Storage chip profits soar, why did the Korean stock market hit the brakes first? Morgan Stanley reveals the "stock price turning point" in the super cycle

According to Odaily Financial News APP, a new research report from Wall Street financial giant Morgan Stanley, titled "Memory Chips: Differentiating Between Traditional Cycles and Supercycles," reveals that the unprecedented global AI boom is creating a stronger and longer-lasting supercycle demand for memory chips. However, the unavoidable traditional boom-bust cycles characterized by supply expansion, price changes, and profit volatility still persist.

As US earnings season is set to begin, the Nasdaq hit a record high last week. However, the Philadelphia Semiconductor Index fell by 0.41% on October 9th and remains some distance from the historic high set in June. In contrast, the Korean stock market's performance diverged sharply from the Nasdaq. Chip manufacturing giant TSMC's preliminary Q3 revenue increased by about 50%, while memory chip giant Samsung Electronics' preliminary operating profit reached approximately 107.4 trillion Korean won, up 782.5% year-on-year, which is 8.825 times higher than the same period last year. Yet, the strong growth of these two chip giants could not alleviate the valuation pressure faced by memory-concept stocks amid surging 10-year and 30-year US Treasury yields. In Asian trading on Monday, Korea’s benchmark KOSPI Index—regarded as a barometer for AI computing power—fell more than 2%, with a drop of nearly 8% since October 6th, completely diverging from the Nasdaq, the bellwether for technology stocks.

Led by Shawn Kim, a strategist team at Morgan Stanley that accurately warned of cycle deterioration and downgraded SK hynix’s rating and target price in 2024 (and precisely predicted the memory supercycles since H2 2018 and 2025), the report explains this typical divergence between the Nasdaq Composite Index and the Korean market, which is 50% comprised of SK hynix and Samsung. It attributes this to the massive AI inference-driven memory workloads, frequent HBM upgrades, and continued expansion of AI server memory demand, all lifting the profit baseline of the industry. Tight DRAM and NAND supply may persist until the end of 2027. However, high profits spurring expansion, customer digestion of existing computing power, and profit upgrades slowing down could lead share prices to turn ahead of earnings data.

The Morgan Stanley strategist team led by Shawn Kim emphasizes that the phase where stocks continuously rise and outperform the market may transition to a period of volatility, correction, or underperformance even before corporate profits peak. “Can memory stocks begin to underperform before the upside in earnings revisions peaks? We think so.” Morgan Stanley notes that the core reason is that the market trades in anticipation of slower price increases, profit revision slowdowns, and future supply expansion—meaning share prices may adjust ahead of current record profits. In referenced historical cycles, stock prices tended to turn weaker some 4–8 months before profit peaks.

Shawn Kim’s Morgan Stanley strategist team maintains a positive view that Korea’s tech sector still has investment appeal but explicitly ranks the three-tier semiconductor allocation sequence as follows: “AI GPU/TPU/ASIC-like AI compute chips first, high-performance network chips next, and finally, selected ‘memory chip plays’ as the third tier.”

Morgan Stanley says the earnings season will truly test whether these memory chip giants can continue to deliver incremental profits beyond expectations after record highs. That is, the Morgan Stanley team led by Shawn Kim first favors GPUs, ASICs, foundry, and advanced packaging—core AI infrastructure computations—and then optical interconnects and high-performance network infrastructure. Their stance on memory chips is to selectively seek structural growth and tactical investment opportunities among major manufacturers.

Memory Price Surge Continues: Supercycle Faces the Ultimate Profit Test

Undoubtedly, the unprecedented AI wave accelerating global AI data center construction has indeed changed the structure of memory chip demand. The Morgan Stanley strategist team led by Shawn Kim expects the HBM market to reach around $100 billion by 2028. Long-term supply contracts, price floors, and increased client deposits improve revenue visibility, while HBM takes up traditional DRAM capacity. Model weights, context, and caching required for inference drive the co-expansion of HBM, server DRAM, and enterprise SSDs. However, Morgan Stanley underlines that the market has not fully priced in the extraordinary profits: the chart below shows Samsung and SK hynix’s 2027 projected P/E ratios at just 3.7x and 3.5x, respectively, and Micron at 6.1x, reflecting investor concerns over future profit rollback risks.

