Memory Prices Are Still Rising, but Chip Stocks Have Plunged. Is the Supercycle Over?
Memory prices keep climbing and chipmakers are posting record results, yet the stocks have already fallen hard. The market is looking past the boom to the next wave of supply and to what customers can afford to spend.
Memory prices are still rising, and chipmakers are posting record results. Yet the stocks have already fallen hard. The market has started to worry less about the current boom and more about what comes next: more supply, and whether customers can keep paying for it.
The 20-second briefing
- The upcycle is still running. TrendForce projects fourth-quarter 2026 contract prices up 10% to 15% for conventional DRAM and 15% to 20% for NAND, quarter over quarter. This is a slowdown in the pace of increases, not a price decline.
- Cracks appear first in consumer markets. Memory for AI servers is strong, but some NAND spot prices have fallen. The boom is not hitting every product equally.
- The real test is 2028. Long-term contracts and HBM demand are positives. But capacity additions, Chinese supply growth and big tech cash flow will decide the next cycle.
1. Prices are rising. Why worry about a peak?
A memory cycle is usually questioned when price gains start to fade, not when prices hit their highest level.
DRAM price growth slowed from the 90% range in the first quarter of 2026 to the 50% to 60% range in the second. TrendForce's fourth-quarter 2026 outlook calls for conventional DRAM contract prices to rise 10% to 15% and NAND contract prices 15% to 20%. These are quarter-over-quarter forecasts, not confirmed prices.
Contract prices have not fallen. But consumer NAND is behaving differently. The spot price of 512Gb TLC wafers fell 9%, and some MLC products stopped a 21-month run of increases.
Spot prices are short-term trading prices. Contract prices are what customers and suppliers agree on. A drop in spot prices alone cannot declare the cycle over, but whether the weakness spreads to contract prices needs to be checked.
2. AI is strong, but consumer markets are different
Treating memory as a single market makes it easy to miss what is happening now.
Sandisk's data center revenue in the first quarter of fiscal 2027 rose 61% from the prior quarter, while consumer revenue fell 29%.
Intel also warned of weak PC demand and high memory prices. Its fourth-quarter 2026 revenue guidance was $15 billion to $16 billion.
AI data centers need more high-performance memory. PC and smartphone makers, meanwhile, struggle to pass higher component costs on to consumers.
The more suppliers steer capacity toward high-margin HBM and server products, the tighter consumer memory supply becomes. At the same time, high prices are weighing on consumer demand.
A price that is good for memory suppliers is not good for every customer. That is a defining feature of this cycle.
3. Long-term contracts are a safety net, not a permanent shield
SK hynix said it has signed long-term supply agreements (LTAs) with 14 major customers, securing more than 50% of its 2027 and 2028 capacity under contract.
That is a positive change from past memory cycles. When customers lock in future volume, suppliers gain demand visibility and less exposure to short-term order swings.
SK hynix expects HBM4 to overtake HBM3E in both shipments and revenue in the fourth quarter.
But long-term contracts do not guarantee every price. Public information does not confirm that every contract carries take-or-pay terms or price floors.
Competition is also intensifying. By Counterpoint's measure, Samsung Electronics' share of HBM revenue rose from 21% in the first quarter of 2026 to 33% in the second. Over the same period, SK hynix's share fell from 58% to 50%.
Even if the HBM market grows, pricing power among suppliers can shift.
4. The next risk may be supply growth, not a demand collapse
Memory makers are spending heavily on capacity.
Samsung Electronics' 2026 capital spending is forecast at 84.4 trillion won. Micron's net capital spending in fiscal 2026 was $27.37 billion, and SK hynix has signaled higher investment in 2027.
Capital spending does not turn into supply right away. Building fabs and installing equipment takes time.
But if new capacity and yield improvements arrive together from the second half of 2027 into 2028, the picture could change.
Chinese producers are another variable. According to Counterpoint, CXMT's share of global DRAM revenue rose from 4% in the second quarter of 2025 to 10% in the second quarter of 2026. YMTC ranked third in the world by NAND bit shipments in the second quarter of 2026.
That does not mean Chinese companies can replace leading-edge HBM now. But more supply of conventional memory could pressure global prices.
What matters is not the calendar year 2028. It is the moment actual bit supply growth exceeds demand growth.
5. The cash flow of the big tech buyers now matters
Understanding this cycle means looking at the buyers' finances, not just the suppliers' results.
Microsoft recorded free cash flow (FCF) of about $18.6 billion in the latest quarter. Alphabet's third-quarter FCF was a deficit of about $9.1 billion, and Amazon's trailing 12-month FCF was a deficit of about $42.4 billion.
The three companies have different reporting periods and business structures, so this is not a simple profitability ranking.
But the question is clear. Can the companies building AI infrastructure keep investing at the same pace? Can that investment generate enough cash?
Even if AI service revenue grows, capacity to keep investing may weaken if capital spending and depreciation grow faster.
Strong orders for memory suppliers are one thing. Secured long-term returns on their customers' investment are another.
6. A falling stock price does not mean the cycle is over
On October 30, the Philadelphia Semiconductor Index fell 7.2%. NVIDIA dropped 4.2%, AMD 10.9% and Intel 28.6%.
SK hynix shares at the end of October were around 1.42 million won, more than 50% below their June high. Samsung Electronics was around 220,000 won, down roughly 40% from its peak.
Share prices do not reflect only current earnings. Investors weigh future profits and the multiple applied to them.
Memory makers trade at a forward P/E of 2.5 to 3 times 2027 expected earnings, or 5 to 7 times current earnings. The two use different earnings dates, so they cannot be compared directly or averaged.
Memory stocks also tend to look cheap on P/E as earnings approach a cycle peak.
So a big price drop alone neither confirms undervaluation nor declares the cycle over.
Wall Street views also need to be read by subject and timing. Many, including BofA and UBS, see memory stocks as undervalued and the cycle as potentially extendable, while some analysts warn of late-cycle risk from supply growth. On October 30, by contrast, a chip selloff and worries about returns on AI capital spending soured sentiment in the US chip market. The two views differ in subject and timing.
7. Warning conditions for long-term investors
| Indicator | Warning condition |
|---|---|
| DRAM and NAND contract prices | Two straight quarters of decline |
| HBM pricing and long-term contracts | Weaker price floors, contract cancellations |
| Chinese memory supply | Rising share alongside falling prices |
| Big tech cash flow | Widening FCF deficits, surging borrowing |
| New semiconductor equipment orders | Consecutive declines |
These conditions are not automatic sell signals. They are benchmarks for rechecking whether the investment thesis still holds.
Unleveraged semiconductor ETFs such as SOXX and SMH can serve as long-term core holdings. New money can be staged, and overlap between individual chip stocks and ETFs is worth checking.
Investors should be wary of raising purchases too aggressively just because prices fell. A pullback in which the earnings outlook holds differs from one in which future profits deteriorate.
Insight Times Editorial Desk
How we use AI
이 기사는 사람 편집자가 기획하고, 사람과 AI가 같이 초안을 쓰고 수정하고, 최종 사람 편집자가 사실과 표현을 확인해 발행했습니다. 수치와 인용은 본문에 적힌 자료에서 가져왔습니다.
