What a memory supercycle is — when DRAM prices move, half of Korea's exports move
A memory supercycle is a stretch in which DRAM and NAND prices rise much further and for much longer than the ordinary two-to-three-year memory cycle allows, because demand has changed structurally and supply cannot arrive inside the same window. Memory behaves almost like a commodity — identical specifications are interchangeable, so price is set by supply and demand rather than by brand — and the industry is a three-firm oligopoly, with first-quarter 2026 DRAM share at roughly 38 percent for Samsung Electronics, 29 percent for SK hynix and 22 percent for Micron. This cycle is driven by AI datacentre buildouts that demand high-bandwidth memory and high-capacity server DRAM simultaneously, and because HBM stacks dies it consumes far more wafer area per usable bit, so making more HBM reduces the supply of ordinary DRAM at the same time. Bank of America projected 2026 global DRAM revenue up 51 percent and NAND up 45 percent. For Korea the effect is immediate: semiconductors were 46.65 billion dollars, or 47.5 percent, of August 2026 exports
The three lines
- Definition — prices rising further and longer than the normal cycle, from structural demand plus a supply lag
- Structure — memory is a near-commodity priced by supply and demand, made by three firms holding about 90 percent of DRAM
- Korea — semiconductors were 47.5 percent of August 2026 exports, so memory prices move the index, the won and the trade balance together
Key questions
- What does memory supercycle mean
- **A period in which memory prices rise much further and much longer than usual.** Splitting the phrase makes it precise. **'Cycle'** — memory prices always oscillate, and the reason is structural. Prices rise, manufacturers build fabs, output from those fabs arrives **two to three years later**, prices collapse, investment stops, inventory drains, prices rise again. **The lag between the investment decision and actual production is what creates the cycle.** **'Super'** — the amplitude and duration break out of the normal range. Two conditions are required. **① Demand has changed structurally, not temporarily.** A well-selling phone launch is temporary; every datacentre on earth being rebuilt for AI computation is structural. **② Supply cannot respond within the window.** Building a fab and reaching yield takes years, so however high prices go, **volume does not double next year.** When both hold, prices keep climbing. Earlier examples usually cited are the mid-1990s PC build-out and the 2017-2018 server and mobile expansion.
- Why is memory so much more volatile than other chips
- **Because memory is close to a pure commodity, and that single property explains the rest.** DRAM of the same specification is interchangeable between makers, so **no manufacturer can defend price through brand.** Price is set by supply and demand alone: when parts are scarce, sellers name their number; when parts are abundant, they sell below cost. Set CPUs and GPUs beside it and the contrast is clear — those are architecturally distinct, hard to substitute, and their prices do not swing nearly as far when supply shifts. Three further factors compound it. **① Three-firm oligopoly.** First-quarter 2026 DRAM share: Samsung Electronics **38 percent**, SK hynix **29 percent**, Micron **22 percent.** With nearly 90 percent between them, **a single company's capacity decision moves world prices.** **② Fixed costs dominate.** Once a fab exists, the marginal cost of one more wafer is small, so production is hard to halt in a downturn — which drives prices lower still. **③ Both directions have long lead times.** Cutting output and adding output each take one to two years to show up. **The mismatch between when price signals and when supply answers is the engine of the cycle.**
- What makes this cycle different from previous ones
- **The demand is a different kind, and HBM squeezes the supply side at the same time.** On demand: the driver is AI datacentre construction, and hyperscaler capital expenditure requires **HBM and high-capacity server DRAM simultaneously.** Real HBM demand in 2026 was projected at 4.23 billion gigabytes, up 95 percent year on year. **The supply-side variable is the more interesting one.** HBM is built by stacking DRAM dies, so it **consumes far more wafer area per usable bit** than conventional DRAM. The consequence: **making more HBM leaves less ordinary DRAM coming out of the same fab.** Demand rises on one side while supply falls on the other, which is why the price increase spread past server parts into commodity DRAM for PCs and phones. In figures: fourth-quarter 2025 commodity DRAM contract prices rose **45 to 50 percent** quarter on quarter, with a further **60 percent or more** projected for the first quarter of 2026. Bank of America compared 2026 to the 1990s boom, projecting global DRAM revenue **+51 percent** and NAND **+45 percent.** **Rankings inside HBM differ from DRAM overall** — second-quarter 2025 HBM share was SK hynix **62 percent**, Micron 21 percent, Samsung 17 percent.
