📚 US-China Summit Series Pre-Summit Guide: 10 Agendas for Korean Investors / Summit Result: No Grand Bargain / CloudMatrix Expansion Scenario / NVIDIA Earnings and Korea’s Second-Order Supply Chain / AI-RAN and the Korean Supply Chain

The May 14-15 Beijing summit was not a grand bargain. It was a managed truce. Washington and Beijing agreed to avoid immediate escalation, but the core issues — semiconductor export controls, rare earths, Taiwan, and tariff structure — were not solved. The more important signal came outside the room. The U.S. cleared H200 exports, but China has not let purchases proceed. Huawei’s CloudMatrix 384, meanwhile, shows a different Chinese response: weaker chips offset by more chips, more power, and more optical links. Decoupling is not ending. It is moving into a more expensive equilibrium.

Key Takeaways

• The summit was a managed truce. It reduced near-term escalation risk but did not resolve the technology-control structure.
• The H200 episode matters more than the communique. The U.S. has cleared sales to roughly 10 Chinese firms, but reports indicate no H200 chips have shipped because China has not approved purchases.
• CloudMatrix 384 is a system-level workaround. Huawei is not beating NVIDIA at chip efficiency. It is using many more chips, much more power, and a larger optical fabric to reach useful system-level performance.
• ITIF’s model complicates the “decoupling is bad for Korea” narrative. In a full semiconductor decoupling scenario, ITIF estimates U.S. firms could lose USD 77B in China-related sales in the first year, while South Korean firms could gain roughly USD 21B.
• Korea’s exposure is time-dependent. The short term favors HBM, memory, substrates, MLCC and optical links. The medium term is mixed. The long term brings standard fragmentation, Chinese substitution, and dual-supply-chain costs.
• The cleanest second-order alpha is optical interconnect. The U.S. stack needs more CPO and rack-scale optics; the Chinese stack needs even more optical links because it compensates with scale.

1. What the Summit Actually Signaled

The summit should be read as a pause, not a settlement. There were preliminary understandings, purchase intentions, and stability language. But there was no binding package that changes the core technology conflict.

That distinction matters. Tariffs can be negotiated. Semiconductor controls are strategic. If export controls remain structurally intact, then the AI infrastructure map remains split into two partially separate systems.

The political incentives explain the outcome. The U.S. wanted market stability and visible business wins. China wanted time and did not want to look as if it was accepting U.S. terms. Both sides wanted to avoid a visible breakdown; neither side had a strong reason to make a structural concession.

Xi Jinping’s reference to the Thucydides Trap should also be read in that frame. On the surface it is a call to avoid conflict. Underneath it sits a time-horizon argument. If the U.S. restricts exports, China accelerates domestic substitution. If U.S. firms lose China revenue, their R&D capacity can weaken. The conflict is not disappearing; the cost distribution is changing.

2. H200 Refusal Shows the Weakening of the Export-License Carrot

The H200 episode is the clearest signal. The U.S. created a conditional framework for H200 exports to China in January 2026, with oversight and a fee structure. By mid-May, the U.S. had reportedly cleared around 10 Chinese firms to buy H200s, including major internet platforms.

But no chips have shipped. Reporting indicates China has not allowed purchases to proceed. Trump reportedly said China wants to develop its own chips.

This is not just an administrative delay. It means export permission is no longer automatically a carrot. If Beijing refuses the product for strategic reasons, U.S. permission has no commercial value.

For NVIDIA, China H200 revenue should be treated as an option, not as the base case. That makes the next NVIDIA earnings call important. If management confirms that China data-center compute revenue is immaterial or excluded from guidance, the market will need to lower the China reopening option value.

For Korea, the meaning is more nuanced. If Chinese customers do not buy H200s, the immediate HBM linkage is weaker. But if non-China hyperscalers continue to expand AI capex, Korean HBM demand remains strong. The H200 story is less about a direct shipment boost and more about the shape of decoupling.

