Every Nvidia GPU shipped to a hyperscaler data centre contains high-bandwidth memory. That memory is, in the overwhelming majority of cases, made in South Korea. SK Hynix's facility in Icheon, about an hour south of Seoul, is the most important single building in global AI infrastructure — not because of its symbolic significance but because of its production reality. HBM is not a commodity input that can be substituted from multiple sources. It is a precision-engineered stack of DRAM dies connected through thousands of microscopic vertical channels, manufactured to tolerances that require years of process development and billions in capital investment to achieve. SK Hynix commands roughly 50% of the global HBM market. Samsung holds approximately 40%. Together, two Korean companies supply approximately 90% of the memory technology that makes modern AI compute possible.1 That concentration of strategic capability in a single mid-sized country, one located 50 kilometres from a heavily armed adversarial state and deeply embedded in the supply chains of two great powers who are in open technological competition with each other, is one of the most consequential structural facts in global markets. It is also, for investors, one of the most interesting.
Understanding Korea's semiconductor position requires separating two distinct businesses that happen to be housed in the same companies. The first is the HBM business — advanced stacked memory for AI accelerators — which is a high-margin, supply-constrained market where demand has been growing faster than capacity additions. The second is the broader DRAM and NAND flash memory business, a significant portion of which is manufactured in China and sold globally as a commodity input. These two businesses have radically different geopolitical risk profiles, different margin structures, and different investment implications. Much of the confusion in how Korean semiconductor companies are discussed conflates them.
The HBM dominance and how it was built
SK Hynix's emergence as the undisputed leader in HBM is one of the more remarkable competitive stories in the semiconductor industry, and its origins are worth understanding because they illuminate the durability of the advantage. When AI infrastructure buildout began accelerating in earnest in 2022-2023, SK Hynix had already been developing HBM technology for nearly a decade, having worked on the first generation since the early 2010s. The company made the counterintuitive decision to prioritise HBM development when it was still a niche product for high-performance computing, not yet the critical bottleneck in AI training that it subsequently became. That early commitment translated into a process technology lead that proved extremely difficult for Samsung to close.
Samsung's HBM story through 2024 was a painful lesson in the difference between manufacturing scale and process sophistication. Samsung has substantially larger semiconductor revenues and far greater capital resources than SK Hynix. It also spent much of 2023 and 2024 failing to pass Nvidia's qualification testing for its HBM3E product — the generation of stacked memory essential for the H100 and H200 GPU families that dominated the AI infrastructure buildout. While SK Hynix was shipping HBM3E at scale to Nvidia throughout 2024, Samsung's production could not meet the thermal and yield specifications that Nvidia requires. The consequence was a competitive position inversion that would have seemed implausible three years earlier: SK Hynix, the smaller of the two Korean memory giants, posting record operating profit of KRW 47.2 trillion for full year 2025, overtaking Samsung's KRW 43.6 trillion for the first time in the companies' histories.2
The China exposure that doesn't appear in the AI narrative
The narrative that dominates coverage of Korean semiconductors in 2025 is the AI tailwind, and that narrative is well-founded. But it has obscured a separate and significant risk that sits in the other half of the business. Samsung manufactures approximately 35-40% of its total NAND flash memory at its facility in Xi'an, Shaanxi province. SK Hynix produces roughly 40% of its DRAM at its Wuxi, Jiangsu facility and approximately 20% of its NAND at its Dalian, Liaoning plant. These are not marginal operations. They are core manufacturing bases for the legacy memory products — general-purpose DRAM and NAND flash used in consumer electronics, servers, and industrial applications — that still account for a substantial share of both companies' revenue.3
From October 2022, when the US first imposed sweeping semiconductor export restrictions on China, both Samsung and SK Hynix operated their Chinese facilities under a regime of validated end-user waivers — exemptions that allowed them to import US-origin chipmaking equipment into China without individual export licences. These waivers were renewed periodically and represented a pragmatic accommodation of the reality that abruptly cutting Korean access to their Chinese fabs would cause significant supply disruption in commodity memory markets globally. In September 2025, the US Bureau of Industry and Security removed both companies from the validated end-user programme, effective December 31, 2025. The era of frictionless operation of Korean fabs in China ended.4
The replacement regime, announced in late December 2025, is an annual licensing system. Samsung and SK Hynix were granted licences to continue operating their existing Chinese facilities through 2026, but under terms that explicitly prohibit capacity expansion or technology upgrades. The US position, stated clearly in the BIS announcement, is that Korean companies may maintain their Chinese production at current levels but may not grow it, improve it, or make it more advanced. The practical implication is that their Chinese fabs become frozen in technological time — producing progressively less competitive legacy products as process nodes advance elsewhere — while the investment capital for new capacity and next-generation technology flows to Korea, the United States, and potentially other allied-country locations.5
The US position is precise: Korean companies may operate their Chinese fabs at existing levels, but may not expand capacity or upgrade technology. The fabs will not be shut down. They will be made obsolete, gradually and deliberately.
