Why the current chip boom, price surge will last longer than before

Samsung Electronics Executive Chairman Lee Jae-yong, left, SK Group Chairman Chey Tae-won, right, and Nvidia CEO Jensen Huang at the Korea-U.S. business roundtable held in Washington, D.C. on Aug. 25, 2025. [PRESIDENTIAL OFFICE]

[MONEY MOVER]


Capital moves in and out of Korea, driven by a range of economic and geopolitical forces. In our "Money Mover" series, we explore key market developments that could shape investment decisions and influence the flow of global funds. — ED.
 
Korea’s semiconductor industry is riding a boom unlike anything in its history. What was once a cyclic business tied to smartphones and PCs has suddenly become the backbone of the global AI race, with Samsung Electronics and SK hynix sitting at the center of a global battle for computing power that is reshaping markets and investment flows worldwide.
 
The two memory giants, which together accounted for 36.9 percent on Kospi's total weight as of Friday, have driven the country's bourse to some of the world's highest gains this year.
 
Samsung Electronics historically became the first Korean company to surpass the 1-quadrillion-won market value mark as its shares climbed to a record 169,100 won on Wednesday, while SK hynix recently hit 909,000 won per share, lifting its market capitalization to 661 trillion won — nearly 10 times its peak in the last semiconductor upcycle in 2018.
 

 
Many analysts now argue that the current “semiconductor supercycle” is not a temporary surge but a new normal, forcing investors to rethink how they evaluate risk, profitability and long-term growth in the tech sector. Yet questions linger: When will Big Tech’s massive AI spending translate into real profits? And what happens to Korean chipmakers if the AI investment wave slows?
 
Beyond the worldwide scramble to secure supplies of high bandwidth memory (HBM), shortages in conventional dynamic random access memory (DRAM) and NAND flash memory are also pushing prices higher as chipmakers prioritize AI-related products over consumer electronics. Despite bubble fears, both Samsung and SK hynix delivered record-breaking results in 2025, with SK hynix even surpassing Samsung in full-year operating profit for the first time, powered by high-margin products such as next-generation HBM and server DDR5.
 
 In this climate of soaring expectations and lingering doubts, investor sentiment remains split between believers in a lasting AI transformation and skeptics warning of an inevitable correction. To shed light on what lies ahead, the Korea JoongAng Daily spoke with Peter Lee, managing director and head of Korea Research at Citigroup Global Markets Korea Securities, Kim Dong-won, senior managing director and head of research at KB Securities and Roh Geun-chang, executive vice president of Hyundai Motor Securities’ research center, about how this AI boom could prove a decisive turning tide for Korea's chipmakers — one that has already surpassed anything seen before.
 

Q. How would you characterize the structural differences between the current semiconductor “supercycle” and previous upcycles, particularly in terms of demand drivers, capital intensity, and durability of end-market demand?
A. Lee: The last time semiconductors went through such an upcycle was from 2001 to 2007. Led by NAND, it was triggered by a shift to MP3 players from CD players and digital cameras from analog cameras. Typically, memory cycles fluctuate every one to two years, but if the upturn is to last longer, that is when new, unprecedented demand emerges. 
 
We’re currently reexperiencing one such phase, as AI demand transitions from the training stage to inferencing. While HBM was largely used in the training stage, conventional DRAM — on top of HBM and graphics DRAM — is used more extensively in the inferencing stage.  
 
Kim: Looking ahead, the widespread adoption of agentic AI and the full-scale emergence of physical AI are expected to generate new sources of demand for AI semiconductors, making this cycle likely to exhibit prolonged structural growth.  
 
Demand for AI semiconductors appears to reflect a tangible transformation across industries, making a prolonged supply shortage of both DRAM and NAND increasingly inevitable. As AI models evolve from the training phase to the inferencing phase, along with the expansion of agentic AI services, AI memory demand is projected to rapidly broaden from HBM-centric usage to server-class memory.  
 
Structural expansion of the physical AI market — which is generating new demand for high-end memory capable of demonstrating high performance with low power in applications such as humanoid robots and autonomous driving — is underpinning a steady upward trajectory in demand for AI chips.


Roh: AI data centers now use a much broader range of semiconductors than in past cycles. In particular, HBM paired with GPUs is consuming significantly more memory, while expanding AI inference is boosting demand for both server DRAM and multiple types of SSDs. As a result, the amount of memory content per data center server has at least doubled compared to the past.
 
On the supply side, chipmakers remain cautious after the 2023 downturn, keeping capacity tight even as demand has surprised to the upside. This suggests the current upcycle is likely to be prolonged, with unusually strong profitability for memory producers. Given that supply is running about 10 percent below demand — and that even a 5 percent shortage can lift prices 40 to 50 percent — memory prices could potentially rise by as much as 2.5 times before the cycle turns.
 

