Wall Street financial giant Morgan Stanley's latest research report titled "Memory Chips: Distinguishing Traditional Cycles from Supercycles" shows that the unprecedented global AI boom is creating a memory chip supercycle with stronger demand and longer duration, but the inevitable traditional boom-bust cycle patterns of supply expansion, price changes, and earnings volatility still exist.

With US stock earnings season about to begin, the Nasdaq hit a record high last week, but the Philadelphia Semiconductor Index still fell 0.41% on October 9 and remains some distance from its all-time high set in June. South Korea's stock market has moved in a completely different direction from the Nasdaq. Chip manufacturing giant TSMC's third-quarter preliminary revenue grew about 50%, and memory chip giant Samsung Electronics' preliminary operating profit reached approximately 107.4 trillion Korean won, a year-over-year increase of 782.5%, equivalent to 8.825 times the same period last year. However, the strong growth results of these two chip giants failed to eliminate the valuation pressure continuously faced by memory chip themed stocks amid surging 10-year and 30-year US Treasury yields.

In Monday's Asian session, the benchmark KOSPI Composite Index of South Korea's stock market, known as the "AI computing power bellwether," fell more than 2%, down nearly 8% since October 6, completely diverging from the tech stock bellwether Nasdaq. Morgan Stanley strategist Shawn Kim, who accurately warned of cycle deterioration in 2024 and downgraded SK Hynix's rating and target price, and accurately predicted the memory supercycle since 2018 and the second half of 2025, led a strategy team that explained this typical divergence between the Nasdaq Composite Index and South Korea's stock market, where SK Hynix and Samsung account for 50% of the weight. Namely, massive memory workloads brought by AI inference, frequent HBM updates and iterations, and continuously expanding AI server memory demand are raising the industry's profit center, with DRAM and NAND supply tightness potentially lasting until the end of 2027. However, high profits stimulating capacity expansion, customers digesting existing computing power, and slowing earnings revisions could all cause stock prices to turn before operating data does.

The strategy team led by Shawn Kim emphasized that the phase where stocks shift from continuous gains and outperforming the market to consolidation, pullback, or underperformance may come earlier than corporate earnings peak. "Can memory stocks start underperforming before earnings revisions peak? We believe they can," Morgan Stanley stated. The core reason behind this is that the market is trading ahead of slowing price increases, decelerating earnings revisions, and future supply increases. Even if current profits are still hitting new highs, stock prices may adjust first. In the historical cycles cited in the report, stock price weakening led earnings peaks by approximately 4 to 8 months.

The Morgan Stanley strategy team led by Shawn Kim maintained a positive judgment that the South Korean tech industry "still has investment appeal," but clearly ranked the three-tier semiconductor allocation order as "AI GPU/TPU/ASIC type AI computing chips first, high-performance networking chips second, and the third-tier semiconductor allocation level is 'selected memory chips.'" Morgan Stanley stated that what the upcoming earnings season truly needs to verify is whether these memory chip giants can continue to generate incremental results exceeding expectations after their record profits. That is, the Morgan Stanley strategy team led by Shawn Kim first favors GPU, ASIC, wafer foundry, advanced packaging, and other core AI infrastructure computing industry chains, secondly favors optical interconnect and high-performance networking infrastructure supply chains, and adopts a selectively curated allocation stance on memory chips, focusing on finding structural growth and tactical investment opportunities in major original manufacturers.

Memory Price Increases Continue, Supercycle Enters Profit Realization Examination

Undoubtedly, the unprecedented global AI boom of accelerating AI data center construction has indeed changed the demand structure for memory chips. The Morgan Stanley strategist team led by Shawn Kim stated that the HBM market is expected to expand to approximately $100 billion by 2028. Multi-year supply agreements, price floors, and some customer deposits increase revenue visibility, while HBM also crowds out traditional DRAM capacity. Model weights, context, and caching required for inference drive the joint expansion of HBM, server DRAM, and enterprise SSDs. However, Morgan Stanley emphasized that the market has not given a complete valuation premium to extraordinary profits. The charts below show that Samsung Electronics and SK Hynix's 2027 expected P/E ratios are only 3.7 times and 3.5 times respectively, while Micron's is 6.1 times, reflecting that investors are still discounting future earnings decline risks.

