Memory Stock Pullback Not a Cyclical Reversal! BofA: Meta Selling Compute ≠ Order Cuts, Korean Expansion Still Distant — DRAM & NAND Price Hikes to Extend into Next Year

Over the past week, global memory stocks have experienced a significant pullback. Markets quickly attributed the decline to three main factors: Meta potentially selling excess compute capacity (suggesting data center overbuilding), Apple evaluating the inclusion of CXMT DRAM, and the South Korean government announcing a large-scale semiconductor plan, signaling future supply oversupply.

These market-cited reasons all point to "demand peaking, supply flooding, and the end of a supercycle." However, Bank of America's recent research report, "Global Memory Tech," argues: "These risks are overestimated by the market. Cloud capex, Korean semiconductor exports, and DRAM/NAND contract prices show no directional reversal in the memory cycle."

The report states that what has truly changed is that the memory industry, after significant price hikes and stock re-rating, has entered a new phase—"fundamentals remain strong, trading difficulty has sharply increased, and the focus has returned to profit realization and stock selection."

First, Meta selling compute capacity does not equal cutting memory orders. The market fears that Meta opening its data centers to external customers implies prior over-purchasing of AI servers. However, BofA cites supply chain feedback indicating that memory chipmakers, NAND controller, and package substrate suppliers are instead observing continued strengthening of Meta's orders for HBM, LPDDR5, and enterprise SSDs.

Meta's external compute offering is more likely a move toward asset monetization and business diversification (extending into cloud services), not a sign of severe overcapacity. Meta has also reportedly committed $10 billion to build its largest data center outside the U.S. in Canada, with no change to its mid-to-long-term expansion plans, and AI infrastructure capex direction remains unchanged.

Regarding reports that Chinese memory giant CXMT may enter Apple's supply chain, BofA views this more as a bargaining chip. The likelihood of Apple adopting Chinese suppliers at scale in the short term is low. First, U.S. semiconductor restrictions on China create compliance uncertainty in the supply chain. Second, Apple requires mobile DRAM to deliver over 10Gbps transfer speeds, ~1.1V low-power design, and ECC error correction. While CXMT has submitted LPDDR5X samples, large-scale, stable fulfillment still requires validation. Third, core DRAM patents have long been monopolized by Samsung, SK Hynix, and Micron, posing potential infringement litigation risks for large-scale adoption.

Even if CXMT secures orders for entry-level models like the iPhone 18e, the actual procurement volume impact would be minimal due to limited sales volume in China. Apple is more likely to list CXMT as a potential qualified supplier to use as leverage in price negotiations later this year and next, rather than fundamentally altering the global DRAM supply-demand landscape in the short term.

Finally, South Korea's large-scale expansion plan does not equate to short-term supply oversupply. Seoul plans to invest approximately 800 trillion KRW to build a semiconductor cluster in the southwest, which the market interprets as a peak-cycle signal. BofA counters that this plan will not generate large-scale effective supply before the early 2030s. Current priorities remain expansions in Yongin and Pyeongtaek. A decade-long industrial plan cannot be equated directly with supply oversupply in the next 2–3 years. Long-term expansion is worth monitoring but is not a basis for declaring this memory cycle has peaked.

Recent BofA surveys of Japan's supply chain also show industry executives are more positive than cautious investors. Q2 2024 memory ASPs were strong, especially for NAND. Q3/Q4 ASPs are expected to exceed Q2 levels, and DRAM and NAND shortages may persist into next year. Long-term supply agreements are increasing but are primarily volume-based. Capex and wafer starts remain restrained, indicating actual expansion and customer stockpiling have not yet entered a supply-oversupply phase.

Samsung Electronics announced preliminary Q2 results on Tuesday (July 7), reporting consolidated revenue of approximately 171 trillion KRW (up 129.3% YoY) and operating profit of about 89.4 trillion KRW (up 1,810% YoY), exceeding market consensus by ~8.6 trillion KRW.

While business segment details were not disclosed, in conjunction with rising Q2 DRAM/NAND contract prices and a surge in semiconductor exports, Samsung's memory division is seen as the core profit driver. BofA's prior forecast of "memory division operating profit exceeding market expectations" is thus indirectly confirmed.

