**, one of very few Taiwanese firms building its own AOM (automated optical metrology) tools, selling into wafer fabs and OSAT houses.
The nominal topic is the inspection opportunity created by advanced packaging. What stayed with me was how candidly he described his own company — candidly enough that it barely sounded like a media appearance.
**Original episode**: Statementdog EP542, "Yield Guardian of Advanced Packaging: AOM Hunting for Wafer Defects" (2026-07-30, in Mandarin).
## The notes I took
**AOI vs AOM is really the economics of full inspection versus sampling.** AOI checks every unit: high volume, cheap per look. AOM is in-process sampling at up to 150× magnification, where depth of field is a millimetre or less and someone walking past the machine blurs the image — which is why the platform needs vibration isolation. His blunt framing: everyone builds AOI because a fab needs dozens of them; nobody wanted AOM because a fab needs one or two.
**They deliberately picked the corner nobody wanted.** When they entered in 2016 the company had twenty or thirty people, and they explicitly decided not to touch AOI: "Others have done it for decades. What would we chase them with? And they probably have patents." So they took in-process control, then spent eight years grinding with customers. Advanced packaging is what changed the math — **a mistake doesn't scrap one die, it scraps the HBM stacked next to it** — and OSATs started installing precision inspection stations in volume.
**Why "round to square" makes warpage worse.** On a round wafer, internal stress travels roughly the same distance in every direction. On a square panel, the distance from centre to a corner differs from centre to an edge, so stress is uneven and it warps more. Inspection tools flatten it with vacuum chucks to hold focus. Silicon photonics is nastier still — **light paths run along the back side, so you can't touch it** — and the flattening trick is off the table.
**The X-ray dilemma, solved by borrowing from lung scans.** Stacked layers and metal shielding block visible light, so you need X-ray. But absorbed X-ray energy causes heating or phase change — **HBM that's been imaged can start leaking three or six months later, and the data is gone**. The path forward borrows from medicine: low-dose CT already images lungs at roughly one-seventeenth to one-twentieth of the old dose, and the semiconductor question is now the same one — what's the lowest dose that still shows what I need?
**The medical project deserves its own telling.** In 2018 a professor at NTU asked whether they could analyse low-dose CT lung images. The engineers said three months. Half a year later it still didn't work — **it found lesions, but with so many false positives that a single scan came back with nine hundred red dots of which one was real**. Useless to a doctor. The team was disbanded back onto semiconductors. Six months later they tried a different representation: instead of feeding the model three hundred flat slices one at a time, **stack the slices into a 3D volume and train on that**. It worked. The product went on to be the first in Taiwan with TFDA clearance and, to date, the only one of its kind with US FDA clearance, deployed at forty to fifty hospitals.
**"Kill a hundred rather than miss one" creates a second business.** Customers set inspection to over-reject, generating heaps of false positives that humans then re-screen. Their answer is **a second AI layer**: of a hundred images flagged bad, clear eighty and hand a human only twenty. **The garbage produced by the last generation of AI is the market for the next one.**
**In-situ inspection sounds elegant and is hell to build.** Customers want inspection embedded directly into process tools: less cleanroom footprint, lower cost, one less handling step. But identical model numbers **differ internally by year of manufacture, with changes the customer was never told about** — "what used to screw in doesn't screw in, what never fouled the arm now fouls the arm" — so each unit has to be measured and fitted individually. His words: **"This is practice. It doesn't make money."** Five or six years of it.
**The cold water he poured himself.** X-ray inspection needs such high resolution that its wafers-per-hour can't keep up with an inline production rate — so it is **structurally limited to sampling**. Meaning even if 3D IC volumes explode, this demand **will not scale one-for-one with chip volume**: "it does increase linearly, but not as much as total volume, because it's a sampling concept."
**On the KLA line everyone quotes, he added a layer.** KLA has said back-end process control could grow as fast as front-end wafer fab equipment. He agrees it's plausible, but points out that market size is **units × price**: the front end has fewer tools at much higher prices (a high-end machine can approach NT$1bn), the back end more units at lower prices (their tools run NT$20m-plus versus NT$70–80m for front-end). The two parameters move in opposite directions, so **you can't call the winner by looking at one of them**.
