Memory Supercycle: Has It Peaked, or Just Hit a Plateau?
Memory prices are still rising, but the pace of increase has slowed sharply. PCs and phones are seeing demand destruction while AI servers and HBM remain undersupplied.
This week's call: "Price acceleration has cooled, but the earnings peak hasn't been confirmed"
| Checkpoint | Current read | Peak signal |
|---|---|---|
| DRAM / NAND price growth | Still rising, but the pace is decelerating sharply | High |
| Supplier capex | Expanding, concentrated in HBM and server DRAM | Mid-to-high |
| Inventory | Split: suppliers running lean, consumer channels building up | Mid |
| AI vs. non-AI demand | AI strong, PC and smartphone demand clearly weak | High |
| China's market share | Already meaningfully higher in commodity DRAM and NAND | Mid |
| Valuation | P/E looks low, but debate over 2027-28 earnings durability is widening | Mid |
Think of the memory cycle as a mountain. Right now looks less like the descent from the summit and more like the stretch where a steep climb ends and you check whether it's turning into a plateau. That distinction matters for stock prices: even if the industry is doing well, if the rate of improvement slows, share prices can wobble before earnings do.
1. The clearest yellow light: prices are still rising, just much more slowly
| Metric | Value |
|---|---|
| Commodity DRAM, Q1 2026 actual | +93-98% QoQ |
| Commodity DRAM, Q2 2026 | +58-63% QoQ |
| Commodity DRAM, Q3 2026 forecast | +13-18% QoQ |
The key point isn't that prices fell. The second derivative, the acceleration of the increase, has turned down. TrendForce still expects the supply shortage to continue into the third quarter, but says elevated prices themselves are suppressing PC and smartphone demand, and buyer resistance to price has strengthened.
NAND is moving the same direction. TrendForce's forecast decelerated from +55-60% in the first quarter and +70-75% in the second quarter to just +10-15% in the third. Notably, the consumer spot market is seeing sluggish trading and flat prices, while the server and data center contract market remains relatively strong. Two separate cycles are already running inside what looks like one memory market.
2. So why is it still too early to call this a "peak"
Typically, a memory downcycle begins when supply floods the market and inventories pile up. What's happening now looks different. TrendForce assessed second-quarter DRAM supplier inventory as being at a historic low, and said additional output is being prioritized for servers. Server DRAM contract prices are also forecast to rise 13-18% in the third quarter.
In other words, consumers are buying less because prices are high, but data centers are still trying to secure volume even at high prices. That gap is what's extending the cycle. Gartner forecasts 2026 PC shipments will fall 10.4% and smartphone shipments 8.4% because of high memory costs. Counterpoint's forecast, provided to Reuters, puts the smartphone shipment decline as steep as 13.9%. Demand destruction is already real, but that shock hasn't yet spread to server memory.
3. Why AI could make this memory cycle longer than past ones
| Metric | Value |
|---|---|
| Citi 2027 DRAM supply-demand balance | -8.7% (demand +30%, supply +19% forecast; negative means shortage) |
| Citi HBM bit demand growth | +62% (2027) to +69% (2028), year over year |
Citi's September outlook captures the core of the bull case. If AI shifts from a train-once model to continual learning, where systems keep absorbing new data, that would require not just HBM but server DDR5 and enterprise SSDs simultaneously. Citi projects the DRAM supply shortfall widening to -8.7% in 2027 and -9.7% in 2028, with the shortage potentially persisting through 2031.
These numbers aren't a settled future. But the important part is that demand is broadening beyond dependence on any single type of GPU. As AI inference and agentic workloads grow, CPU servers need more RDIMM and enterprise SSDs for bulk storage as well. That's weakening the old rule of thumb that "smartphones have turned down, so all of memory will turn down soon."
4. China has already become a downward pressure on commodity memory prices
China's catch-up is no longer a distant story. By Counterpoint's numbers, CXMT's global DRAM revenue share rose from 4% in the second quarter of 2025 to 10% in the second quarter of 2026. In the same quarter, Samsung held 38%, SK hynix 25%, and Micron about 24%.
In NAND, YMTC hit 14% of global bit shipments in the second quarter of 2026, putting it in the world's top three by that measure, though its revenue share remains lower because its mix still skews toward low-cost consumer products. That's the precise nature of the China threat: it's less a variable that immediately undercuts profitability in HBM and cutting-edge server memory, and more a force pressing down the long-term price ceiling for commodity DRAM and NAND.
The dangerous window is around 2028. If Korean and US producers' new capacity ramps up at the same time CXMT and YMTC keep expanding commodity share, and AI demand slows more than expected, oversupply could build quickly.
5. If P/E ratios are low, why are stocks still worried about a peak
Memory stocks have historically tended to trade at their lowest P/E multiples right when earnings are strongest, because the market front-runs the earnings decline of the next downcycle. Micron, for instance, is trading at a low single-digit P/E by recent reports, even after its 2026 surge.
So the real question isn't "6x or 9x earnings." It's whether 2027 earnings estimates keep rising, or whether the upward revisions start to stall. That's why Morgan Stanley raised the possibility in July of some rotation out of semiconductors and into hyperscalers. Meanwhile, other houses like Citi see the supply shortage lasting through 2031. Right now, the debate on Wall Street is less "bubble or not" and more how many years the earnings run can last.
Insight Times Editorial Desk
