Following Storage Sell-off, Focus on "Three Major Catalysts" Expected to Reverse Market Sentiment

Memory stocks underwent a period of concentrated selling in March, with seemingly sufficient reasons cited—the sustainability of cloud vendor hardware spending, the potential overestimation of AI memory demand, and the impact of Google's TurboQuant algorithm compression technology on HBM demand. These issues converged, compounded by geopolitical tensions in the Strait of Hormuz, leading to a continuous deterioration of sentiment in memory stocks.

According to the "Chasing Wind Trading Desk," Jay Kwon, an Asia-Pacific semiconductor analyst at JP Morgan, stated in a recent report that this downturn is a "divergence between sentiment and fundamentals." His assessment is that the memory industry is transitioning from a phase of "rapid infrastructure expansion" to one of "optimization and refinement of earnings quality." While short-term catalysts are lacking and the risk-reward ratio is unfavorable, mid-to-long-term valuations are considered highly attractive. Memory stocks are currently trading at a mere 1.1x expected price-to-book ratio for FY27 and a price-to-earnings ratio of 2 to 6x. He maintains overweight ratings on Samsung, SK Hynix, and Kioxia, recommending buying on dips.

The report acknowledges the real threat of TurboQuant but suggests it's not as fatal as perceived, likely leading to "optimization of system-level efficiency" rather than "reduction in memory consumption." Furthermore, foreign ownership of Samsung has fallen to a nearly ten-year low of 48.6%, and SK Hynix's is also in a low historical range at 53.1%. However, retail investors have actively increased their holdings in the first quarter of 2026, net purchasing approximately KRW 47 trillion of Korean memory stocks.

In the next one to three months, the reversal of market sentiment hinges on the realization of three key factors: whether cloud vendors' AI-related revenue and capital expenditures can continue to exceed expectations, whether HBM specifications can be revised upward (directly impacting visibility into 2027 earnings), and whether the disclosure of long-term agreement (LTA) terms from memory manufacturers can provide sufficient downside protection for profits.

TurboQuant: A Real Threat, But Perhaps Not as Fatal as Feared

The release of Google's TurboQuant technology was one of the primary triggers for the sell-off in storage stocks in March. This technology can compress the Key-Value Cache (KV Cache) in AI systems by four to six times with minimal loss of accuracy, raising market concerns about a decline in hardware demand.

JP Morgan adopts a cautious rather than pessimistic stance, offering three counterarguments:

1. Compressing KV Cache might reduce offloading demands to different storage tiers, thereby enhancing overall system efficiency.
2. Developers are likely to use the freed-up storage capacity to expand context windows, thus maintaining demand for HBM and system memory.
3. According to Jevons Paradox, cost reductions will drive broader adoption of AI, ultimately increasing overall computing power and storage demand.

Simultaneously, TurboQuant has certain limitations. The demonstrations of this technology and NVIDIA's KVCT were based on models with parameters ranging from 1.5 billion to 70 billion from 18 to 24 months ago. Its scalability to modern environments with clusters of interactive agents using multi-hundred billion or even trillion-parameter models remains highly questionable. Additionally, the reading and processing of compressed data require additional GPU computing resources.

Therefore, the impact on the memory industry is more likely to be an "optimization of system-level efficiency" rather than a "reduction in memory consumption." The logic is somewhat similar to when DeepSeek was first released; the market's initial reaction is often an overreaction.

Investor Sentiment and Fund Flows: Foreign Exodus, Retail Influx

Fund flow data reveals the current structural divergence in the market.

Foreign investors continue net selling: In the first quarter of 2026, foreign investors net sold approximately KRW 64 trillion of Korean memory stocks (Samsung Electronics: ~KRW 44 trillion, SK Hynix: ~KRW 20 trillion), a stark contrast to the net purchase of approximately KRW 7 trillion in the full year 2025. Foreign ownership in Samsung Electronics has dropped to 48.6%, the lowest level since the end of 2015 (a significant decline from the peak of 52.6% in November 2025 and much lower than the over 58% in 2019). Foreign ownership in SK Hynix stands at 53.1%, a slight increase from the low of 52.6% in February 2026, but generally on a downward trend (the peak in September 2025 was 56.3%).

Retail investors aggressively increase positions: In contrast to foreign investors, retail investors actively increased their positions in the first quarter of 2026, with net purchases totaling approximately KRW 47 trillion (Samsung Electronics: ~KRW 32 trillion, SK Hynix: ~KRW 15 trillion), a complete reversal from their net selling stance of approximately KRW 10 trillion in the fourth quarter of 2025.

Institutional investors remain stable: Korean domestic institutions largely maintained their holdings throughout 2025 and turned to significant net purchases in the first quarter of 2026, totaling approximately KRW 6 trillion (Samsung Electronics: ~KRW 1 trillion, SK Hynix: ~KRW 5 trillion).

Looking at hedge fund movements, the first half of March was primarily characterized by deleveraging. However, short interest has recently shown a marked increase—the average daily short volume for Samsung Electronics and SK Hynix has risen by 2.2 times and 1.3 times, respectively, compared to the first two weeks of March.

Three Major Catalysts: The Turning Point for Sentiment in the Next 1 to 3 Months

JP Morgan points out that during the earnings season in April and May, three catalysts are expected to reverse the current negative market narrative:

Catalyst One: Can cloud vendors' AI revenue growth support hardware capital expenditure expectations? The current market conflict lies in the projected slowdown of cloud vendor hardware capital expenditure growth from 65% in 2026 to 15% in 2027, while the overall memory market growth rate is expected to decrease from 226% in 2026 to 33% in 2027. If cloud vendors' AI-related revenues and orders exceed expectations, this "cliff-like slowdown" concern can be partially mitigated.

Catalyst Two: Updates on HBM specifications. The upward revision of HBM content in servers for 2027 and the shipment pace of ASIC chips will directly determine the depth of the memory industry's supply-demand gap in 2027. The market is currently concerned about memory average selling prices peaking, but counter-signals are emerging in the supply chain—GPU server demand may be revised upward, and the ASIC pipeline is expanding. Regarding rumors of scaled-down production of NVIDIA's Rubin Ultra GPU compute chips, the report believes the impact on HBM usage can be disregarded.

Catalyst Three: Substantive disclosure of LTA long-term contracts. Micron has already announced its first strategic customer long-term agreement, and the market awaits similar follow-ups from Samsung and SK Hynix. What investors truly want to see are clauses that protect profits at the bottom of the cycle—price floors, volume guarantees, prepayment arrangements, and breach penalty mechanisms. If these terms are sufficiently robust, the valuation system for the memory industry could be systematically re-rated.