Is This AI Memory Supercycle Really Different? The Valuation Trap for Long-Term Investors
Memory is reshaping around AI through 2030. But a good industry and a good entry price are separate questions.
The 20-second read
1. What is different about this memory boom
The memory market used to track PC and smartphone sales. Makers added supply, prices fell, production cuts followed, and prices recovered.
AI data centers have changed the nature of demand. As GPU compute improves, high-bandwidth memory (HBM), which feeds data to the chip quickly, matters more. AI servers need not only HBM but also DDR5 and high-capacity NAND flash.
Gartner's August 2026 forecast has the memory market growing from $220.1 billion in 2025 to $837.3 billion in 2026 and $1.0755 trillion in 2027. That does not mean shipments are surging by the same amount. Rising selling prices explain a large share of the revenue growth.
| Global memory revenue forecast | Revenue |
|---|---|
| 2025 | $220.1 billion |
| 2026 (forecast) | $837.3 billion |
| 2027 (forecast) | $1.0755 trillion |
2. Why more HBM means tighter supply of commodity DRAM
HBM stacks several DRAM chips vertically and links them at very high speed. It takes more die area, more manufacturing steps and more packaging work than standard DRAM.
When a memory maker raises its HBM share, the same wafer input can yield fewer bits of commodity DRAM. This is known as the wafer penalty.
TrendForce cites this capacity race and delays in stabilizing yields on next-generation products in forecasting tight supply through 2027. It expects the average selling price of HBM to rise 121% in 2027 from a year earlier. That is a blended ASP forecast that includes a higher mix of expensive HBM4. It does not mean the price of the same product rises 121% across the board.
The structure favors memory makers but raises costs for PC and smartphone makers. If prices climb too far, those makers may cut memory capacity in consumer products, or buyers may delay upgrades.
3. Can SK hynix's 80% operating margin last?
SK hynix reported third-quarter 2026 revenue of 104.6155 trillion won and operating profit of 83.3338 trillion won. That is an operating margin of about 79.7%.
The company says it has secured long-term contracts for more than 50% of its 2027 to 2028 capacity. That may give better earnings visibility than in past cycles. It does not mean all future volumes and prices are fixed.
| Company | Strengths | Risk for investors to check |
|---|---|---|
| SK hynix | HBM lead, packaging technology, customer ties | Durability of high profitability |
| Samsung Electronics | DRAM production base, foundry, integrated packaging | HBM yields, customer qualification, profitability |
| Micron | Leading-edge DRAM, data-center-focused product strategy | Return on large capacity expansion |
For US investors, Micron (MU) is the main way to own the memory cycle directly. But a US listing alone does not justify a higher valuation than its rivals.
Micron reported fiscal fourth-quarter 2026 revenue of $54.23 billion. Fiscal third-quarter revenue was $41.46 billion. Investors should not confuse the company's fiscal quarters with calendar quarters.
4. What may change from 2027 to 2030
Competition is widening from HBM volume to next-generation HBM4 yields, base-die design and advanced packaging capacity.
In 2029 to 2030, CXL-based memory expansion and sharing, and research into next-generation 3D DRAM, could become important variables. The CXL 4.0 specification was released in 2025 and raises transfer speed to 128 GT/s. A published spec is not the same as large-scale commercial adoption.
If memory pooling spreads, servers can share memory more efficiently. That could create new demand for server memory. It could also reduce duplicate purchases of existing memory.
5. Long-term valuation turns on normalized earnings, not peak earnings
When a high operating margin and a low forward P/E show up together, one question comes first.
Can today's earnings hold in 2030?
Say the share price is 100 and EPS is 20. The P/E is 5x. If memory prices normalize and EPS falls to 10, the P/E becomes 10x at the same share price.
A low P/E can signal undervaluation. It can also mean the market expects earnings to fall.
| Hypothetical 2030 scenario | EPS | P/E applied | Implied price |
|---|---|---|---|
| Bull | 25 | 8x | 200 |
| Base | 15 | 7x | 105 |
| Bear | 8 | 6x | 48 |
Assumes a current share price of 100 and current EPS of 20. This is an illustrative model, not a price target, probability or return forecast for any company. Dividends excluded.
Monte Carlo simulations also depend on the assumptions for volatility and expected return. One source cited a 10,000-run simulation with a probability of gains, but it gave no input parameters or result distribution, so the figure cannot be verified.
Feeding a model the assumption that the industry is safer than before, then getting a high probability of gains, does not mathematically prove that industry risk has fallen.
FAQ
Q. If AI keeps growing, will memory prices keep rising?
- Wider AI adoption is positive for demand, but capacity additions and efficiency technologies can restrain price gains.
- Rising demand and continuously rising prices are different things.
Q. Do long-term contracts end memory downturns?
- Contracts improve volume visibility. They do not remove the risk of price adjustments, customers cutting investment or changes in contract terms.
Q. If you hold to 2030, is today's low P/E enough?
- Estimate normalized EPS that can hold across several industry conditions, not peak-cycle EPS.
- Also weigh future capital spending, free cash flow and changes in share count.
In short, memory in the AI era is becoming a bigger and more technically complex market. But a good industry does not always come with a good price. Believing in the supercycle is one thing. What you pay for its earnings is another.
- In the past, PC and smartphone sales drove the memory market.
- This time, AI data centers are lifting demand for HBM, DDR5 and high-capacity NAND together.
- HBM takes up more chip area and puts a heavier burden on processing and packaging than standard DRAM.
- If manufacturers raise their HBM share, the supply of commodity DRAM made from the same wafers can shrink. This is called the wafer penalty.
- It is a forecast that the average selling price of HBM in 2027 will be 121% higher than the year before.
- The figure also reflects a growing share of high-priced HBM4. It does not mean the price of the same product will rise 121% across the board.
- That is hard to say. The company said it has secured more than 50% of its 2027 to 2028 capacity through long-term contracts.
- However, that does not mean all future volumes and selling prices are fixed. The risk of price adjustments or changes in contract terms remains.
- Not necessarily. A low P/E can be a sign of undervaluation.
- It can also be low because the market expects earnings to fall. The key question is whether today's earnings will still hold in 2030.
What to watch
Insight Times Editorial Desk
