The license

On December 30, 2025 — the last business day of the year — the US Bureau of Industry and Security issued annual licenses to Samsung Electronics and SK Hynix, authorizing them to continue operating their semiconductor fabrication facilities in China for another 12 months.^[2]

The day before, on December 31, BIS had revoked the Validated End-User (VEU) authorizations that had previously governed these operations. The VEU program allowed pre-approved companies to receive controlled technology without individual export licenses. Its revocation meant that Samsung and SK Hynix — which had been operating their Chinese fabs under VEU since 2018 — now needed explicit, annually renewable licenses from Washington to continue making memory chips in China.^[1]

If you were reading this as a contract, the terms changed as follows:

Before (VEU): A standing authorization. The company is pre-approved. Equipment, technology, and materials flow without transaction-level review. The authorization is indefinite, subject to compliance audits.

After (annual license): A renewable permit. The company must apply each year. Washington reviews the application against current policy priorities. The license can be modified, conditioned, or denied. Each renewal is, in practice, a negotiation.

The practical effect: Samsung and SK Hynix went to sleep on December 30 as companies with standing rights to operate in China. They woke up on December 31 as companies operating in China at Washington's pleasure, renewable annually.

What's in China

It helps to understand what, exactly, the licenses are protecting.

Samsung operates a NAND flash memory fabrication facility in Xi'an, Shaanxi Province. This factory accounts for approximately 35-40% of Samsung's total NAND production.^[3] NAND flash is the storage memory in your phone, your laptop, and every data center in the world. Samsung is the world's largest NAND producer. So roughly a third of the global leader's output comes from a factory that now requires an annual permission slip from the US government to operate.

SK Hynix operates a DRAM fabrication and packaging facility in Wuxi, Jiangsu Province. This facility accounts for approximately 40% of SK Hynix's DRAM production.^[8] DRAM is the working memory in every computing device — the volatile memory that handles active processing. SK Hynix is the world's second-largest DRAM producer and the leading manufacturer of high-bandwidth memory (HBM), the specialized DRAM that Nvidia's AI chips require.

Combined, these two facilities represent a substantial fraction of the world's memory chip supply. They are also, from a US export control perspective, the most consequential pieces of Korean industrial infrastructure operating on foreign soil.

The structural tension:

1. The US wants to restrict China's access to advanced semiconductor technology.
2. Samsung and SK Hynix's Chinese fabs are the most advanced semiconductor facilities in China — but they are Korean-owned and produce for the global market, not for Chinese customers who would use the chips for military or AI applications that concern Washington.
3. Shutting down these fabs would devastate global memory supply, spike prices for every electronic device, and destroy a significant portion of Korean corporate value.
4. So Washington compromises: allow the fabs to operate, but convert the authorization from a standing right to a renewable privilege, thereby gaining leverage over Seoul.

This is not a sanctions regime. It is a dependency architecture. Korea's two largest technology companies — which together account for roughly 20% of South Korea's total stock market capitalization — now need Washington's annual permission to maintain their Chinese operations.

The HBM premium

The story gets more pointed when you look at what SK Hynix specifically makes.

High-bandwidth memory is the bottleneck component in AI training and inference hardware. Nvidia's H100 and B200 GPUs — the chips that OpenAI, Google, Microsoft, and everyone else running large language models needs — each require multiple HBM stacks. SK Hynix is the dominant supplier, with Samsung as the only meaningful competitor. TSMC packages the HBM stacks but doesn't manufacture the DRAM dies.

The HBM market was approximately $38 billion in 2025 and is projected to reach $58 billion by 2027.^[4] SK Hynix alone captured an estimated 50-55% of this market. Its Wuxi facility, while primarily producing conventional DRAM, is part of the integrated supply chain that enables HBM production at SK Hynix's Korean fabs.