## What I took away
**1. When an operator calls his own moat thin, believe him.**
Asked about barriers to entry, he said: "What we'd consider a truly uncrossable technical gap — we think ours is pretty thin." Domestic competitors could build this given time and resources; his real advantages are **a six-or-seven-year head start, some patents, and domain know-how that simply takes time to accumulate**. He then volunteered that the strongest company in the world in this field, with formidable technical and capital moats, got caught anyway.
My assessment of the company went *up* during that passage. Operators have a strong directional bias when discussing moats — **discount the ones who say "wide," believe the ones who say "narrow,"** because admitting you have no wall buys you nothing. That asymmetry is more useful to me than any single number.
**2. His three questions for picking a battlefield are a complete discipline of refusal.**
Is the market big? Do I have an edge in it? Is there an opponent I can't beat or who'd overtake me easily? That filter made him **walk away from three things**: front-end AOI (US incumbents with decades and thousands of staff), infrared inspection (the three scarce skills it needs are already fully committed to visible light), and a customer's invitation to build silicon-photonics AI ("a hundred-person company chasing a hundred-year-old company — too hard").
This is the same discipline I try to run on stock selection: **rule out the fights you can't win first, and only then compare what's left**. The difference is that he does it with company resources, at far higher cost. Declining a customer's request is the kind of thing most people never admit to.
**3. "It's practice, it doesn't make money" belongs stapled to every hot trend.**
Embedded inspection is fashionable and sounds entirely sensible. The real cost hides somewhere nobody puts on a slide: **generational variance between nominally identical machines**. Whoever absorbs that variance does years of unpaid work.
So the next time I see "integrated," "embedded," or "single-vendor," the question is: **who absorbs the integration variance, and are they being paid for it?** If the answer is "the supplier, and no," then that trend will show up in financials much later than in the narrative.
**4. When physics caps demand growth, the theme and the P&L decouple.**
X-ray is too slow for inline, so it samples, so its demand can't scale with chip volume. That's a **physical ceiling, not an execution problem**. A theme can be entirely correct — advanced packaging is booming, inspection genuinely matters more — while the revenue elasticity is far smaller than the story implies.
Same failure mode I ran into writing about tungsten a few days ago: **being right about the theme says nothing about the path by which it reaches a company**.
**5. The leading indicator he volunteered beats any price target.**
Asked what listeners should watch, he said: watch the end customers — the fabs and OSATs — and specifically their **capex and expansion plans in advanced packaging**. An equipment vendor's next six months to three years is visible there.
That's my favourite kind of indicator: **public, checkable, with a real time lag, and named by the person being interviewed**. Watching whether your customer's customer breaks ground is a far cleaner signal than guessing at order books.
## Where my head is
The medical story stuck with me this week.
Same team, same data, same question — six months of failure. Then one change in representation: **stack three hundred flat slices into a single volume and train on that**. Done. The difference wasn't effort. It was **the dimension they were looking at the data in**.
Most of what I've been doing this past six months is that, in a different domain: taking things that were sitting around separately and reassembling them into a shape that can actually be checked. What usually blocks me isn't a shortage of data — it's that I keep looking at it along the wrong axis.
And the other reminder is more direct: **inspection itself always costs money, always slows you down, and always looks like overhead** — right up until the run that scraps the HBM next to it. Investing has the same structure. Holding cash, writing down your criteria, going back to check your answers: all slow, all indistinguishable from waste until the day you need them.
## Further reading
- The episode: Statementdog EP542 (2026-07-30), in Mandarin
- For where process control sits in the semiconductor equipment stack, KLA's earnings calls and investor day materials are public
- The two lines from the *Tao Te Ching* are my own footnote to the episode, not part of it
---
**Disclaimer**: This is a listener's reflection and general education, **not investment advice, an offer, or a solicitation**. Companies and industries mentioned come from the public episode and public sources; nothing here recommends any security or offers a price target. Investing carries risk — judge for yourself against your own circumstances, and consult a qualified professional if needed. Copyright in the original episode belongs to its producers; please go listen and support them.