While AI continues to extend the memory sector's upcycle, the market's trading logic is shifting from broad-based price increases to structural differentiation. The core thesis from a recent major report is that the memory cycle remains favorable, but there is now a preference for original equipment manufacturers over module makers, and for DRAM over NAND.

A recent global technology report has updated the NAND supply-demand model. Its calculations indicate that AI-related NAND demand is projected to grow 60% year-over-year in 2027, driving a global NAND supply-demand gap of approximately 9% that year. While a shortage persists, the market is no longer uniformly tight.

More critical changes are emerging on both ends of demand. Server and AI-related demand remains robust, with long-term agreements providing downside price protection. However, consumer demand is showing signs of a price ceiling. Inventory at module makers and distributors is rising, and smartphone and PC clients are under pressure between sales volumes and margins, with actual order cuts appearing after price hikes.

This means investors need to reassess relative positions within the memory chain. DRAM is tactically favored over NAND due to better LTA terms, higher demand visibility, supply discipline constrained by EUV technology, and potential HBM4E capacity crowding. Within NAND, OEMs are preferred over module makers due to stronger profit resilience and supply control.

AI Demand Remains Key Driver, NAND Shortage Extends to 2027

AI is becoming the core source of incremental NAND demand. AI NAND demand is forecast to rise from 205 EB in 2025 to 400 EB in 2026, and further to 609 EB in 2027. AI's share of total NAND demand is also expected to increase from 18% in 2025 to 32% in 2026 and 41% in 2027.

At the aggregate level, global NAND demand is projected to increase from 1111 EB in 2025 to 1250 EB in 2026, reaching 1484 EB in 2027. Supply for the same periods is estimated at 1128 EB, 1058 EB, and 1347 EB, respectively. This corresponds to supply sufficiency rates of 2% in 2025, negative 15% in 2026, and negative 9% in 2027.

This data suggests the industry shortage is not ending quickly. Even with supply recovering to 27% year-over-year growth in 2027, AI servers, enterprise SSDs, QLC storage, and CSP inventory buffers are sufficient to absorb significant new supply.

However, this does not mean NAND pricing can rise indefinitely. AI demand and consumer demand have clearly diverged. The shortage is concentrated in server and AI-related products, while the capacity for price increases in consumer-grade products is waning.

Price Signals Diverge: Servers Strong, Consumer Segment Peaking

Channel checks indicate that 3Q26 TLC enterprise SSD-related NAND prices increased approximately 30% quarter-over-quarter, while consumer-grade NAND products saw only modest gains. On the DRAM side, server-grade product prices rose about 20% QoQ in 3Q26, while traditional DDR3 and DDR4 saw gains of 30% to 40% due to continued supply tightness and increased AI-related demand.

LTAs are altering price volatility patterns. Memory suppliers and major clients are still negotiating long-term agreements, which typically include price ceilings and floors. The floors help protect OEM profitability and valuation, while the ceilings limit the scope for further significant price increases.

Client attitudes are also diverging. Clients are more willing to pay higher prices for DRAM to secure supply, while NAND price hikes are encountering some resistance. This is related to rising margin pressure on consumer electronics clients.

Inventory is also sending warning signals. Supplier inventory remains at historically low levels, but module maker inventory has increased significantly, and distributor inventory for consumer-grade memory is also elevated. Distributors believe demand has not fundamentally deteriorated, but aggressive price hikes over the past three quarters have raised costs, dampened purchases from smaller buyers, and contracting transaction volumes have increased inventory holding pressure.

Rationale for Favoring DRAM Over NAND

The overall constructive view on the memory cycle is maintained, but tactical positioning favors DRAM. There are four key reasons. First, DRAM LTA terms are more favorable, with clients showing stronger willingness to pay to ensure supply. Second, demand visibility is higher, with AI computing and related server demand remaining the primary support. Third, supply discipline is clearer, with constraints like EUV technology and capacity limiting rapid expansion. Fourth, potential HBM4E capacity crowding may further tighten DRAM supply and demand.

However, memory stocks are still influenced by the "rate of change." If year-over-year price growth plateaus around the fourth quarter of 2026 while 2028 supply-demand dynamics remain unclear, short-term cyclical catalysts may weaken. Yet, the profit visibility afforded by LTAs could still support valuation re-ratings.

Rationale for Favoring OEMs Over Module Makers

Within NAND, OEMs are preferred over module makers, primarily due to profit resilience and supply control. In traditional cycles, module makers typically accumulate low-cost inventory at the cycle bottom and release it during the upcycle, achieving higher profit elasticity. This model also brings pronounced cyclicality; once low-cost inventory is depleted and margins peak, share prices often face pressure.

This cycle is different. If AI-driven shortages persist for three to five years via LTAs, module maker margins could be more stable than in the past. However, module makers still face three constraints: low-cost inventory will be gradually consumed within the year, consumer-grade price increases are expected to narrow in the second half of 2026, and OEMs are allocating more supply to CSP clients, limiting module makers' shipment growth in 2026 and 2027.

Furthermore, hyperscale customers tend to purchase directly from NAND OEMs or sign long-term supply agreements, which may limit module makers' long-term addressable market in enterprise SSDs and AI storage. Enterprise SSD revenue contribution for Asian SSD module makers mostly remains around 10% to 20%, insufficient to fully offset weakness in consumer SSDs.

Key Risks for Investors to Monitor

The first is a slowdown in AI capital expenditure. This represents the biggest macro risk for memory stocks. As long as AI capex does not peak imminently, memory profitability is likely to extend beyond 2027.

The second is pressure on the consumer end. Smartphone and PC clients' tolerance for further price increases is diminishing, order cuts have already appeared, and module maker and distributor inventories are rising. If consumer demand remains weak, the broad-based NAND price increase narrative would be undermined.

The third is a potential supply reversal in 2028. If new wafer capacity accelerates and supply discipline loosens, NAND could face oversupply risks. Conversely, if new AI inference SSD products enter mass production, potentially consuming about three times the capacity of regular SSDs, it could further tighten industry supply.

In conclusion, the memory cycle is not over, but the phase of indiscriminate buying is passing. AI demand continues to support high-end memory and server products, and LTAs improve profit visibility. However, consumer prices are peaking, module maker inventories are rising, and 2028 supply variables are increasing. For investors, the key for the next phase is not determining if the memory upcycle persists, but differentiating who has stronger pricing power, more stable supply access, and clearer demand visibility.