Skip to content 
Here’s a puzzle for you.
A company reports its best quarter in history. Operating profit jumps nineteen times over last year. Revenue more than doubles. The number beats what every analyst on the street was expecting.
And the stock falls 10% in a single day.
That’s exactly what happened to Samsung Electronics this week. And if you’ve been following the AI trade even loosely, this one’s worth understanding properly, because it tells you a lot about where we are in this cycle.
Let’s get into it.
The numbers, first
Samsung’s preliminary Q2 print showed revenue of about 171 trillion won, up 129% year over year. Operating profit came in around 89.4 trillion won (roughly $58 billion), a nineteen-fold jump from the same quarter last year.
To put that in perspective, this single quarter’s profit was higher than Samsung’s entire 2025 profit. One quarter beat one full year.
The operating margin worked out to roughly 52%, compared to just 6% in Q2 last year. That’s not a typo. Samsung went from a 6% margin business to a 52% margin business in twelve months.
If you’d shown these numbers to an analyst in isolation, without telling them the stock reaction, they’d have assumed the shares rallied hard.
Instead, Samsung closed down 8%. SK Hynix fell 7%. Japan’s Kioxia dropped 11%. And the Kospi, Korea’s benchmark index which is heavily weighted toward these two chipmakers, fell as much as 8% before triggering a trading halt, eventually closing down 5%.
So what’s going on?
Why memory chips became the hottest trade on earth
To understand the fall, you first need to understand the rise.
DRAM and NAND are the memory chips that sit alongside every processor, including the AI accelerators powering data centers worldwide. Three companies, Samsung, SK Hynix, and Micron, control almost all of global production. That’s about as concentrated as an industry gets.
When AI labs and cloud providers started racing to build data centers, they didn’t just need more processors. They needed staggering amounts of memory to go with them, especially high-bandwidth memory (HBM), the specialized chips that let AI accelerators move data fast enough to keep up with the workload.
Here’s the catch. Building more memory chip capacity isn’t something you do overnight. Fabs take years to plan and build. So when demand suddenly went vertical, supply couldn’t keep pace. Classic economics kicked in. Prices went up, and they went up hard.
Citi Research noted that average selling prices for DRAM and NAND rose 44% and 53% respectively in Q2 alone, quarter over quarter. Nomura expects a further 25% and 24% rise in the current quarter. This is what happens when demand overwhelms a genuinely constrained supply chain.
To put the scale of this in perspective, spot DRAM prices have climbed nearly 700% compared to January 2025. NAND has risen a little over 300% in that same window. This isn’t a normal cyclical uptick. Research firm IDC has called it “a crisis like no other.”
Naturally, the stocks followed. Samsung’s shares had more than doubled this year before the fall, pushing its market cap past $1 trillion. SK Hynix, Samsung, and Micron had all crossed the trillion-dollar mark back in May. The Kospi itself was up more than 80% for the year, and Samsung and SK Hynix together made up close to 58-60% of the entire index. When two stocks become that dominant in a national market, everything else becomes a side character.
Why can’t chipmakers just build more, faster?
This is the obvious question, and the answer explains why this shortage has dragged on as long as it has.
Building new fab capacity takes years and billions of dollars, whether you’re Samsung, SK Hynix, or Micron. But HBM, the specialized memory that AI chips need, comes with an extra layer of difficulty on top of that.
HBM is made by stacking multiple memory dies, each thinner than a human hair, on top of each other with microscopic precision.
Manufacturers drill thousands of tiny vertical channels through the stack, fill them with copper, and bond the layers together using solder droplets as small as 10 microns.
A single defect anywhere in that stack can ruin the entire unit. That makes HBM slower to produce and far less forgiving than conventional DRAM.
The payoff for all that complexity is speed. Transferring a terabyte of data can take more than 10 seconds on a standard DDR5 chip. A single HBM3 chip does the same job roughly 10 times faster, which is exactly the kind of speed AI systems need to avoid bottlenecking on data.
So even with unlimited capital, you can’t just flip a switch and produce more HBM. It has to be built right, one microscopic layer at a time, and that’s part of why supply has struggled to keep pace with demand.
So why did great news trigger a sell-off?
This is the part that trips people up, so let’s slow down.
Markets don’t react to numbers in a vacuum. They react to numbers relative to what was already priced in. And going into this print, expectations were sky-high.
Samsung’s guidance of 89.4 trillion won did beat the consensus estimate of about 87.3 trillion won. But that’s only a 6% beat.
After a rally where the stock had doubled, investors weren’t looking for “good.” They wanted “spectacular.” A 6% beat, after a run like that, reads less like a win and more like an invitation to take profits.
There’s also a subtler point buried in the results. Samsung’s operating profit was reduced by provisions for employee bonuses, after the company agreed in May to pay its chip division staff 10.5% of operating profit as a special bonus. Some analysts pointed out that without this provision, profit could have crossed 100 trillion won. So even the beat came with an asterisk attached, which gave the skeptics something to hang on to.
Then there’s a structural worry that’s been building for a while: what happens when all this new capacity actually arrives?
It’s not just Samsung’s stock. It’s your next laptop too.
Here’s the part of this story that goes well beyond Wall Street or the Kospi. This memory shortage is quietly making everyday electronics more expensive.
Because AI customers are willing to pay a premium for guaranteed, multi-year supply, chipmakers have been redirecting capacity toward these high-margin orders and away from the conventional memory that goes into laptops, phones, and cars. That’s left everyone else fighting over a shrinking slice of supply.