The bottleneck in HBM upgrades or technological iterations remains concentrated in yield and delivery. The report cites industry forecasts that HBM blended average prices will increase 121% year-over-year in 2027, including contributions from product generation upgrades. HBM4 is priced at $3 to $5 per Gb, and HBM4e at over $5 per Gb. Rubin's ramp-up is slow, with VR200 NVL72 potentially fewer than 10,000 units in 2026, below the originally planned approximately 15,000 units. Blackwell still accounts for approximately 78% of Nvidia's high-end GPU shipments. With advantages in yield and production cycles, 8-layer HBM may become mainstream in 2027. Some inference KV caches are shifting to host memory tiering, further supporting RDIMM and SOCAMM demand.

Additionally, memory chip price increases remain strong, but end-market divergence has already appeared. Morgan Stanley stated that fourth-quarter traditional DRAM contract price increase forecasts were revised up to 10% to 15% quarter-over-quarter, NAND revised up to 15% to 20%, and enterprise SSDs reaching 23% to 28%. Cloud companies adding server memory orders and capacity shifting to HBM may cause South Korea's PC DRAM supply to decline by more than 20% in 2027, with DDR4 also remaining tight. Meanwhile, smartphone and PC demand remains very cautious, with consumer-grade NAND inventory accumulating. QLC enterprise-grade SSDs are beginning to take on some inference cache storage demand, but consumer spec downgrades and Chinese supply growth are further diverging the profit trends of enterprise-grade and consumer-grade memory.

South Korean Stock Market Capital Retreat: Market Begins Asking "Where Is the Next Round of Growth"

The KOSPI index fell 22% in the second half of the year, with trading volume decreasing 70% from its peak at the end of May, and foreign capital withdrawing $131 billion during the year. Samsung and SK Hynix together account for more than half the weight, exposing South Korea's stock market heavily to memory prices and earnings expectations. When Samsung's record profits cannot drive stock prices higher, the focus of capital trading has shifted from current profit scale to whether profits can continue to accelerate.

It is understood that Rayliant has taken the most negative "underweight" investment stance on these two largest global memory chip companies headquartered in South Korea, while another asset management giant Pictet pointed out that short-term funds chasing growth will not re-enter solely because of low valuations. Phillip Wool, Head of Portfolio Management at Rayliant Global Advisors, said: "For most investors, especially those who recently entered the South Korean market for the memory chip theme, the biggest challenge is that the easy profits and excess alpha returns from this theme have basically been fully captured." Jon Withaar, Portfolio Manager at Pictet Asset Management, said: "Especially memory chip stocks have now become value stocks, and their valuations are quite cheap based on current earnings levels." "But retail investors and hedge funds chasing short-term gains will not excessively pursue value, nor will they buy heavily just because an asset is 'cheap.' What they chase is growth."

The simultaneous contraction of institutional and retail investor buying further amplifies the adjustment. The combined 55 trillion Korean won buyback by the two companies is nearing completion, margin balances have dropped from 38.6 trillion to approximately 33 trillion, and pending investment funds have fallen from nearly 140 trillion to approximately 100 trillion. Meanwhile, China's Taiwan stock market, which also focuses on the AI computing power theme and serves as home to TSMC, MediaTek, UMC, and other leaders in the AI computing power industry chain, has risen approximately 70% during the year, outperforming South Korea by approximately 23 percentage points in the third quarter. Institutional investors seem to prefer China's increasingly complete AI computing hardware ecosystem.

These capital flows are also responding to the latest judgment of the Morgan Stanley strategist team led by Shawn Kim, namely that a more reasonable investment judgment is to shift allocation within AI computing power toward AI chips, networking infrastructure, data center optical interconnects, high-performance CPUs, ABF/glass substrates/CCL/electronic cloth/PCB, 2.5D/3D advanced packaging, and a broader complete Nvidia NVL AI rack supply chain. South Korea's market reallocation opportunities still depend on the sustainability of earnings at high levels, new buying, and whether corporate value enhancement reforms can be delivered.

Morgan Stanley Rearranges AI Computing Power Theme Investment Order: AI Chips and Optical Interconnects, Networking First, Memory Shifts to Selected and Tactical Opportunities

Morgan Stanley believes that capacity expansion has not yet resolved the shortage but has already entered the pricing phase of the stock market. The DRAM capital expenditure forecast cited in the report is $98 billion for 2026 and $150 to $160 billion for 2027, growing 82% and approximately 60% respectively. Effective new output will only begin to be released in the second half of 2027, and supply may remain tight until the end of the year. Yangtze Memory's global NAND share is approximately 12% to 13%, and ChangXin Memory's share in China's HBM market is approximately 20%, making them important variables for future supply.

