원문 보기
The single most important company in artificial intelligence isn't Nvidia. It isn't OpenAI. It's a memory maker in South Korea, and it's sold out through 2026. Today, for the first time, you'll be able to buy it in America. Everyone is watching the chip. The chip is starving. And only three companies on Earth can feed it.
For 30 years, owning the company that makes the memory inside every AI chip meant opening a Korean brokerage account. That ends today. On July 10th, SK Hynix lists on the American market, on the Nasdaq, under a brand new ticker, SKHY. It is the largest listing of its kind ever attempted, nearly $30 billion.
For the first time, anyone can own the memory leader directly. So, why does this matter? Why is a memory maker suddenly the most important name in AI? To see it, you have to understand memory itself. And it starts with a word that trips up almost everyone. Memory is not storage. Storage is where files live when the power is off, cheap, roomy, slow.
Memory is the working space a chip uses right now, this instant, fast, expensive, and gone the moment the power stops. On a spec sheet, that is DRAM. A processor can do a calculation in about a nanosecond. Reaching all the way out to memory takes a hundred times longer. For decades, computing got faster and faster, and memory could not keep up.
Engineers call the result the memory wall. On an AI chip, this is the whole game. To write a single word, a language model reads its entire brain out of memory. Then it does it again for the next word, and the next. Most of that time is not spent thinking, it is spent waiting for memory. So, the industry did something radical.
Instead of laying memory out flat and far away, they stacked it into a tower, 16 razor-thin layers high, and parked it millimeters from the chip. This is high-bandwidth memory or HBM. The trick is to go straight down. Thousands of microscopic copper channels are drilled through every layer of the stack, so data moves vertically through the tower instead of crawling around the edges.
The result is a river of bandwidth. One chip moving data 65 times faster than the memory in a high-end desktop. It is brutally hard to build. You are bonding 16 wafer-thin sheets of silicon into one flawless cube. A single bad layer ruins the entire stack. That difficulty is the whole reason this story has only three main characters.
SK Hynix, Samsung, Micron. Together they make almost all of the world's DRAM and every stack of HBM on the planet. There is no fourth option at scale. SK Hynix leads by a wide margin and it is Nvidia's main supplier. In 2025 it overtook Samsung in DRAM for the first time since 1992 and it makes more HBM than anyone.
Micron, the American name, is the fastest grower and every one of the three is already sold out through 2026. How replaceable are they? Barely. Memory is savagely cyclical and the crashes have killed almost everyone who ever tried. Germany's Qimonda, gone. Japan's Elpida, gone. Over 30 years more than 20 makers collapsed into three.
The survivors did not just build better, they outlasted the graveyard. Now, the obvious objection. Everyone is building their own chips. OpenAI with Broadcom, Google, Amazon, Microsoft, Meta. If they can design around Nvidia, can they design around memory, too? No. And the reason is something most people miss.
You can design your own chip because logic can be rented. You draw it and a foundry called TSMC builds it for anyone. But memory has no rental counter. The three makers design and manufacture as one because in memory the process is the product. You cannot buy it as a service at any price. And here is the twist.
Every one of those custom chips still needs HBM and buys it from the same three makers. Google's newest chip carries 192 GB of it. The custom silicon wave does not route around memory. It lines up behind it. So, what actually limits supply? Not raw materials. It is the wafers. An HBM chip eats roughly three times the silicon of ordinary memory.
So, every wafer turned into HBM starves the rest of the market. That is why memory prices are spiking. And why you cannot just make more overnight. A new factory costs $20 billion and takes three years. But that raises a bigger question. The wafers themselves, the pure silicon discs that every chip is printed on.
Where do they come from? Who makes them? And can you own that, too? You can. It is one of the most hidden corners of this entire map, and it is the next choke point we are breaking down. Follow along so you don't miss it. This is the tightest memory shortage in about 15 years. For the first time, there is even a memory index fund with a ticker DRAM that holds all three makers in one basket.
But remember the one honest warning. Memory has always been cyclical. Prices fall as hard as they rise. So, watch the chip if you want. It deserves the attention. But it is a furnace, and a furnace is only as good as its fuel line. The fuel is memory. Three companies sold out, impossible to replace.
That is the memory play. If you made it this far, you already see the market differently than most people do. And this is exactly what Stack Maps does. We take the biggest stories in the world and pull them apart layer by layer down to the companies that actually build them. Memory was just one layer.
So, follow, subscribe, and explore the map. Because the best ideas are never on the surface. They are always one layer down.
Comments 0