The mechanism by which the annual license creates potential leverage:

1. SK Hynix needs its Wuxi DRAM facility to maintain overall production volumes and supply chain efficiency.
2. SK Hynix's HBM production — which happens in Korea — depends on the company's ability to allocate DRAM capacity flexibly across its global footprint.
3. If Washington restricts or conditions the Wuxi license, SK Hynix's production planning is disrupted, which could ripple into HBM supply.
4. HBM supply disruptions would affect Nvidia's GPU production, which would affect every major AI company's training capacity.
5. So the annual license on a DRAM factory in Wuxi is, at least in theory, a potential pressure point on AI hardware supply chains.

I might be wrong about how directly Wuxi's production links to HBM allocation. SK Hynix doesn't disclose its internal capacity planning in that detail. But the structural logic — that constraining one node in an integrated production network creates effects at other nodes — is sound.

The tariff layer

On top of the export control regime, there is a tariff regime, and the two don't always point in the same direction.

Semiconductors imported into the US are subject to a Most Favored Nation tariff cap of 15%.^[9] This has effectively been the ceiling, even as tariffs on other Chinese imports escalated. The reasoning is self-interested: the US doesn't manufacture enough memory chips domestically, so taxing memory imports raises costs for every American technology company without creating domestic jobs.

However, in 2025 the Commerce Department launched a Section 232 investigation into semiconductor imports, arguing that dependence on foreign-produced chips constitutes a national security risk.^[6] Section 232 allows the president to impose tariffs beyond MFN rates on national security grounds — the same authority used for steel and aluminum tariffs.

If Section 232 tariffs are applied to semiconductors, the rate could go as high as 25%. For AI-specific chips (advanced logic and HBM), a 25% tariff would represent a meaningful cost increase for data center buildouts.

The contradiction:

1. The US restricts Samsung and SK Hynix from upgrading their Chinese fabs (export controls limit technology transfers to China).
2. The US simultaneously encourages Samsung and SK Hynix to build new fabs in the US (via CHIPS Act subsidies — Samsung is building a $17 billion fab in Taylor, Texas; SK Hynix has a $3.87 billion packaging facility planned for Indiana).
3. The US might also impose tariffs on the chips that Samsung and SK Hynix produce in Korea and export to the US (Section 232).
4. So the message to Korean chipmakers is: don't make chips in China (we'll restrict you), do make chips in America (we'll subsidize you), and by the way, we might tax the chips you make at home (Section 232).

If you were a Korean semiconductor executive trying to plan a five-year capital allocation strategy, the policy environment reads less like a set of rules and more like a set of suggestions that change annually.

The megacluster response

Seoul's response has been to double down on domestic capacity. In 2025, the Korean government announced a 622 trillion won (approximately $470 billion) semiconductor megacluster — a coordinated investment plan spanning government subsidies, tax incentives, and private capital commitments from Samsung, SK Hynix, and their supply chain partners.^[7]

The plan calls for building multiple new fabs in the Yongin-Pyeongtaek corridor south of Seoul, creating what would be the world's largest concentration of semiconductor manufacturing in a single geographic zone.

The strategic logic is defensive:

1. If Chinese fabs are at risk of being shut down by US export controls, build replacement capacity at home.
2. If US fabs are expensive and subject to unpredictable policy demands, keep the center of gravity in Korea.
3. If tariffs make Korean exports to the US more expensive, maintain cost competitiveness through scale and government support.
4. The megacluster is a hedge against all three risks simultaneously.

But it's also a 10-year play. The new fabs won't be operational until 2028–2030, and the timeline carries execution risks beyond mere scheduling: skilled labor shortages in semiconductor manufacturing, power and water infrastructure demands for advanced fabs, and the concentration of roughly $470 billion in capex in a single geographic corridor. In the interim, Samsung and SK Hynix remain dependent on their Chinese operations, which remain dependent on Washington's annual license, which remains dependent on the US-China relationship, which remains dependent on — among other things — a war in the Middle East that has reshuffled everyone's diplomatic alignments.

The KOSPI problem

On March 3, 2026, the KOSPI — South Korea's benchmark stock index — fell 7.2%.^[5]

The proximate cause was the Iran conflict and its ripple effects: oil prices spiking (Korea imports virtually 100% of its oil), shipping routes disrupted (Korea is an export economy entirely dependent on maritime trade), and a general risk-off move in Asian markets.