Morgan Stanley strategists stated that architectural upgrades are also redistributing value. In the AI rack scenarios shown in the report, memory's relative cost share drops from 29% to 19%, while cross-rack networking is expected to push optical interconnect infrastructure's share from approximately 10% to 20%, supporting the networking supply chain in gaining a larger investment share. The Morgan Stanley research report shows that the institution's strategists calculated the hardware cost composition of different generations and configurations of AI racks as follows: memory share from 29% to 19%, comparing VR200 NVL72 with 54TB memory configuration against Rubin Ultra NVL72 configuration. High-performance data center networking infrastructure including optical interconnect systems rises from 10% to 20%, comparing the Rubin Ultra NVL72 single-rack configuration against the hypothetical configuration of the RU NVL576 eight-rack combination, which requires larger-scale cross-rack interconnection.

Morgan Stanley stated that AI investment opportunities are expanding from computing chips further to high-speed interconnects and system-level high-efficiency data transmission architectures that determine overall cluster efficiency. More importantly, multi-rack optical interconnect already has official basis. Nvidia disclosed that Rubin Ultra NVL576 will connect eight racks of 72 GPUs each into a 576-GPU NVLink domain, with optical connections used to solve cross-rack distance, high-speed signal transmission, and power consumption issues. A group of data center optical product suppliers such as Lumentum have seen continuously strong stock prices recently, not only because their capacity for the next 2 to 3 years has already been booked, but also because market news indicates that "AI chip overlord" Nvidia (NVDA.US) is adjusting its next-generation AI rack computing system architecture, namely that the next-generation AI computing cluster Rubin Ultra may reduce HBM configuration per rack and instead use optical interconnect technology architecture to connect multiple AI racks. As HBM/DRAM memory systems and NAND memory chip products have seen continuously skyrocketing prices in recent years, driven by long-term memory supply constraints, rising costs, and continuous AI cluster expansion, the focus of AI computing infrastructure investment seems to be shifting from GPU/TPU single-point performance and HBM capacity to overall data center architecture efficiency, with high-speed optical interconnect potentially being the most critical direction.

Morgan Stanley also believes that low P/E ratios have not eliminated cycle risk. The report points out that Asian memory companies' earnings forecasts have been revised up by an average of 432% this year, but the breadth of earnings upgrades for the three major original manufacturers has already declined. Historically, memory stocks can weaken months before earnings peaks. Morgan Stanley even believes that to sustain the existing stock upward logic, DRAM contract prices would need another approximately 70% quarter-over-quarter increase, a threshold that has become difficult to achieve. Morgan Stanley stated that continued price increases and slowing price increase rates can occur simultaneously, and the latter is sufficient to weaken earnings revisions and valuation expansion. The key future variables are long-term order fulfillment, return on investment after capacity expansion, and shareholder distribution capability. This is why Morgan Stanley's stock selection focus has shifted to sustained earnings revisions and per-unit value growth.

The Morgan Stanley team led by Shawn Kim's most preferred computing targets are Nvidia, TSMC, Broadcom, MediaTek, Global Unichip, Aspeed, and Montage. Networking includes Corning, LITE (Lumentum), Coherent, Furukawa, Fujikura, Soitec, Keysight Technologies, Accton, as well as TSMC, ASE, FOCI, AllRing, MPI, Winway, Hon Precision, and other CPO supply chain companies. On the memory side, attention is on ChangXin Memory's domestic substitution, Winbond and Macronix legacy-spec products, while Samsung, SK Hynix, and Kioxia are tactical choices. Supporting directions, namely key hardware support segments spanning AI computing, networking infrastructure components, and memory components, cover CCL/ABF, MLCC, packaging and testing, and semiconductor equipment, including Unimicron, Nan Ya PCB, Zhen Ding, Ibiden, Samsung Electro-Mechanics, GCE, Shengyi, Murata, Advantest, Tokyo Electron, Ebara, Canon, ASE, AEIS, MKS, NAURA, AMEC, ACMR, and ASML. The institution is most cautious on downstream hardware facing high memory costs, overbuilding, and high valuation pressure.