South Korea's semiconductor exports reached $44.8 billion last month, up 21% MoM and 199% YoY—marking six consecutive months of triple-digit growth. This is nearly three times the 2025 monthly average of $14 billion. The surge in export value reflects rapidly rising ASPs. The core issue remains price hikes and tight supply, not inventory buildup or clear demand contraction.

BofA estimates that DRAM ASPs will increase sequentially by 53%, 17%, and 7% from Q2 to Q4 2024. While the pace slows as the base rises, the upward trend remains intact. Spot prices for 16Gb DDR5 are around $47, and 16Gb DDR4 around $75—both above previous cycle highs. As major manufacturers continue shifting wafer capacity to HBM and server DRAM, supply of traditional DDR4/DDR5 is simultaneously contracting. DDR4, in particular, faces structural shortages as Samsung, SK Hynix, and Micron gradually phase out production, pushing contract prices to the $35–40 range. The technology premium (DDR5 being more expensive than DDR4) has nearly disappeared, reflecting that legacy process retirement is outpacing customer product transitions.

For NAND flash, BofA expects price growth to slow but absolute prices to remain high. The 512Gb NAND wafer spot price peaked in March and stabilized or slightly declined in Q2 but has risen over 50% year-to-date—about eight times the February 2025 low. Contract prices are around $25, ten times the $2.5 February 2025 low. After sharp increases in Q4 2023 and Q1 2024, Q2 2024 monthly NAND contract price increases slowed to 1–5%, indicating customers are nearing price tolerance limits and price hikes are returning to normal—not a trend reversal.

Server memory continues to hit record highs: 64GB server DRAM modules (DDR5) are around $1,400, DDR4 around $1,100—both historical highs. DDR5 contract prices rose again last month, while DDR4 remained flat. Demand for high-end memory in AI servers and data centers shows no signs of weakening.

Amazon, Microsoft, Alphabet, and Meta—the four hyperscalers—are the core new demand drivers for memory.

BofA estimates these four companies will collectively spend around $700 billion in capex in 2026, up ~80% YoY, approaching $1 trillion in the next two years, with no visible signs of downward revision. Their combined revenue is expected to grow 15–20%, with cloud business revenue growing 35–40% (AWS margin >35%, Azure >40%, Google Cloud 30–35%). As long as cloud businesses maintain high profitability, these giants will have incentives to continue expanding AI infrastructure, translating into more demand for HBM, server DRAM, and enterprise SSDs.

Unlike historical cycles driven by PC and smartphone inventory restocking, this cycle's demand structure is more complex. HBM, server DRAM, enterprise SSDs, AI inference infrastructure, hyperscaler capex, and structural shortages from DDR4 capacity exits are all overlapping. This means that observing only PC and smartphone shipments is no longer sufficient to judge the full cycle. Even if some consumer electronics demand is pressured by price hikes, AI servers can continue absorbing high-end products, keeping overall supply tight.

However, internal divergence is intensifying: HBM, server DRAM, enterprise SSDs, and advanced packaging-related products continue to benefit, while manufacturers overly reliant on consumer NAND and mobile DRAM will feel demand elasticity decline more quickly.

Since 2026, SanDisk and Kioxia have seen gains exceeding 800%, and DRAM makers have also risen significantly. With high prices, high expectations, and high gains occurring simultaneously, any noise around orders, capex, supply, or pricing can trigger sharp volatility.

BofA concludes that fundamentals remain positive, but stock prices have passed the "blind buying" phase. Future performance will depend on three factors: whether memory prices can sustain profit upgrades, whether actual profit growth can digest valuation expansion, and whether new capex remains restrained to avoid worsening long-term supply expectations. The past week's pullback may not signal a cycle peak but rather the market's attempt to "reprice for the next phase."

BofA acknowledges that Meta selling compute, CXMT entering the "Apple chain," and Korean long-term expansion are not negligible risks. But at least current data does not overturn three key facts: hyperscaler capex continues to rise, DRAM/NAND prices remain at historical highs, and new advanced capacity will take years to form effective supply.

Currently, the memory industry's fundamentals remain strong, and the era of indiscriminate re-rating is gradually ending. Investors' next question is no longer "will it keep rising?" but "which memory products are still in shortage, which companies can deliver profits, and which stocks have already