Samples of sixth-generation high bandwidth memory (HBM4) and prior generation HBM3E are displayed at Samsung Electronics booth set up for Semiconductor Expo 2025 held at Coex in southern Seoul on Oct. 22, 2025. [YONHAP]

 
What is your view on the likely duration of the current upturn stage, and what key indicators would signal its peak or inflection point?
Lee: This upcycle began around 2024 and is expected to last longer than the previous 2001-2007 upcycle. Whereas the earlier cycle was driven by NAND, the current acceleration phase represents a synchronized upcycle across both DRAM and NAND. As AI evolves to increasingly replace human labor, its overall impact is expected to be far more profound than the changes that drove the previous cycle. AI models will continue to be updated, while physical AI and on-device AI are expected to develop in parallel, supporting a sustained phase of structural growth.
 
Kim: Memory supply shortages driven by capacity constraints among suppliers are expected to persist at least through 2027. The upswing is expected to expand from an HBM-centered phase in 2024-2025 to server memory, including server DRAM and eSSD, in 2026-2027, leading to an unprecedented level of supply tightness.
 
The timing of the peak will depend on supply and demand dynamics. On the supply side, the scale of shipment increases following the completion of SK hynix’s Yongin Cluster Phase 1 fab and Micron’s New York fab, and on the demand side, the extent to which Big Tech companies’ AI businesses translate into monetization and sustained AI-related capital expenditures.


Roh: The biggest risk to the current upcycle would be if major investors such as OpenAI or SoftBank run out of funding and slow or stop their buildouts — that would likely mark the end of the upcycle.
 
Even if demand remains solid, memory prices cannot rise forever. If chips start to account for more than about half of total server costs, customers are likely to become more cautious about further investment, especially once their initial infrastructure buildout is largely complete. For now, AI data center spending is expected to hold up through the end of this year, but there is a possibility of a sharper slowdown toward late in the fourth quarter if customers begin inventory adjustments or pause new investments — a risk the industry is watching closely.
 

SK hynix's Yongin semiconductor campus in Gyeonggi is under construction, which is scheduled to be completed in the first half of 2027. [SK HYNIX]

Given recent market concerns — including skepticism around the pace at which AI investments are translating into revenue — do you view current AI infrastructure spending as sustainable? Additionally, when do you expect U.S. hyperscalers to meaningfully monetize their AI investments?
Lee: Concerns about an AI bubble persist. However, until AI reaches a certain level of maturity, Big Tech companies are expected to continue investing in a bid to gain market leadership. While some view current investment levels as excessive, the picture changes if those investments ultimately turn out to generate greater value. To the extent that this dynamic plays out, semiconductor companies stand to be the primary beneficiaries.
 
Kim: The massive AI infrastructure spending by Big Tech companies is not aimed at near-term monetization but rather stems from competition to secure market leadership in the AI services market. Big Tech firms have already been developing and refining a range of AI monetization strategies, and have achieved a meaningful degree of success through paid subscriptions and integration into broader AI tool ecosystems.
 
While there may be periods of moderation in AI infrastructure investment, a sudden or sharp contraction in the near term appears unlikely. Moreover, even assuming the existence of an AI bubble, it is not expected to burst easily in the absence of a significant external shock.
 
Roh: There is a possibility of an AI investment bubble, particularly as data center spending could prove to be duplicated or excessive. However, semiconductor companies cannot judge when such overinvestment might end, which is why they have kept capital spending for facility investments relatively cautious. The conservative stance contributes to maintaining unusually strong price momentum in this cycle.
 
Unlike past cloud computing investments, many companies now feel they must invest in AI to survive, even if returns are uncertain. As a result, firms are prioritizing deployment first and leaving questions about profitability for later.
 
While funding concerns could eventually slow investment, the risk is unlikely in the first half of this year since major players secured capital late last year. By year-end, however, fundraising could become more difficult, especially as OpenAI faces rising competition from players like Google and Anthropic.
 
In short, profitability is not driving today’s investments — fear of falling behind is. But if monetization takes too long, investment behavior could eventually shift.
 

Nvidia CEO Jensen Huang receives an SK hynix HBM4 semiconductor wafer as a gift from SK Group Chairman and Korea Chamber of Commerce and Industry Chairman Chey Tae-won at the APEC CEO Summit held at the Gyeongju Arts Center in Gyeongju, North Gyeongsang, on Oct. 31. [JOINT PRESS CORPS]

 
There are concerns that Samsung and SK hynix may have already peaked, as their share prices are trading at all-time highs. Based on your outlook for memory pricing, demand and competitive positioning, how much further upside do you see for Samsung Electronics and SK hynix shares from current levels?
Lee: The upside potential lies in AI demand rising further if adoption continues to grow faster than expected and more devices are equipped with AI capabilities. The downside risk, on the other hand, could come from weakness in PC and smartphone demand.  
 