The bottlenecks for HBM upgrades or technical iteration remain largely in yield and delivery. The report, citing industry forecasts, predicts HBM blended average price will grow 121% year-on-year in 2027, with product generational upgrades contributing. HBM4 is expected at $3–5/Gb, and HBM4e at over $5/Gb. Rubin’s ramp-up is slow; in 2026, VR200 NVL72 shipments may be less than 10,000 units, below the original plan of around 15,000. Blackwell still accounts for about 78% of Nvidia’s high-end GPU shipments. Thanks to yield and production cycle advantages, 8-layer HBM could become mainstream by 2027. Some inference KV cache shifts to host memory tiering, further supporting RDIMM and SOCAMM demand.

Additionally, memory chip price hikes remain pronounced, but end-market divergence has emerged. Morgan Stanley states that the Q4 contract price increase forecast for traditional DRAM has been revised up to 10%–15% QoQ; NAND to 15%–20% QoQ; and enterprise SSDs to 23%–28%. Cloud players are adding more server memory orders, and production capacity continues to shift to HBM, which may cause Korean PC DRAM supply to drop by over 20% by 2027, with ongoing tightness in DDR4. At the same time, smartphone and PC demand remains cautious, leading to consumer NAND inventory buildup. QLC enterprise SSDs are beginning to absorb part of the inference cache demand, but consumer spec downgrades and growing Chinese supply further differentiate profit trends between enterprise and consumer storage.

Korean Stock Funds Retreat Massively: Market Begins to Ask “Where’s the Next Growth?”

KOSPI Index fell 22% in the second half, with trading volume down 70% from the late May peak and foreign capital outflows totaling $131 billion YTD. Samsung and SK hynix account for over half the market cap, exposing the Korean market heavily to memory price and profit expectations. When even record profits at Samsung can’t lift the stock price, the focus of the capital has shifted from current profit levels to whether profits can keep accelerating.

It is reported that Rayliant has taken an “underweight” (the most negative) position on these two Korea-based global memory chip giants, while another asset management giant, Pictet, notes that short-term capital chasing growth won’t jump back in just because valuations are low.

Phillip Wool, Portfolio Manager at Rayliant Global Advisors, said: “For most investors, especially those who only entered the Korean market recently for the memory chip theme, the biggest challenge is that most of the easy profit and excess alpha from this theme has basically been earned.”

Jon Withaar, Portfolio Manager at Pictet Asset Management, said: “Especially for memory chip stocks, these have now turned into value stocks, and at current profit levels, their valuations are quite cheap. But neither retail nor short-term-focused hedge funds will aggressively chase value, nor will they buy heavily just because something’s ‘cheap.’ They are seeking growth.”

Both institutional and retail investor demand has shrunk, amplifying the adjustment: a combined KRW 55 trillion in buybacks is nearly complete for the two companies, margin balances dropped from KRW 38.6 trillion to about KRW 33 trillion, and investible funds fell from nearly KRW 140 trillion to around KRW 100 trillion. Meanwhile, the Taiwan stock market, also focused on the AI compute power theme and hosting leaders like TSMC, MediaTek, UMC, and Hon Hai in the AI compute industry chain, is up about 70% year-to-date, outperforming Korea by 23 percentage points in Q3. It appears institutions increasingly prefer the growing full-stack hardware ecosystem for AI compute in China’s market.

This capital flow echoes the latest view from Morgan Stanley’s strategists, led by Shawn Kim: a more rational investment decision is to allocate within the AI computing theme toward AI chips, network infrastructure, data center optical interconnects, high-performance CPUs, ABF/glass substrates/CCL/electronic cloth/PCB, 2.5D/3D advanced packaging, and the broader Nvidia NVL AI rack supply chain. Korean market reallocation opportunities still depend on earnings sustainability, new buying power, and whether corporate value-enhancing reforms materialize.

