Its market value briefly crossed $1 trillion. Analysts expect the next quarter's profit to nearly double again. The entire investment question is whether this cycle is different, and there is a serious argument on both sides.
Why the memory business has always been a trap
To understand why record results are ambiguous rather than simply good, you have to understand what kind of business this normally is.
Memory chips, DRAM and NAND, have historically been commodities. One company's DRAM is largely interchangeable with another's, which means they compete on price, and the industry runs in brutal boom-bust cycles. The pattern is mechanical and repeats for decades: demand rises, prices spike, every manufacturer earns huge margins, and those huge margins trigger a wave of capacity investment across the whole industry. That new capacity arrives all at once, roughly two years later, and floods the market. Prices collapse. Margins go negative. Everyone stops investing. Supply tightens. The cycle begins again.
The cruel feature of this pattern is that peak profitability is the most dangerous moment, not the safest, because it is precisely when the industry is building the capacity that will destroy the next cycle's pricing. A memory company posting record margins is a memory company whose competitors are all, at that same moment, deciding to build more fabs. That is why a 72% operating margin in memory cannot simply be read as good news. Historically it has been the signal that the top is near.
So the reflexive read on SK Hynix's spectacular results is caution. The question is whether the reflexive read is wrong this time, and there are genuine reasons it might be.
The case that this cycle really is different
The bull argument is not just hype, and it rests on a specific structural claim: that HBM is not a commodity the way DRAM has always been.
High-bandwidth memory, the product driving these results, stacks memory chips vertically to feed data to AI accelerators fast enough to keep them working, and SK Hynix controls roughly 58% of the global HBM market by revenue, well ahead of Samsung and Micron near 21% each. That dominance in the high-margin segment, even though SK Hynix trails Samsung in the broader DRAM market, is the whole story. The company leads exactly the part that matters for AI.
Several things distinguish HBM from ordinary memory. It is genuinely harder to make, with advanced packaging that is difficult to scale quickly, which slows the capacity flood that normally ends cycles. It is co-developed with customers years in advance, so SK Hynix secured long-term agreements rather than spot orders, locking in demand visibility that is rare in a historically boom-bust business. And it is not really interchangeable: as SK Group's chairman put it, if the HBM were swapped for another product, the AI system may not function properly, so what used to be a peripheral component has become a core component. A component designed into a customer's architecture, on a multi-year contract, with packaging competitors cannot easily replicate, behaves less like a commodity and more like a specialty product with pricing power.
If that framing holds, the normal cycle logic weakens, because the mechanism that ends memory booms, easy capacity expansion crushing prices, is exactly the thing HBM's manufacturing difficulty and design-in relationships impede. The chairman's own reported view is that the wafer shortage could persist until 2030, with expanding capacity taking four to five years. That is the structural bull case, and it is coherent.
The case that a cycle is still a cycle
The bear argument does not dispute any of that. It disputes whether it is enough.
Start with the capacity that is coming. SK Hynix is investing 19 trillion won in a new Korean plant, and Samsung and Micron are pouring money into catching up in HBM specifically because the margins are so attractive. HBM is harder to scale than commodity DRAM, but "harder" is not "impossible," and 72% margins are a ferocious incentive for competitors to solve the packaging problem. The thing that makes HBM special today is exactly the thing every competitor is now racing to replicate, and memory history says they eventually will, at which point the scarcity premium compresses.
Then there is demand concentration, which is where the risk is sharpest and where this connects to a broader worry about the AI buildout. The United States accounts for about 65% of SK Hynix's revenue, and Nvidia alone contributed around 15% in the first quarter, rising toward 24% for full-year 2025. That is a spectacular position while AI capital spending is booming, and a dangerous one if it slows. SK Hynix's fortunes are now tightly bound to a handful of American hyperscalers continuing to spend on AI infrastructure at the current pace. If that spending decelerates, whether because AI monetization disappoints or because the debt-fueled data-center buildout tightens, the demand for HBM softens quickly, and a company selling most of its output to a few customers in one end-market feels that immediately.
The honest bear case is not that HBM is a normal commodity. It is that HBM demand depends entirely on sustained, enormous AI capital spending, and that memory has never once escaped its cyclicality permanently, only delayed it. "This time is different" is the most expensive sentence in the industry's history, and it has been true for a while right up until it wasn't.
What the two sides actually disagree about
Strip it down and the bull and bear are not arguing about SK Hynix's quality, which both concede is exceptional. They are arguing about two things: how long HBM's manufacturing moat holds off competition, and how durable AI infrastructure spending turns out to be. Those are the only questions that matter, and neither is knowable now.
Notably, the bull case depends on both holding, while the bear case only needs one to break. If competitors close the HBM gap, margins compress even with strong demand. If AI spending slows, volumes fall even with the moat intact. SK Hynix wins big only if the technological lead persists and the demand persists. That asymmetry is worth weighing: the upside requires two things to go right, the downside requires only one to go wrong.
An investor buying here is making two bets whether they realize it or not: that HBM's moat lasts longer than commodity memory's moats ever have, and that AI capital spending stays elevated for years. If both hold, the current valuation may prove cheap. If either fails, the record margins are a peak to be sold, not a level to be extrapolated. This time might be different. It usually is, until it is not.