HP has said memory now makes up roughly 35% of the materials cost of building a laptop, up from just 15-18% a single quarter earlier. Dell has already raised prices on servers and PCs. Counterpoint Research estimates smartphone material costs could rise 15% or more in coming quarters because of this. IDC is projecting the global smartphone market will shrink 12.9% in 2026, which would be the sharpest drop on record for the industry. Gaming console makers like Sony and Nintendo have flagged the same pressure on their own product pricing and launch timelines.
Even automakers aren’t immune. Honda has cited chip shortages for cutting production volumes in North America. Tesla’s Elon Musk has floated the idea of Tesla making its own memory chips, which tells you how seriously this is being taken outside the semiconductor industry itself.
This is really the flip side of Samsung’s blowout quarter. The profits chipmakers are booking are, in a very direct sense, coming out of the margins of every company that needs memory but isn’t an AI hyperscaler with a multi-year contract.
There’s a structural shift behind this too. A decade ago, the top 10 customers of SK Hynix and Micron accounted for roughly a fifth of their combined revenue. Today, that same top tier of customers, now dominated by AI-centric companies, accounts for closer to half. Data center demand has grown from about 32% of global DRAM consumption five years ago to around 50% today, and Bloomberg Intelligence expects AI servers alone to cross 60% of global consumption by 2030.
The oversupply question nobody can answer yet
Right now, the shortage is real. But the profits pouring in are being aggressively reinvested into new capacity, and eventually, that capacity comes online.
Industry data from TrendForce shows fab capacity is set to increase meaningfully starting in 2027. Samsung, SK Hynix, and Micron are all racing to expand at the same time, together planning to invest a combined $2 trillion to build out chip production in South Korea over the coming years.
Demand forecasts still look strong through the first half of next year, but they start tailing off toward the year end, according to TrendForce and BofA estimates. Since markets are forward-looking, some of today’s selling is really investors repricing a 2027 oversupply risk, today.
There’s also a demand-side worry that’s less about chips and more about what’s happening one layer up. Data from Bloomberg shows LLM token spending, essentially how much AI companies are actually paying to run their models, has been pulling back lately. If enterprises get better at doing more with less, and start optimizing costs instead of throwing compute at every problem, the hardware trade eventually feels it too.
JPMorgan flagged that investors are increasingly questioning whether AI memory can keep claiming an estimated 52% share of cloud providers’ capex this year, and over 70% next year. That’s a big number to sustain.
Adding to the unease, there were reports that Meta plans to sell excess computing capacity to external customers as part of a broader cloud push, a signal some read as early evidence that AI infrastructure spending might be outpacing actual usage.
It’s also worth remembering that memory has always been a boom-and-bust business, long before AI entered the picture. Revenue swings tend to be followed by matching swings in capital spending, as chipmakers scramble to add capacity in the good years and slam the brakes in the bad ones.
As recently as 2023, Micron and SK Hynix both lost billions of dollars when a pandemic-era demand boom evaporated faster than expected, leaving them stuck with a glut of capacity nobody wanted. That memory (pun intended) is exactly why chipmakers are expanding more cautiously this time around than their AI customers would probably like. Nobody wants to repeat 2023.
Is this the top, or just a breather?
Opinions are split, and reasonably so.
The bear case says the AI-memory trade has run too far, too fast, on speculative momentum rather than fundamentals, and a real correction is overdue as leverage unwinds. There’s even a chart doing the rounds comparing the semiconductor rally to the 2011 silver bubble, mapping the two price patterns on top of each other.
But the comparison has a hole in it. Silver’s rally was driven largely by a speculative narrative around a supply squeeze, with little underlying earnings power to justify it. Samsung’s rally, on the other hand, is backed by genuinely record-breaking, verifiable earnings. The capex, the demand, and the transformative nature of AI infrastructure are real, not narrative.
That’s the more constructive read: this looks less like a bubble popping and more like a very steep rally taking a breather after getting ahead of itself. As one analyst at CLSA put it, “I don’t think this story is over. It’s just a bit of a correction on the way up. Nothing goes up forever.”
That history is really the crux of the debate. Nobody yet knows for certain whether this is another familiar memory cycle that will eventually swing back down, or whether AI represents a genuinely structural shift in demand that keeps memory prices elevated for years. That single question is what every investor in this space, and every company that depends on affordable chips, is quietly trying to answer right now.
What to watch next
If you’re tracking this space, a few things matter more than the daily stock move:
- 2027 supply data. This is the year fab capacity meaningfully increases. Watch whether demand forecasts hold up alongside it.
- Long-term contracts. Whether Samsung and SK Hynix follow Micron’s lead and lock in multi-year purchasing agreements could materially change how the market values them.
- AI capex trends at hyperscalers. If token spending and infrastructure investment keep growing, the memory story holds. If it slows, memory demand slows with it.
- SK Hynix’s US listing. The company is preparing a near-$30 billion ADR offering on Nasdaq, one of the largest-ever share issuances by an Asian company. How it’s received will say a lot about how global investors are currently pricing this trade.
The bigger picture here isn’t really about one bad day for Samsung. It’s a reminder that even genuinely great earnings can’t always outrun expectations that got too far ahead of themselves. And in a market as concentrated and leveraged as Korea’s chip sector has become, the swings, in both directions, tend to be bigger than the fundamentals alone would suggest.
Note: This piece discusses securities that do not trade on major US exchanges, including Samsung Electronics and SK Hynix, which are listed in South Korea. Investors should be aware of the additional risks, including currency and liquidity risk, associated with international shares.
