Memory Prices Hit a Plateau, But the Shortage Isn't Over
DRAM and NAND prices are still rising, but the pace of gains has slowed sharply. AI servers and HBM remain in short supply even as PCs and smartphones show clear demand destruction.
This week's takeaway: price acceleration has stalled, but peak earnings haven't been confirmed
| Checkpoint | Current read | Peak signal |
|---|---|---|
| DRAM / NAND price growth | Still rising, but the pace is decelerating fast | High |
| Supplier capex | Expanding, focused on HBM and server DRAM | Medium-high |
| Inventory | Split: low at suppliers, rising in consumer channels | Medium |
| AI vs non-AI demand | AI strong, PC and smartphone clearly weak | High |
| China market share | Already gaining meaningfully in commodity DRAM and NAND | Medium |
| Valuation | P/E looks low, but debate over 2027-28 earnings durability is widening | Medium |
Think of the memory cycle as a mountain. Right now looks less like the descent from the summit and more like the stretch where the climb flattens into a plateau, and the question is whether it holds there. That distinction matters for stock prices. Even when the industry backdrop stays good, a slowdown in the rate of improvement can hit shares before it shows up in earnings.
1. The clearest yellow light: prices are still rising, just much more slowly
Commodity DRAM price growth, quarter over quarter:
- 1Q26 actual: +93% to +98%
- 2Q26 actual: +58% to +63%
- 3Q26 forecast: +13% to +18%
The key point isn't that prices fell. It's the second derivative, the rate of increase, that has cracked. TrendForce still expects the supply shortage to continue in the third quarter, but says high prices themselves are suppressing PC and smartphone demand, and buyer resistance to further price hikes has strengthened.
NAND is moving the same way. TrendForce's forecast shows growth decelerating sharply, from +55% to +60% in the first quarter and +70% to +75% in the second, down to +10% to +15% in the third. Consumer spot markets in particular are seeing weak trading and flat prices, while server and data center contract markets remain relatively strong. Two separate cycles are already running inside what looks like a single memory market.
2. So why is it still too early to call this a "peak"
A typical memory downcycle starts with a supply glut and rising inventory. That is not what's happening now. TrendForce says DRAM supplier inventory hit a historic low in the second quarter, and additional output is being allocated to server products first. Third-quarter server DRAM contract prices are also expected to rise 13% to 18%.
In other words, consumers are buying less because prices are high, but data centers are still trying to lock in volume even at higher prices. That gap is stretching out the cycle. Gartner forecasts that high memory costs will cut 2026 PC shipments by 10.4% and smartphone shipments by 8.4%. A forecast from Counterpoint provided to Reuters puts the smartphone shipment decline as high as 13.9%. Demand destruction is already real, but the shock hasn't spread to server memory yet.
3. Why AI could stretch the memory cycle longer than past cycles
Citi's September outlook:
- 2027 DRAM supply-demand balance: -8.7% (demand +30%, supply +19%; a negative number means shortage)
- HBM bit demand growth: +62% in 2027, rising to +69% in 2028
Citi's forecast captures the core of the bull case. If AI models move toward continual learning, constantly absorbing new data rather than training once and stopping, that creates simultaneous demand for HBM, server DDR5, and enterprise SSDs, not just one product. Citi projects the DRAM shortage widening to -8.7% in 2027 and -9.7% in 2028, with the shortfall potentially persisting through 2031.
These numbers aren't a locked-in future. What matters more is that demand is broadening beyond a single product tied to GPUs. As AI inference and agentic workloads grow, CPU servers need more RDIMMs and more eSSD storage capacity too. That weakens the old rule of thumb that once smartphones roll over, memory overall follows shortly after.
4. China has already become a downside variable for commodity memory pricing
China's push up the value chain is no longer a distant story. Per Counterpoint, CXMT's global DRAM revenue share climbed 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 roughly third place worldwide, though its revenue share is lower since its mix still skews toward cheaper consumer products. That's the precise nature of the China threat. It's less a near-term risk to HBM and cutting-edge server memory profitability, and more a variable capping the long-term price ceiling for commodity DRAM and NAND.
The dangerous window is around 2028. That's when new capacity from Korean and US producers ramps up in earnest. If CXMT and YMTC's commodity share keeps growing at the same time AI demand cools more than expected, oversupply could build quickly.
5. If P/E looks low, why are stocks still afraid of a peak
Memory stocks have historically traded at their lowest P/E right when earnings are best, because the market prices in the next downcycle's earnings decline ahead of time. Micron, even after its 2026 surge, is trading at a low single-digit P/E based on recent reports.
So the real question isn't whether the multiple is 6x or 9x. It's whether 2027 earnings estimates keep getting revised up, or whether that upward revision cycle starts to stall. That's the reasoning behind Morgan Stanley's July note flagging a possible rotation out of semiconductors and into hyperscalers. Other houses, like Citi, see the shortage lasting through 2031. Right now the debate on Wall Street isn't really "bubble or not." It's closer to: how many years does this earnings run last.
Insight Times Editorial Desk
