Memory stocks have been hit hard. Micron, Samsung, and SK Hynix are all well off recent highs, and the sell-off is accelerating. But here's the important distinction that most retail investors miss: this is not a story about AI demand weakening. Hyperscalers are still spending at record rates. NVIDIA's order backlog is still years long. The GB200 is still shipping to the biggest buyers on earth.
The sell-off is entirely about supply — and the fear that the memory cycle, which powered one of the most violent upcycles in sector history, may be cresting. Four things are happening simultaneously, and together they have spooked the market.
1. The HBM Supply Surge — Everyone Is Ramping at Once
For the past two years, the memory trade was simple: HBM was the only memory product that mattered for AI, SK Hynix had a near-monopoly on it, yields were terrible, and supply was chronically short. That scarcity was the foundation of the entire memory bull thesis.
That foundation is now cracking — not because demand has weakened, but because all three major memory makers have simultaneously solved their HBM yield problems and are ramping hard:
- SK Hynix is scaling HBM4 production and improving HBM3E yields to record levels, meaning more usable chips per wafer.
- Samsung, which struggled with HBM3E yield issues for much of 2025, has reportedly passed NVIDIA's qualification and is entering mass production of HBM3E — adding a significant new supply source that didn't exist six months ago.
- Micron is ramping HBM3E aggressively, shipping to NVIDIA and targeting share gains from a position of near-zero exposure two years ago.
The result: the global HBM supply pool is expanding rapidly from three directions at once. Investors who were pricing in perpetual scarcity are now having to reprice for a market that could shift from shortage to adequacy within two to four quarters.
2. SK Hynix's Guidance — The Canary in the Coal Mine
SK Hynix is the world's #1 HBM supplier, with roughly 50–55% of global HBM share and NVIDIA as its anchor customer. When SK Hynix speaks about memory demand, the market listens — which is why its recent guidance language unsettled investors.
The company flagged softer near-term numbers, citing a combination of factors: inventory digestion at some customer tiers, HBM pricing pressure as Samsung re-enters the supply mix, and conventional DRAM weakness as Chinese supply from CXMT starts weighing on commodity pricing. None of these factors are catastrophic in isolation. Together, in the context of a stock that had priced in flawless execution for two years, they were enough to trigger a sharp de-rating.
The critical question is whether Hynix's guidance represents a temporary air pocket — a one or two-quarter pause before AI capex drives the next leg of HBM demand — or the beginning of a more sustained margin compression cycle. The answer depends almost entirely on what happens to HBM4 pricing as Samsung scales up.
3. CXMT's $8.6B IPO — China's Biggest Wildcard
On July 27, ChangXin Memory Technologies (CXMT) begins trading on Shanghai's STAR Market in what is set to be China's largest semiconductor IPO ever — targeting at least $8.6 billion in proceeds. For memory investors, this is the supply wildcard that keeps getting bigger.
CXMT is currently running approximately 265,000 wafer starts per month in conventional DRAM (DDR5 and LPDDR5X). The IPO proceeds fund its expansion to 350,000 wspm by year-end and 500,000 wspm by 2028 — which would represent roughly 17% of all global DRAM supply. At that scale, CXMT becomes the world's third-largest DRAM manufacturer by volume, overtaking Micron.
The near-term competitive impact is primarily on conventional DRAM pricing — not HBM. CXMT is targeting HBM3E production in 2027, which puts it 1–2 generations behind the incumbents. But conventional DRAM still accounts for meaningful revenue at all three majors, and the pricing dynamics in that segment are materially affected by CXMT's aggressive capacity additions.
| Memory Maker | CXMT Threat Level | HBM Revenue Mix | Conventional DRAM Exposure | Verdict |
|---|---|---|---|---|
| SK Hynix | Moderate | Very High (~50–55% HBM) | Low | Most insulated |
| Samsung | High | High (~35–40% HBM) | High (dominant DDR5) | Caught in the middle |
| Micron (MU) | Highest | Low-Medium (~10% HBM) | Very High | Most exposed |
4. Is This a Cycle Peak — or a Buying Opportunity?
Here is the framing that matters most: the market is not selling memory stocks because AI is weak. It is selling them because surging prices and expanding supply look like the definition of a cycle peak. That is classic memory-cycle nervousness after a massive two-year rally — and it has historically created both real tops and painful false alarms.
The case for a genuine cycle peak rests on three pillars. First, HBM supply is catching up from three directions simultaneously. Second, SK Hynix — the strongest player in the best position — is already flagging softer numbers. Third, CXMT is adding conventional DRAM supply at a scale that will weigh on commodity pricing regardless of what happens with HBM.
The case against a deep downcycle is equally compelling. AI capex from the hyperscalers is not slowing — Microsoft, Google, Meta, and Amazon have all raised capital expenditure guidance for 2026 and 2027. Every new NVIDIA chip requires more HBM, not less. GB300 requires HBM4 in quantities that current production cannot yet satisfy. And the smartphone and PC DRAM cycle is in the early stages of its own recovery, adding a demand tailwind that is entirely separate from AI.
- Hyperscaler capex still rising — AI compute demand structurally intact
- GB300 / next-gen AI chips require HBM4 in volumes not yet producible
- Smartphone and PC DRAM recovery adds a second demand driver
- HBM pricing holding up better than commodity DRAM — premium tier insulated
- Micron at lower multiples than Hynix offers best risk/reward on a recovery
- CXMT not in HBM at scale — the highest-margin product is still protected
- All three HBM makers ramping simultaneously — supply glut incoming
- SK Hynix weak guidance signals pricing pressure has already started
- Samsung re-entering HBM supply = structural ASP compression ahead
- CXMT $8.6B raise funds aggressive conventional DRAM expansion through 2028
- Memory cycles historically overshoot in both directions — the fall can be violent
- AI capex timing uncertainty — any slowdown hits HBM demand first and hardest
How to Position in This Environment
The memory sector sell-off has created differentiated opportunities that did not exist two months ago. Not all memory stocks face the same risks, and the spread between the most insulated and most exposed names has widened significantly.
SK Hynix (000660.KS) remains the highest-quality play in the sector. Its HBM dominance is structural, its customer relationships with NVIDIA are deep, and its technology lead in HBM4 is measured in years rather than quarters. The guidance miss is a short-term catalyst, not a structural shift. For long-term investors, Hynix weakness is historically a better entry point than a reason to exit.
Micron (MU) is the most complex call. It is the most exposed to CXMT competition in conventional DRAM, has the smallest HBM share, and trades at a premium that prices in a successful HBM3E ramp that is still a work in progress. But Micron has also executed extraordinarily well over the past two years and has a US-listed, geopolitically advantaged position that Samsung and Hynix cannot claim. If HBM3E ramps on schedule, the current sell-off looks like an overreaction.
Samsung (005930.KS) faces the most complex competitive landscape — CXMT below in conventional DRAM, SK Hynix above in HBM. It also has the most to gain if its HBM3E yield improvements accelerate faster than the market expects. Samsung is the highest-risk, highest-optionality play in the group.
The sell-off is rational but likely overdone on the HBM leaders
This is a supply-driven correction in a demand-intact environment — the most historically recoverable type of memory downturn. The risk is a deeper conventional DRAM price war from CXMT that bleeds into blended ASPs across all three majors. The opportunity is SK Hynix and Micron at valuations that price in a more severe downcycle than current AI capex trends justify. Watch HBM4 pricing data and CXMT yield ramp timelines — those two datapoints will determine whether the bears are right about cycle peak, or whether 2027 sets up the next leg of the memory supercycle.