- How does this reach my money
- **Through three channels, and they do not all point the same way.** **① Korea's economy — the most direct.** August 2026 semiconductor exports were **46.65 billion dollars**, up 209.0 percent year on year and **47.5 percent of all exports** — roughly the size of cars, ships, petrochemicals, steel and biotech combined. The trade balance, the growth rate and the currency all lean on that one line. **② Equities — the index tracks two stocks.** On September 7, 2026, when the KOSPI opened more than 3 percent higher, what moved were Samsung Electronics and SK hynix. It works in reverse too: on September 2 those two fell about 4 percent and the index dropped 273 points. **You may think you are watching an index; you are largely watching a memory price forecast.** **③ Your wallet — here it is bad news.** Rising memory prices raise the cost of PCs, phones and graphics cards. Device makers can absorb only so much before it reaches shelf prices. If you plan to add RAM or buy a laptop, **a supercycle is simply a cost increase.** **The same phenomenon reads as a tailwind for exports, volatility for the index and a burden for consumers.**
When "supercycle" appears in market coverage it reads as good news. **But the second half of the word is cycle.**
1. Read the two halves separately
Cycle — memory prices oscillate, and the reason is built into the industry.
| Phase | What happens | Elapsed |
|---|---|---|
| Prices rise | Inventory drains, prices climb | — |
| Investment decision | Manufacturers commit to fabs | Months |
| Supply floods | Output from those fabs arrives | 2-3 years after the decision |
| Prices collapse | Oversupply | — |
| Investment stops | Expansion halts | — |
The third row creates the cycle. A price signal is received and acted on, but the action reaches the market two to three years later. That gap is the engine.
Super — the amplitude and duration break out of the usual range. Two conditions:
- ① Demand changes structurally. A strong product launch is temporary. Every datacentre being rebuilt for AI is structural.
- ② Supply cannot arrive in time. Prices can double without volume doubling next year.
2. Why memory specifically
One sentence: memory is close to a pure commodity.
DRAM of the same specification is interchangeable between manufacturers. There is no brand premium to defend price with, so price is set by supply and demand alone. Scarce means name your number; abundant means below cost.
The comparison with CPUs and GPUs makes it obvious. Those are architecturally distinct and hard to substitute, so their prices do not swing as far when supply shifts.
The industry structure compounds it.
| Factor | Detail | Effect |
|---|---|---|
| Three-firm oligopoly | Q1 2026 DRAM: Samsung 38% · SK hynix 29% · Micron 22% | One company's capacity decision moves world prices |
| Fixed-cost dominance | Once the fab exists, the marginal wafer is cheap | Hard to stop in a downturn, so declines run deeper |
| Long lead times | Cuts and expansions both take 1-2 years to show | Response is always late |
3. HBM is the new variable, and it works on supply
The demand driver — AI datacentres — is the familiar part. The less familiar part is that HBM tightens supply at the same time.
HBM stacks DRAM dies, so it consumes far more wafer area per usable bit than conventional DRAM.
Which produces this: making more HBM means less ordinary DRAM comes out of the same fab.
Demand rises on one side while supply falls on the other, which is why this price increase did not stay in servers and spread into commodity DRAM for PCs and phones.
| Metric | Value | Type |
|---|---|---|
| Q4 2025 commodity DRAM contract price | +45-50% QoQ | Recorded |
| Q1 2026 | +60% or more | Projection |
| 2026 global DRAM revenue (BofA) | +51% | Projection |
| 2026 global NAND revenue (BofA) | +45% | Projection |
| 2026 real HBM demand | 4.23bn GB (+95%) | Projection |
Read the right-hand column. Most of the numbers that appear in supercycle coverage are forecasts rather than settled figures.
Rankings inside HBM differ from DRAM overall: on a second-quarter 2025 basis, SK hynix 62 percent, Micron 21 percent, Samsung 17 percent. The same company occupies different positions in the two markets — which matters when looking at individual stocks.
4. In Korea the price flows three ways
| Channel | How it appears | Direction |
|---|---|---|
| Exports | August 2026 semiconductor exports $46.65bn (+209.0%), 47.5% of all exports | Tailwind |
| Equities | The index follows Samsung Electronics and SK hynix. KOSPI +3% on September 7; −273 points on September 2 when those two fell ~4% | Volatility |
| Consumption | PC, phone and graphics card prices rise | Burden |
The second row deserves attention. You may believe you are watching the KOSPI; functionally you are watching a memory price forecast. The 47.5 percent export share is that structure stated plainly.
The third row is usually dropped from the story. If you are planning to add RAM or buy a laptop, a supercycle is a cost increase.
5. What to keep attached to the word
- Cycles end. A supercycle is still a cycle. Today's capacity arrives as supply eventually; only the date is unknown.
- Separate growth rates from absolute levels. Half of a +209 percent figure comes from how low the year-ago base was. The 46.65 billion dollars stands regardless.
- Forecasts get quoted as facts. Most of the table above is research-house projection. Cited without source and date, it becomes a wrong sentence within months.
- Price is not the only variable. The United States is weighing product-level semiconductor tariffs, and a large share of that 46.65 billion dollars would be directly exposed.
- Forty-seven and a half percent from one product is itself the risk. A structure that is very good when things are good is equally bad when they are not.
Sources
- Counterpoint Research — Global DRAM and HBM market share, quarterly
- SK hynix Newsroom — 2026 market outlook: a memory supercycle led by HBM
- Global Economic — Memory boom: SK hynix and Samsung head for a record 2026
- Hankyung Business — Memory supercycle to 2027, with a further 15% price increase
- KB — Is DRAM cheapest today? Chipflation and memory semiconductors explained
- Ministry of Trade, Industry and Energy — August 2026 export and import trends