Breaking the Cycle: The "Super Bull Run" of Memory Chips (NAND/DRAM) to Extend to 2028

As of 2026, the global semiconductor memory market is undergoing an unprecedented structural shift. Unlike the "short cycles" of the past that relied on PCs and smartphones, the current demand driven by Artificial Intelligence (AI) is regarded as a turning point comparable to the "steam engine revolution." Major investment banks and industry players have recently warned that not only will DRAM and NAND Flash remain in severe shortage through 2026, but this trend is likely to continue all the way until 2028.

I. AI: A "Black Hole" Devouring Production Capacity

In traditional memory cycles, demand was driven primarily by consumer electronics and highly sensitive to macroeconomics. That pattern has now been shattered. Nomura Securities points out that over the past decade, the share of consumer-grade memory demand has plummeted from 60% to less than 30%, replaced by aggressive bets on AI infrastructure by big tech companies.

The core engine driving this shortage is the four major US Cloud Service Providers (CSPs) . Estimates suggest that their AI-related spending in 2026 has already reached $700–725 billion. This massive capital is like a "black hole," absorbing global memory capacity. The demand from AI servers is staggering — a single AI server requires roughly 8–10 times more DRAM than a traditional server, and over three times more NAND. Additionally, as the focus of AI investment shifts from "training" to "inference," edge devices such as AI PCs and AI smartphones are further amplifying memory demand.

II. Long-Term Agreements Lock In Supply, Spot Market Sees Extreme Scarcity

Extreme capacity tightness is reshaping the industry's business model. To avoid running out of ammunition in the coming years, North American CSPs are no longer buying on the spot market as before. Instead, they are proactively paying substantial deposits to the major suppliers — Samsung, SK Hynix, and Micron — signing long-term agreements (LTAs) lasting 2 to 3 years. Industry insiders describe the situation: if a customer asks for "1" unit of supply, the supplier may only deliver "one-fifth to one-third." The market's core logic has shifted from "negotiating price" to "fighting for allocation."

III. The "Time Lag" of Capacity Expansion: Distant Water Cannot Quench Near Thirst

Although memory suppliers are flush with cash, the laws of physics and construction timelines create fundamental supply bottlenecks. According to multiple research firms, memory chip capacity expansion takes considerable time:

Slow capacity release: From land acquisition and cleanroom construction to equipment installation and yield ramping, new capacity takes at least 2–3 years to come online.
Equipment bottlenecks: Lead times for key equipment range from 1 to 1.5 years.
Crowding-out effect: HBM (High Bandwidth Memory) has become a priority for suppliers. The low yields and capacity consumption of HBM are taking up significant DRAM wafer capacity, severely constricting supply of DDR5 and DDR4.

Industry forecasts indicate that even if major fabs begin expansion plans in 2026, the new supply that could truly alleviate the market crunch will not be available until late 2027 or early 2028.

IV. Price Trends: From "Cyclical" to "Structural" Increases

The supply-demand imbalance is directly reflected in prices. In early Q1 2026, Nomura aggressively raised its forecasts, predicting that Q2 prices for commodity DRAM and NAND would rise by 51% and 50% respectively — far higher than previous estimates. Suppliers are updating price lists almost every month, with quarterly increases reaching as high as 40%. Looking ahead to the second half of 2026, although the pace of increases may moderate compared to the first half, the upward trend remains firmly in place.

V. Outlook and Risks: Concerns Beneath the Feast

Despite broad optimism extending to 2028, risks remain.

First, the risk of a "demand cliff" . Jason Kau, General Manager of Silicon Motion, warns that if the AI infrastructure investment by North American CSPs slows down or reaches a phase of completion — given that such massive capital expenditure cannot be sustained indefinitely — and if no new killer applications emerge to take the baton, memory demand could face a sharp downturn.

Second, cost pass-through pressure. The high prices of memory chips have already begun to strain Android phone and PC manufacturers. This could lengthen consumer replacement cycles or erode the market share of smaller brands, while giants like Apple with strong bargaining power may further expand their dominance.

Conclusion

Between 2026 and 2028, memory chips will be in a clear "seller's market." For downstream companies, this means sustained BOM cost pressure and supply chain management challenges. For investors and the semiconductor industry chain, it represents a long-cycle opportunity driven by AI with relatively high certainty.