But there was a semiconductor-specific overlay. The market was pricing the possibility that the Iran crisis could accelerate US-China decoupling, which could in turn increase pressure on Korean fabs in China and compress the timeline on which Samsung and SK Hynix might have to choose between their Chinese operations and their US relationships.

Samsung's stock fell 9.3% on the day. SK Hynix fell 11.2%. Together, these two companies represent roughly 25% of the KOSPI's market capitalization. Their decline accounted for more than half of the index's drop.

The mechanism is potentially recursive:

1. War in the Middle East disrupts the global economy.
2. Economic disruption intensifies US-China rivalry.
3. Intensified rivalry makes the annual license for Chinese fabs less certain.
4. Less certainty about Chinese fabs reduces the value of Korean semiconductor companies.
5. Reduced semiconductor valuations drag down the Korean stock market.
6. If sustained, a weaker stock market could constrain the Korean government's ability to fund the megacluster.
7. A delayed megacluster would mean longer dependence on Chinese fabs.
8. Which would mean more exposure to the annual license.

This is a reinforcing dynamic, or at least the beginning of one. Steps 1 through 5 are observable; steps 6 through 8 depend on duration and severity. The way out is for the war to end, US-China tensions to stabilize, and the license renewals to become pro forma rather than political. None of those things is guaranteed.

What this is

Korea's semiconductor situation is a particular kind of statecraft problem — one in which a middle power's core industry sits at the intersection of two great powers' strategic competition, and every possible move has costs. The options below are stylized — no boardroom decision is this clean — but they frame the trade-offs:

Option A: Stay in China. Maintain the Xi'an and Wuxi fabs, accept the annual license dependency, hope Washington doesn't tighten the terms. Risk: Washington tightens the terms.

Option B: Exit China. Wind down the Chinese fabs, take the write-down (Samsung's Xi'an facility alone is worth an estimated $15–20 billion in replacement cost), and rebuild capacity in Korea and the US. Risk: the exit takes years, destroys shareholder value, and cedes the Chinese market to domestic competitors like YMTC and CXMT.

Option C: Diversify everything. Build the megacluster in Korea, build the US fabs, maintain the Chinese fabs as long as permitted, and hope that geographic diversification provides enough resilience to survive any single point of failure. Risk: capital overextension, execution complexity, and the possibility that Washington conditions the next license renewal on reductions in Chinese capacity.

Samsung and SK Hynix appear to be pursuing Option C, which is probably the right answer if you can afford it. Whether they can afford it depends on HBM demand staying strong, DRAM and NAND pricing remaining supportive, and the Korean government maintaining its subsidy commitments through what may be an economically challenging period.

One plausible reading of Korea's semiconductor situation: a temporary awkwardness created by the current geopolitical moment, manageable through diplomacy and capital allocation. Another way: a permanent feature of a world in which the two largest economies are trying to decouple their technology supply chains, and every country that sits in between gets squeezed.

The annual license renewal for Samsung and SK Hynix's Chinese fabs is scheduled for December 2026. Between now and then, the companies need to navigate a war in the Middle East, a tariff investigation at Commerce, a stock market correction, and the ongoing construction of a $470 billion domestic manufacturing complex.

---

Things happen

Samsung's Xi'an fab was originally built in 2014. SK Hynix's Wuxi facility opened in 2006. YMTC, China's domestic NAND champion, is currently producing at 128-layer technology, roughly two generations behind Samsung's 236-layer V-NAND. CXMT, China's DRAM entrant, is at DDR4 quality, roughly three years behind SK Hynix's HBM3E. The CHIPS Act allocated $52.7 billion for US semiconductor manufacturing subsidies. Intel received $8.5 billion. TSMC's Arizona fab received $6.6 billion. Samsung's Taylor, Texas fab received $6.4 billion. SK Hynix's Indiana packaging facility received $458 million. South Korea's semiconductor exports accounted for roughly 20% of the country's total exports in 2025. The won fell 2.1% against the dollar on March 3. Nvidia's market capitalization declined by approximately $200 billion on the same day.