Is Memory Sector On The Top?

The memory sector's performance last week continued to polarize the market: bulls firmly believe HBM is ushering in a new profit cycle, while skeptics argue spot price gains for DRAM/NAND have clearly deviated from fundamentals, resembling the bubble phase seen in 2017–18.

I lean toward the opposite of the latter view: Prices are indeed overheated, but the core drivers stem from genuine structural shifts in supply and demand, not speculation.

How should I express it?

DRAM spot prices rose 5% week-over-week (16Gb DDR4 surged 9% to $37.5), hitting a 25-year high; NAND spot prices climbed even more sharply, with 1Tb/512Gb/256Gb rising 15-22%. Despite thin trading volumes, supply shortages drove up inventory replenishment demand. NVIDIA's Oct-Q revenue reached $5.7 billion (Data Center: $5.1 billion, YoY +66%), with Jan-Q guidance at $6.5 billion (YoY +65%). HBM content increases by 50% starting with Blackwell Ultra, delivering a veritable feast of orders for memory manufacturers.

Spot prices are indeed abnormal, but supply and demand provide a plausible explanation.

The surge in DRAM spot prices represents an extreme outlier within the historical long-term cycle:

• 16Gb DDR4: US$37.5, YoY +1076%

• 16Gb DDR5: US$25, YoY +424%

• 8Gb DDR4: US$13.6, YoY +774%

However, "exaggerated price increases" do not equate to "bubbles." Examining the supply side reveals that the tightness in spot supply is structural:

(1) Major manufacturers proactively reduce legacy production capacity

SK Hynix and Samsung are shifting their production focus to LPDDR5, GDDR7, and HBM, causing 8Gb/16Gb DDR4 supply to plummet to a decade low. For an AI-driven, capital-intensive memory company, significant capex investments primarily target data center upgrades and dedicated HBM chip procurement (limited cleanroom space allows only single-digit capacity growth, mainly for infrastructure or legacy fab upgrades rather than greenfield expansion). Some argue the spot market bubble could burst at any moment because Korean manufacturers are cutting legacy DDR4/DDR5 capacity to pivot toward LPDDR5/GDDR7/HBM. There is at least some truth to this. However, Kioxia's inventory clearance in Q3 (bit growth up 30% QoQ) and the slow migration to new nodes by Korean manufacturers have pushed NAND spot prices up by over 15%.

(2) OEM inventory levels are low.

Mobile phone and PC manufacturers have been compressing inventory for over a year to preserve cash flow. This year, they have begun restocking, leading to a surge in urgent orders on the demand side.

(3) The spot price itself is the "marginal buyer price."

Spot goods account for 10% of the supply shortage, yet it is precisely those 10% of buyers facing the tightest constraints who determine the price.

These three factors combined have created a situation where spot prices are rising but goods remain scarce. Spot prices are the result of supply-demand mismatches, not driven by sentiment.

HBM is the most certain variable in this cycle.

If spot price increases require explanation, then HBM certainly does not.

The DRAM capacity requirements for AI servers are growing at a non-linear rate: Blackwell Ultra: HBM usage increased by 50% compared to the previous generation. Rubin Ultra: HBM configurations may reach the 1TB level. HBM sales are projected to reach $21 billion in 2025 (+118% YoY) and $29 billion in 2026 (+40%).

HBM represents for the three major memory manufacturers:

• Profit margins have remained consistently above 60% over the long term.

• Supply shortages are expected to persist at least through 2026–27.

• $SK Hynix, Inc.(HXSCF)$ Market share >50%, and expansion of 1nm node production will further boost ASP and market share.

The storage industry's predicament over the past decade stemmed from declining ASPs and excessive competition. HBM has repositioned the industry back toward a "technology barrier → profit barrier" paradigm.

Is the price nearing its peak? Historical data suggests: no.

A common view among opponents is that DRAM unit prices have already approached their 2018 peak, and the cycle is about to top out.

However, the data does not support this conclusion:

If we consider a longer timeframe:

The current price has not even reached the historical median.

Another point overlooked by the market: Contract prices remain at US$9–10, and multiple OEMs are signing long-term contracts for 2026–28 (with a 30–50% premium).

The existence of long-term contracts indicates that the industry chain now views "high prices" as a structural trend rather than short-term fluctuations.

China's demand is strong, but its supply has not altered the global landscape.

The DeepSeek-style leap forward by Chinese manufacturers is gradually fading, with some voices questioning their actual production volumes—though I believe this criticism is unwarranted. However, the capital markets' reaction to the significant increase in 2026 capex has been overly suspicious and prone to seeing ghosts where none exist. Meanwhile, Hynix continues to advance steadily: HBM sales projected at $21/die in 2025 (YoY +118%), $29/die in 2026 (+40%), with operating margin holding above 60%; data center gross margin remains resilient at 75%.

Although China's production capacity is rapidly expanding, the memory market is only in its second year of an upward cycle (starting late 2023). Historically, prices rose for six years in the 1990s (from $5 to $16). The current unit price of $8 (Sep-25) remains well below the inflation-adjusted peak of $35. Spot prices are inflated, yet contract prices remain low (<$10). Even if spot prices drop by 50%+, contracts still hold nearly 50% upside potential.

The return of memory to the AI-driven growth narrative is the most compelling rationale for my strong optimism regarding SK Hynix. As for mitigating risks, the challenge extends beyond external supply constraints to internal factors as well. For instance, 86% of NVIDIA's revenue comes from US/European clients—perhaps we should start calling it AI memory. I wonder if I'm the first to coin this term.

The AI-driven momentum in memory is evident, with HBM orders gaining traction and margin efficiency improving. NVIDIA's HBM content increased by 50% last month, yielding noticeable results—BlackRock Ultra users saw training times shorten by over 5%. To leverage HBM as a surprise attack against traditional memory, AI capabilities remain indispensable. Rapid ramp-up and supply are crucial, but maintaining price stability amid slow supply growth demonstrates genuine technical prowess.

Has the stock price already reflected all the positive news?

Taking SK Hynix as an example: 2026E operating profit is estimated at W65tn (conservative version), with some sellers projecting W100tn. While the current 2026E P/E ratio stands at approximately 12x and the stock price has hit new highs, earnings growth is accelerating even faster. For cyclical stocks like this where "earnings acceleration > stock price growth," valuation peaks often materialize later.

The storage sector's year-to-date stock price gains align with spot market increases, yet the market still treats it as a "cyclical stock." However, considering the structural shift in High Bandwidth Memory (HBM), the critical question is no longer "when will the cycle peak?" but rather: Is the industry transitioning from cyclical characteristics to AI infrastructure attributes?

This point may not yet be fully priced into the market.

$Southern Double Long Samsung Electronics (07747)$ $Micron Technology (MU)$ $Samsung Electronics (SMSD.UK) $ $Western Digital (WDC)$ $KIOXIA HLDGS CORP(KXHCF)$