Korea's impossible position and how it is navigating it
The structural difficulty facing Korean semiconductor policy is not that the US-China bifurcation presents a choice between bad options. It is that the two largest Korean semiconductor companies are so deeply embedded in both supply chains that any policy choice produces asymmetric costs for one set of relationships. China is Korea's largest export market. The US is Korea's primary security guarantor. Korean chipmakers' Chinese fabs serve global commodity memory markets that Chinese customers depend on. Korean chipmakers' HBM production serves the AI infrastructure of US hyperscalers that China's technology industry cannot easily access. Optimising for one relationship structurally compromises the other.
The Lee administration's navigation of this tension has been notably pragmatic. Korea has moved to align its export control framework more closely with US standards — a prerequisite for eventually receiving the Foreign Direct Product Rule exemption that Japan, the Netherlands, and 31 other countries enjoy, which would reduce the regulatory friction on Korean companies' technology exports. Progress on FDPR exemption status would materially improve Korean companies' operational position by removing the requirement to seek US authorisation for each technology transfer. As of late 2025, South Korea was still working toward that status, with discussions ongoing.6
Meanwhile, both Samsung and SK Hynix have significantly accelerated their investment in domestic Korean capacity. SK Hynix's Yongin semiconductor cluster — a massive multi-fab complex in Gyeonggi province — represents one of the largest single semiconductor capital investment programmes in the world, designed specifically to house the next generations of HBM production. Samsung's investment in its Taylor, Texas facility in the United States, alongside continued domestic expansion, reflects the same strategic imperative: build advanced capacity in jurisdictions where it is insulated from geopolitical friction. The Chinese fabs are becoming legacy operations by design — not abandoned, but deliberately left behind as the companies invest their growth capital elsewhere.7
The China countermove and its limits
China's response to the HBM export controls has followed its established pattern for semiconductor restrictions: accelerated domestic development investment, aggressive talent acquisition, and retaliatory export controls on materials. Changxin Memory Technologies, China's primary HBM development effort, has made measurable progress. China has also imposed export restrictions on germanium, gallium, and other rare earth materials that are inputs to semiconductor manufacturing — applying pressure at the supply chain points where it has leverage.
The honest assessment of where Chinese HBM capability stands, as of late 2025, is that it remains multiple generations behind the commercial frontier. Korea's SK Hynix is producing HBM3E and working on HBM4. China's domestic producers are working on earlier HBM generations with considerably lower yield and capacity. The technological gap is real, and it is not a gap that closes quickly — HBM manufacturing requires not just advanced lithography but extremely precise through-silicon via technology, thermal management, and yield optimisation that represent years of accumulated process knowledge. The US export control regime, by restricting access to US-origin equipment and materials needed to advance that process knowledge, is specifically designed to widen that gap or at minimum prevent it from closing at the pace that unfettered access would allow.8
The risk to Korean companies from China's domestic semiconductor development is therefore asymmetric. In commodity DRAM and NAND flash — the legacy memory markets — Chinese domestic producers like CXMT and YMTC are genuine competitive threats that are already affecting pricing and market share in lower-specification products. In HBM, the threat is real over a five-to-ten year horizon but is not a near-term commercial risk. The export controls that restrict Korean access to China actually, in a narrow sense, benefit Korean HBM producers by limiting China's ability to produce competitive alternatives. The technology whose export is being restricted is also the technology that is most valuable to protect.
The investment framework
The Korean semiconductor investment case in the context of geopolitical bifurcation resolves into three distinct time horizons, each with different dominant risks.
In the near term, through 2026, the primary variable is HBM demand from the AI infrastructure cycle. This is driven by US hyperscaler capital expenditure, Nvidia's GPU production ramp, and the competitive dynamics between SK Hynix and Samsung for Nvidia's qualification. Samsung's clearance for HBM3E supply to Nvidia, achieved in late 2025, partially addresses the competitive gap that had been the central concern about the company's position. Both Korean companies are supply-constrained on advanced HBM, which means near-term earnings are more sensitive to production ramp and yield improvement than to demand. The geopolitical friction around Chinese fabs is a background risk but not the primary near-term earnings driver.
In the medium term, the Chinese fab transition is the dominant issue. Samsung and SK Hynix are managing an orderly wind-down of Chinese fab investment while accelerating domestic and US capacity. The cost of this transition — stranded capital, production disruption, relocation expenses — is real but manageable given the earnings levels both companies have been generating. The more important question is whether the annual licensing regime for Chinese fab operations remains stable or deteriorates. A sudden revocation of operating licences would be significantly more disruptive than the gradual managed transition currently underway. The political durability of the annual approval system — which depends on US-Korea trade relations remaining constructive — is therefore a non-trivial risk variable.
In the long term, the question is whether Korea's semiconductor leadership is durable enough to survive both the competitive challenge from Chinese domestic development and the capital intensity required to maintain process leadership at successive generations of HBM. SK Hynix's early investment in HBM technology is the most important competitive moat in Korean equities today. Maintaining that moat requires continued process leadership, continued alignment with leading AI chip designers, and continued access to the equipment and materials ecosystem that supports advanced memory manufacturing. All three are currently intact. None should be taken for granted.