Morgan Stanley Reorders AI Compute Investment: AI Chips and Optical/Networking First, Memory Turns to Select Plays and Tactical Opportunities

Morgan Stanley believes that while expansion has not alleviated the shortfall, it has already entered the pricing-in phase for the stock market. The report quotes DRAM capex forecasts of $98 billion for 2026 and $150–160 billion for 2027, up 82% and around 60%, respectively. Effective new supply will not be released until H2 2027, so tightness may last through year’s end. YMTC’s share of the global NAND market is around 12%–13%, and Changxin holds about 20% of China’s HBM market—important variables for future supply. Morgan Stanley strategists say architectural upgrades are also redistributing value. Within the illustrated AI rack scenarios in the report, memory’s relative hardware cost share will drop from 29% to 19%, and multi-rack networking could lift the optical interconnect share from around 10% to 20%, supporting a larger investment allocation to networking supply chains.

Morgan Stanley’s research reveals hardware cost breakdowns for various generations and configurations of AI racks:

Memory share decreases from 29% to 19%: Comparison is between VR200 NVL72 with 54TB memory and the Rubin Ultra NVL72 configuration.

High-performance data center network infrastructure (including optical interconnect systems) share increases from 10% to 20%: Comparison is between a single-rack Rubin Ultra NVL72 configuration and a hypothetical RU NVL576 8-rack setup, which requires more extensive cross-rack interconnection.

Morgan Stanley says AI investment opportunities are expanding from compute chips to high-speed interconnects and system-level efficient data transmission architectures that determine the effectiveness of the entire cluster. More importantly, multi-rack optical interconnects now have official support: Nvidia has disclosed that Rubin Ultra NVL576 will connect eight racks—each with 72 GPUs—into a single 576-GPU NVLink domain, using optical links to solve cross-rack distance, high-speed signal transmission, and power consumption challenges.

Lumentum and other data center optics vendors have seen consistently strong share prices recently—not only because capacity for the next 2–3 years has been booked, but also as market chatter suggests Nvidia (NVDA.US) is reshaping the next generation of AI rack compute system architectures. The next-gen Rubin Ultra AI cluster may reduce per-rack HBM configuration and instead connect multiple AI racks using optical interconnect technology. In recent years, prices for HBM/DRAM memory systems and NAND storage chip products have surged. With persistent supply constraints, rising costs, and continued AI cluster expansion, the investment focus of AI compute infrastructure seems to be shifting from single-point GPU/TPU performance and HBM capacity toward overall data center architectural efficiency, with high-speed optical interconnects as potentially the key direction.

Morgan Stanley also believes low P/E ratios don’t eliminate cycle risks. The report notes that profit estimates for Asian memory companies have been revised up 432% on average this year, but upward revisions have lost breadth among the three major makers. Historically, memory stocks can weaken months ahead of earnings peaks. Morgan Stanley even says that to maintain the current stock rally narrative, DRAM contract prices would need to surge another 70% QoQ—an unlikely hurdle now. It notes that price increases and the slowing pace of price hikes can occur simultaneously, with the latter enough to weaken profit upgrades and valuation expansion. Key forward variables are long-term order fulfillment, post-expansion investment returns, and shareholder distributions.

This is also why Morgan Stanley’s stock-picking focus shifts toward those with sustained profit upgrades and increasing per-machine value. Preferred compute names from Shawn Kim’s team include Nvidia, TSMC, Broadcom, MediaTek, Alchip, Aspeed, and Montage. In networking, picks include Corning, LITE (Lumentum), Coherent, Furukawa, Fujikura, Soitec, Keysight Technologies, Accton, plus the TSMC, ASE, FOCI, AllRing, MPI, Winway, and Hon Precision CPO supply chain. On the memory side: focus on Changxin domestic products, Winbond and Macronix for legacy specs, and take Samsung, SK hynix, and Kioxia as tactical plays. For supporting segments—key hardware enablers across AI compute, networking infrastructure, and memory—preferred names span CCL/ABF, MLCC, packaging/testing, and semiconductor equipment, including Unimicron, Nanya PCB, Zhen Ding, Ibiden, Samsung Electro-Mechanics, Kaimei Electronics, Shengyi, Murata, Advantest, Tokyo Electron, Ebara, Canon, ASE, AEIS, MKS, Naura Technology, AMEC, ACMR, and ASML. Be most cautious on downstream hardware facing high storage costs, overbuilding, and valuation pressures.