What China’s HBM catch-up should teach Korea
Two headlines in June should shape Korea’s artificial intelligence (AI) debate. The first was a victory lap. SK hynix overtook Samsung Electronics by common-share market capitalization for a day, powered by high-bandwidth memory (HBM) chips used in AI systems. Samsung has a fair caveat: If preferred shares are included, it remains larger. But the market signal was clear. AI has turned memory from a cyclical commodity into critical infrastructure.
Sejin Kim
The second headline was less comfortable. Korean and industry reports suggest China’s ChangXin Memory Technologies (CXMT) is moving faster in HBM than many in Seoul expected. Korea still leads. But the margin is narrowing, and that should change how Seoul thinks about its AI goals.
Start with what the CXMT story actually shows. HBM3 is no longer the frontier. Nvidia’s Rubin platform uses HBM4, and Korean firms are already pushing into HBM4 and HBM4E. SK hynix has shipped samples of 12-layer HBM4E chips to major customers, while Samsung has showcased HBM4 and HBM4E products for Nvidia’s next-generation platforms. CXMT is trying to close the gap, but the hard part is not only making a chip. It is making it at yield, stacking it reliably, managing heat, passing customer qualification and shipping at scale.
Korea should take that news seriously but not panic. China narrows gaps through patient industrial policy, large pools of capital, aggressive capacity expansion and talent recruitment. Reuters reported that CXMT planned to raise about $4.2 billion through a Shanghai initial public offering to fund DRAM expansion, including HBM. Industry sources cited by DigiTimes estimate that more than 200 Korean engineers now work at CXMT. Reuters has also reported on Korean prosecutions tied to alleged leaks of memory technology to CXMT. A protected Korean market did not stop any of this.
The wider picture is clear. The Information Technology and Innovation Foundation's 2026 Hamilton Index found that China now produces nearly one-quarter of global output across the 10 advanced industries it tracks and leads in seven of them. Its share rose from 3.5 percent in 1995 to 24.9 percent in 2022. Korea performs well in this index, ranking second by location quotient, reflecting its specialization in advanced industries. But specialization is not insulation. The same system that gives Korea strength also exposes it to China’s scale pressure.
This is where Korea risks getting in its own way. Seoul’s AI Action Plan for 2026 through 2028 has a real pro-innovation agenda, but it also points toward data localization for AI training on personal data, including private cloud infrastructure with servers located in Korea. The AI Basic Act adds obligations for high-impact, generative and high-performance AI systems. Some safeguards may be justified. But the question is whether Korea is building trust in AI or raising the cost of competing in it.
Modern AI development does not fit neatly inside national borders. Training and inference depend on distributed graphics processing unit clusters, cloud regions, model partnerships, data pipelines, security controls and specialized hardware supply chains. Forcing sensitive activity onto domestic servers may sound like a way to maintain government control over technology, but the primary effect is to make Korea a smaller and more expensive market for the global cloud providers and AI firms it wants to attract. Encryption, access controls, audit rights and risk-based safeguards can protect sensitive data without making geography a substitute for security.
Korea is not large enough to host a self-contained frontier AI economy. What it has is a place in the global AI stack that few others can match: HBM, advanced memory, manufacturing discipline, telecom networks, industrial data, data centers and engineering talent. The answer is to become the partner an allied AI ecosystem cannot build without.
The proof is already visible. At the Asia-Pacific Economic Cooperation summit in Gyeongju in October 2025, Nvidia said it would supply more than 260,000 of its most advanced AI chips to Korea’s government and major companies, including Samsung, SK Group, Hyundai and Naver. This is what AI competitiveness looks like: chips, factories, cloud, software and industrial deployment moving together.
Nvidia’s June partnership with SK hynix makes the same point. The companies announced a multiyear technology partnership to co-develop next-generation memory for global AI factories and align advanced memory supply with Nvidia’s AI infrastructure road map. The partnership also extends into semiconductor design, manufacturing simulation and digital twins for autonomous fabs. Korea’s value is not in standing apart from the AI supply chain. Its value is in becoming indispensable to it.
That is the lesson in the CXMT number. The frontier moves whether Korea builds walls or not. China will keep using scale, capital, talent and industrial policy to close the gap. Korea should answer with speed, openness and sharper specialization. It should strive to tightly integrate itself inside a broader democratic technology alliance, not pursue efforts to fragment the market. The country that keeps its place in this race will be the one that runs hardest with partners, not the one that retreats behind a market too small to defend.
Daniel Castro is president of the Information Technology and Innovation Foundation (ITIF), a leading science and technology policy think tank, and founder and director of ITIF’s Center for Data Innovation. Kim Se-jin is a tech policy analyst at ITIF’s Center for Korean Innovation and Competitiveness.