When a single company's quarterly profit can fall by ninety percent and then, two years later, multiply several times over, something unusual is going on. That is the normal rhythm of the memory-chip business, and because two Korean firms sit at the centre of it, it is also the rhythm of Korea's exports, its currency, and its stock market.
If you follow global markets at all, you have seen the headlines swing from "memory glut" to "memory shortage" and back. Here is the machinery underneath them.
The short version: memory chips are a commodity built in factories that cost tens of billions of dollars and take years to complete. Supply cannot respond quickly to demand, so prices — and therefore profits — overshoot wildly in both directions. Korea is unusually exposed to that cycle because memory is its single largest export.
Memory is a commodity; logic chips are not
It helps to split the chip world in two.
Logic chips — the processors designed by Apple, Nvidia, Qualcomm and manufactured mostly by TSMC — are differentiated products. A customer designs around a specific chip and cannot swap it out casually. Pricing is negotiated, relationships are sticky.
Memory is different. DRAM (the working memory in a phone, PC or server) and NAND flash (storage) are close to interchangeable across suppliers within a given specification. Buyers qualify several vendors and buy on price. That single fact — substitutability — is what makes memory behave like oil or steel rather than like software.
Commodities with fixed short-run supply have violent prices. Memory adds a twist: the supply is not just fixed in the short run, it is lumpy. You cannot add ten percent of a fab. You build a whole one.
Why supply always arrives at the wrong time
A new memory fab is a multi-year, multi-billion-dollar commitment. The decision to build is made when prices are high and cash is pouring in. The output arrives two or three years later — often exactly when demand has cooled.
The loop looks like this:
| Phase | What happens | What it looks like in the numbers |
|---|---|---|
| Shortage | Demand outruns installed capacity | Contract prices rise fast, margins expand |
| Investment | Everyone approves new fabs at once | Capital spending jumps across the industry |
| Arrival | New capacity comes online together | Supply growth overshoots demand growth |
| Glut | Prices fall below cash cost for some players | Losses, production cuts, capex frozen |
| Reset | Frozen capex means no new supply for years | The next shortage is quietly being created |
Note the cruel symmetry: the cure for high prices is investment, and the cure for low prices is the absence of it. Because the lag is measured in years, the industry is permanently reacting to conditions that have already changed.
There is a second amplifier. Memory makers have enormous fixed costs and modest variable costs. Once a fab is built, running it at full tilt is nearly always better than idling it, even at poor prices. So supply does not gracefully taper when demand softens — it keeps coming until someone is losing serious money. That is why memory margins swing from deeply negative to extraordinarily high rather than settling anywhere in between.
Where Samsung and SK hynix sit
Korea's two memory makers are not marginal participants. In the second quarter of 2026, TrendForce data reported by Korean media put Samsung Electronics at roughly 39% of global DRAM revenue, SK hynix at about 25%, and US-based Micron at about 23% — meaning the two Korean firms together accounted for well over half the market, with essentially one non-Korean competitor of scale. (Figures as of Q2 2026; shares move every quarter.)
That concentration cuts both ways. In a shortage, pricing power is exceptional. In a glut, there is nowhere to hide: the product is a commodity and the fixed costs do not care about the business cycle.
It also means their behaviour is the cycle. When the major producers cut output, the glut ends sooner. When they all expand at once, the glut is guaranteed. Watching announced capital spending is often more informative than watching current prices.
HBM changed part of the story
High Bandwidth Memory is the one genuine break from commodity logic in recent years. HBM stacks DRAM dies vertically and connects them with thousands of vertical channels, giving the enormous data throughput that AI accelerators need. It is sold into a small number of demanding customers, qualified over long cycles, and produced with far lower yields than standard DRAM.
Three consequences follow:
- It is less commoditised. Qualification with a major AI chipmaker is a relationship, not a spot purchase. That makes revenue somewhat stickier than conventional DRAM.
- It consumes ordinary capacity. HBM uses more wafers and more process steps per usable bit. Capacity redirected to HBM is capacity not making conventional DRAM — which tightens the conventional market as a side effect.
- It reshuffled the hierarchy. HBM leadership has not simply followed overall DRAM scale; it has depended on who qualified first with the big AI customers. This is why headline DRAM share and HBM share can tell different stories about the same quarter.
HBM does not abolish the cycle. It changes which part of the factory is scarce, and it raises the stakes on being technically first rather than merely large.
Why this matters beyond the chip industry
Semiconductors are Korea's dominant export. In 2024 they made up roughly a quarter of total exports, and through 2026 AI-driven demand pushed that share well above 40% in monthly customs data — an unusually concentrated position for a large economy. (Korea Customs Service and KITA data; shares vary month to month.)
So the memory cycle transmits outward:
- To the trade balance. Korea's monthly export figures are, to a first approximation, a memory-price report with other industries attached.
- To the currency. Export earnings and foreign investor flows into the two large chip names are both cycle-sensitive, and both push on the won.
- To the index. Samsung Electronics and SK hynix are heavyweights in the KOSPI. When the cycle turns, the index often turns with it, regardless of what the rest of corporate Korea is doing.
This is also the honest explanation for a familiar pattern: Korean equities can look statistically cheap for long stretches. Part of that discount reflects governance and structure, but part of it is simply that earnings this cyclical deserve a lower multiple than steady ones. A peak-cycle price-to-earnings ratio of five is not a bargain; it is a warning that the "E" is temporary.
What to watch, structurally
Rather than tracking daily prices, the useful indicators are slow ones:
- Announced capital spending by the major producers. Large simultaneous increases plant the seeds of the next glut, roughly two to three years out.
- Inventory language in quarterly results — whether customers are drawing down stock or rebuilding it.
- The split between conventional DRAM and HBM capacity, which determines how tight the ordinary market gets.
- Korea's monthly export data, published early each month, which reads as a near-real-time proxy for global memory demand.
- Whether new entrants are scaling. Chinese producers moving up the DRAM stack would change the long-run supply picture more than any single quarter of demand.
None of these tell you what happens next week. All of them tell you where in the loop the industry probably is — which, for a business this cyclical, is the more useful question.
The memory cycle has never been repealed. It has only ever been delayed, and the delay is always paid for later.
This is general information, not investment advice.
