Not long ago, tidying a desk drawer, I found a stick of RAM I had pulled from an old computer. A green board about the size of a pack of gum, black chips fastened to it in a neat row. I set it on my palm and looked at it for a long while. The hottest thing in the world right now is a distant descendant of this — the semiconductors that run artificial intelligence, or more precisely, a special kind of memory called HBM that sits pressed up close against the compute chip. Take this thing in your hand and trace the chain back, and you can see how narrow the place where value pools can become.
HBM is just what it says it is: DRAM memory chips stacked upward like an apartment building — eight stories, twelve, sixteen. When flat land runs out, you build into the sky. But to stack chips, you have to join floor to floor — connect them electrically and fasten them physically. This joining process is called bonding. And this, precisely, is where the value pools in the entire AI semiconductor chain — the bottleneck.
The Bottleneck Is in Assembly
Why would the joining be the bottleneck? The international standard fixes the total height of this apartment building at 0.775 millimeters. Into that space — less than a tenth the thickness of a fingernail — the base die and all sixteen stories have to be crammed. At one point I stopped writing, took out a ruler, and peered at that width. It was the gap between two tick marks, a sliver the eye can barely take hold of. If, inside it, a chip is fastened even slightly crooked, or warps in the heat, the whole expensive semiconductor becomes a reject. So the true bottleneck of the great chain called artificial intelligence pools, unexpectedly, not in the blueprints or the wafers but in the very last step: assembly.
The machine that performs this assembly is called a TC bonder, and in the market for these machines one company's name looms overwhelmingly large. Hanmi Semiconductor.
| Hanmi Semiconductor (TC bonders for HBM) | Figure |
|---|
| Global market share | about 71% (2025) |
| Exclusive supply to SK hynix | 100% for 8 years |
| 2025 revenue | highest in company history |
| Bonder market size | around $700 million (2025) → $1.5–2 billion (2028, industry estimate) |
The Paradox of the Small Wall
Here comes the most interesting turn in this chain. Hanmi Semiconductor holds 70% of the world's bonding equipment. And yet the company has exactly three customers. Because only three companies in the world make HBM: SK hynix, Samsung, Micron. To hold a link that rules the world, and have three customers to sell to. I call this the paradox of the small wall. The wall is the highest on earth, but so few customers pass in front of it that the power to name a price comes to only half of what it should.
In 2025, Hanmi Semiconductor actually tried swinging that half-measure of power. When SK hynix, saying it would diversify its supply chain, began placing orders with another equipment maker as well, Hanmi notified it of a steep increase in delivery prices. But in a market with only three customers, the card of naming a hard price only pushed that customer further toward the competition. Even where value pools at a single link, it turned out, the one holding the link cannot name whatever price he likes.
What Does the Wall Lean On?
What's more, the very ground this wall stands on is shifting. To stack the floors any higher, the current way of joining runs up against its limits. So the industry is trying to move to an entirely different bonding technology that fuses copper directly to copper — and in that field, European and American equipment makers are ahead. The wall Hanmi Semiconductor holds leans on today's assembly method, and on this structure of three customers. Both are things that could someday change.
Even so, what this chain tells us is clear. On the dinner-table chain, value scattered like water; here it pools with frightening force at a single narrow link called assembly. When we trace a chain back, what we should be looking for is not the biggest company or the most famous one. It is the link so narrow that no one can take its place. That is always where the value pools. Only — you must also count how many customers are standing in front of that link.
The People Who Don't Leave
| Company | Employees | Average pay | Average tenure | Revenue per employee | Operating profit per employee |
|---|
| Hanmi Semiconductor | 737 | 65.3 million won | 9.7 years | about 780 million won | about 340 million won |
Based on FY2025 disclosures. Consolidated and separate figures are nearly identical for this company, so these are consolidated results divided by headcount.
At the end of this chain, I count the people. 737. That is the size of the company holding 70% of the world's bonding equipment. A headcount about the size of an ordinary high school's student body has its hands on the throat of the great chain called artificial intelligence. And in one year, those 737 people produced 340 million won of operating profit apiece, while their average pay was 65.3 million won. A gap of more than five times. At a company holding a narrow link, the place where the value pools is the company — not the pay envelopes of the people who actually tighten that link. The two numbers state this plainly, without raising their voice. Of course, as the 9.7 years of tenure attest, people do not leave easily. But what expression to wear while reading this gap — I have not yet decided.
Semiconductors are not the only chain where artificial intelligence has made value pool in one place. For those chips to run day and night takes staggering amounts of electricity, and in the chain that makes and carries that electricity, something similar is happening right now. The next journey turns to how the electricity chain, so long a supporting actor, suddenly took the lead.
This piece is part of the "Korea Value Chain" series and is an original work. It is an analysis based on public data and industry sources; some figures and assessments are estimates. It is not intended as a direct basis for investment decisions.
By Alexandro Lee · July 2026
댓글
댓글 쓰기