Article — "Samsung Nears 80% HBM4 Yield as Google and Nvidia Test Intel Foundry": Article — "Samsung Nears 80% HBM4 Yield as Google and Nvidia Test Intel Foundry": Everyone is about to read Samsung's 80% HBM4 yield as the moment it finally caught SK hynix on process. It isn't.
SK hynix's HBM4 yield, per industry sources, has also already reached the 80% range. The two companies are effectively tied on the number that was supposed to separate them.
So the figure reshuffling the market isn't a yield at all. It's the gap between what Nvidia and the hyperscalers need and what the two Korean leaders can physically produce.
The figure reshuffling the market isn't a yield at all. It's the gap between what buyers need and what the two Korean leaders can physically produce.
Last issue, we described Google and Nvidia circling Intel's foundry as a hedge against a sold-out TSMC. The same dynamic now has a hard number attached to it—and the same logic, it turns out, governs the memory layer sitting one level up in the stack.
Start with the tell. Kim Ki-tae, who runs HBM sales at SK hynix, told investors that "HBM4 competitiveness is completed only when it encompasses not just simple specifications but also the ability to supply in large volumes with stable yields and quality."
Read that carefully. The dominant supplier is no longer arguing that it makes the best HBM. It's arguing that it can make the most of it.
That is a defense you mount when the specification war is over and the volume war has begun. When your process lead evaporates, capacity becomes the only moat left to defend.
And capacity is exactly where Samsung has aimed everything. It is reportedly expanding an end-fab at Hwaseong using idle clean rooms, building a 1c DRAM line at Pyeongtaek's P4, and laying foundations for two more plants, P5-1 and P5-2.
None of that is a process breakthrough. It's square footage—the physical ability to ship what customers are begging for.
UBS has already priced in what that buys. It reportedly expects SK hynix to hold 2026 with a 48% share of HBM bit shipments, then lose the lead next year as Samsung edges ahead 41% to 39%.
Notice what flips the ranking. Not a better chip. A second source that can finally supply at scale, in a market where the leader cannot cover demand alone.
Samsung's own target makes the ambition explicit: lift its HBM share by year-end to roughly its 38% DRAM share. The company is betting that if it can produce, buyers will qualify it—because they have no choice.
This is the counterintuitive mechanic. Scarcity, not superiority, is dismantling the monopoly. When demand runs years ahead of supply, a buyer cannot afford to punish a second source for being second-best; being available is the qualification.
Now watch the identical logic play out one layer down, in foundry.
Google has reportedly commissioned Intel's foundry to produce more than three million TPUs in 2028, against projected 2027–2028 TPU output of over six million units.
Nvidia, by contrast, is reportedly only testing—evaluating whether Intel's process fits its 2028 Feynman GPU, which packs four graphics dies into a single package. No order. A hedge.
The contrast is the point. A firm commitment and a cautious test are two responses to the same condition: a TSMC that is structurally sold out for years.
TSMC's chairman has reportedly said as much—that global chip supply will struggle to keep pace with AI demand for years, even as U.S. fabs come online. Its July sales reportedly hit a record NT$467.58 billion, up 44.7% year on year, on lines that are already fully loaded.
That is the ambient weather here, not the story. The story is what a sold-out leader does to everyone downstream of it.
It forces buyers to qualify alternatives they would otherwise ignore. Google isn't sending three million TPUs to Intel because Intel's 18A beat TSMC on paper. It's doing it because TSMC has no room, and three million TPUs have to be built somewhere.
So Samsung's HBM4 and Intel's foundry are being validated on the same currency: availability. Not a technical leap—a capacity gap the incumbents can't close before 2028, because a new fab takes more than three years from groundbreaking to wafer-out.
Which is precisely why this reordering is more fragile than it looks.
Consider the uncomfortable corollary for the challengers. The scarcity that rewards a second source also lowers the bar it had to clear. Qualification won on desperation is not the same as qualification won on merit—and buyers know the difference even when they can't act on it yet.
The moment supply catches up, the arithmetic inverts.
That moment is already dated. Large-scale new memory capacity arrives around 2028. Samsung has reportedly raised its HBM4E reliability-test yield past 70% and pulled mass production forward. New fabs everywhere come online at once.
When they do, the bar rises back to where it was. A buyer who qualified Samsung or Intel because there was no alternative can, once alternatives exist, revert to the supplier it actually preferred—and leave the second source holding capacity it built on a shortage that has ended.
Digitimes has reportedly floated the scenario: the memory crunch could tip into a glut by 2028 as chipmakers expand. Every plant Samsung is now pouring foundations for is sized to a demand curve that may not hold once the leaders catch up.
This is the trap inside the triumph. Samsung and Intel are winning share in a window defined entirely by other people's inability to supply. The validation feels like a coronation. It may prove to be a lease—one that expires the day scarcity does.
The financial market is already hedging both ways. Foreign investors reportedly net-sold 5.9183 trillion won of Korean chip stocks in a single week, from the 3rd to the 7th, with SK hynix and Samsung the two largest sellers—cutting exposure at the exact moment the share-shift narrative turned most bullish.
That contradiction is the honest read. The bit-share reversal UBS projects is real, but it is built on scarcity, and scarcity is a condition, not a competence.
SK hynix understands this better than anyone, which is why Kim Ki-tae keeps talking about "stable yields and quality" alongside volume. He is reminding buyers that when the shortage lifts, the reasons they preferred the incumbent will still be there.
The suppliers being qualified today should hope they can convert desperation into loyalty before 2028. Because the same math that handed them the door will, on schedule, ask whether they ever deserved to walk through it.