External factors include Korea’s market conditions. Recently, there have been strong market interests in Korean stocks, helped by factors including the third revision to the Commercial Act, requiring companies to retire treasury shares within fixed deadlines, and the government's move to encourage firms to boost dividend payouts.
 
Kim: With the advent of the physical AI era, a significant increase in the deployment of high-capacity memory and high-performance, low-power memory such as LPDDR5X is expected. In the NAND sector, rising demand for ICMS (Inference Context Memory Storage), which will be used in Nvidia’s Vera Rubin AI computing platform, is expected to provide significant upside to future earnings.
 
The key drivers for further stock gains will be production capacity and the strategic positioning of memory semiconductors. Given the already deepening supply shortages, memory is increasingly recognized as a strategic core asset, making current supply capacity and the ability to secure additional future production key factors in determining future stock performance.  
 
Roh: Stock prices are already at unprecedented levels, so past comparisons are no longer very meaningful, as the market has moved into a new “new normal.” When that ultimately ends is unclear, but if clear warning signs appear, a correction would be possible.
 
Near term, any pullback is likely to be mild through late April, when first-quarter results are released, and earnings expectations could be revised higher at that point. However, memory prices have already surged — roughly 50 percent in the fourth quarter of 2025 and anticipated to further rise to 70 to 80 percent in the first quarter this year — so the pace of increases should slow from the second quarter onward, which could raise volatility from May.
 
From May, slower price gains and customer concerns about high memory costs could lead some buyers to delay or cancel orders in the second half of the year.
 
 

A view of inside Samsung Electronics’ Hwaseong semiconductor campus in Gyeonggi [SAMSUNG ELECTRONICS]

 
How do you assess geopolitical risks — particularly potential U.S. semiconductor tariffs or export controls — and their likely impact on global supply chains and Korean chipmakers specifically?
Kim: The chance of imposing a high-rate tariff as a bargaining tool to pressure Korea into fulfilling its U.S. investment commitments seems limited. Thanks to the learning effect from the past year, there is little concern that such threats would translate into actual harm, and the Korean National Assembly is expected to soon propose legislation related to U.S. investment. Even if higher-than-expected tariffs are imposed, Korean semiconductors are strategic assets with no viable substitutes. Cost increases could be passed on to customers through price adjustments, but as competitors face similar conditions, the relative impact would be limited.  


Roh: I have consistently said for several years that tariffs are largely irrelevant in this case. The reason is that AI semiconductors are not exported directly to the United States — we supply them to TSMC, which then packages them and delivers them to Nvidia and other U.S. customers. Moreover, TSMC is currently building manufacturing facilities in the United States, which effectively means these products enter the United States tariff free. As a result, this issue has very little practical impact.
 
 
Similarly, direct exports of conventional memory semiconductors to the United States are not that significant. Most of these chips are embedded in smartphones or PCs, and U.S. customs does not disassemble finished products to impose separate tariffs on individual memory components. Therefore, U.S. tariffs have had — and are likely to have — minimal effect on Korean memory exports.
 
 

An aerial view of SK hynix's M16 semiconductor plant in Icheon, Gyeonggi [SK HYNIX]

 
How long do you expect Samsung and SK hynix to retain their leadership positions in the HBM market? What technological, competitive or policy-related factors could potentially disrupt their dominance?
Lee: There is nothing immediately threatening, but China could become a key point of concern. China is aggressively pursuing localization to achieve self sufficiency amid the dispute with the United States. 
 
China has also advanced rapidly in software development, and is already outperforming Korea in that area. However, hardware development is taking longer, as it depends heavily on both software and infrastructure, which require all processes, equipment, and materials to be fully in place. This complexity is one reason we have been able to maintain a technological lead over China.
 
Kim: Micron’s expansion of production capacity and an attempt to increase market share with government support could be seen as a potential threat, though there are more upsides. Domestic memory companies are evaluated to be ahead of overseas competitors in both continuously meeting customers’ higher specification requirements and providing the desired volumes amid the continuous rise of memory specification, making the comparative advantage unlikely to be threatened.  
 
Demand for specialty memory is also expected to remain strong over the next three to five years as the AI paradigm rapidly shifts to physical AI, which requires high bandwidth, low-power memory such as HBM, LPDDR5X, and GDDR7.


Roh: At this point, it appears highly unlikely that Samsung and SK hynix will lose their leadership in the HBM market in the foreseeable future. The market structure is more likely to remain a stable “Big Three” framework rather than shift dramatically in favor of any single player. Micron is also investing heavily in HBM and will continue to be a meaningful competitor, but its overall production capacity remains smaller than that of the two Korean firms. In short, barring a major technological disruption or regulatory shock, the current leadership structure in HBM by the two Korean firms is likely to persist.
 

BY JIN MIN-JI, LEE JAE-LIM [lee.jaelim@joongang.co.kr]