3. CloudMatrix 384: Losing at the Chip Level, Fighting at the System Level

Huawei’s CloudMatrix 384 shows how China is responding. According to public reports based on SemiAnalysis work, the system uses 384 Ascend chips and can reach roughly 300 PFLOPs of dense BF16 compute, compared with about 180 PFLOPs for NVIDIA’s GB200 NVL72.

But the comparison is not a clean victory. CloudMatrix reportedly consumes roughly 559kW, compared with about 145kW for GB200 NVL72. It is a brute-force design: more chips, more racks, more power, and more optical transceivers.

This tells us two things.

First, China has not achieved chip-level parity. NVIDIA’s efficiency, software ecosystem, CUDA moat, and training stack remain powerful.

Second, China may not need parity for every use case. For domestic policy markets, state cloud, sovereign AI, and “good enough” inference, a less efficient but domestically controlled system can be acceptable.

The most interesting investment implication is optical interconnect. A brute-force cluster needs a much larger communication fabric. CloudMatrix reportedly relies on thousands of high-speed optical transceivers. If weaker chips require more scale-out and scale-up connections, then optical components become a shared bottleneck.

The U.S. stack also moves in the same direction. Rubin, NVL systems, Spectrum-X, and CPO all increase rack-scale optical demand. The U.S. uses optics to improve efficiency; China uses optics to compensate for chip gaps. Different reasons, same direction.

4. ITIF’s Decoupling Model: Korea Is a Short-Term Absorber

ITIF’s November 2025 report gives a useful quantitative frame. In a hypothetical full semiconductor decoupling scenario, U.S. semiconductor firms could lose about USD 77B of China-related sales in the initial year. ITIF estimates that South Korean firms could gain about USD 21B of those lost sales, more than the EU, Taiwan, Japan, mainland China, or other regions.

The memory line is especially important. In ITIF’s table, a large share of the U.S. memory-related loss is absorbed by South Korea. This is why “decoupling is bad for Korea” is too simple.

In the short run, Korea can benefit. The country is not the direct AI accelerator rival to NVIDIA. It supplies memory, HBM, substrates, MLCC, testing, packaging, and other bottleneck components. If U.S. firms lose parts of China while non-China AI capex rises, Korean component suppliers can still benefit.

But the long-run view is less comfortable. The longer decoupling persists, the more China substitutes. Standards fragment. Korean firms may need U.S.-aligned and China-aligned product paths, compliance structures, inventory systems, and qualification cycles. That raises cost and can pressure margins.

The right answer is time-segmented:

• 6-12 months: favorable for Korean memory and AI components.
• 1-3 years: mixed, as China substitution begins to pressure pricing and standards.
• 3-5 years: structurally riskier, because dual supply chains and fragmented standards raise costs.

5. What It Means for Korean Stocks

The first-order beneficiaries remain SK hynix and Samsung Electronics. SK hynix has the cleanest HBM exposure. Samsung Electronics has HBM4, DDR5, eSSD, foundry optionality, and a broader AI-infrastructure platform story. But both are already heavily discovered. For new money, chasing after a large move is less attractive than buying on macro or earnings pullbacks.

The cleaner second-order alpha may sit in the less crowded supply chain.

Optical interconnect is the most interesting. OE Solutions and related optical names should not be framed only as NVIDIA suppliers. They are exposure to a world where both the U.S. and Chinese AI stacks need more high-speed connections.

Samsung Electro-Mechanics is another example. AI servers need MLCCs and FC-BGA substrates regardless of whether the rack is U.S.-standard or China-standard. The company is not the GPU winner, but it is a component bottleneck.

Hana Micron belongs in the structural-improvement bucket. If its Vina pricing structure and overseas margins continue to improve, it is not simply a memory-cycle beta. It is a post-processing margin-upgrade story.

6. Four Market Misreadings

The first misreading is that decoupling is a single big-bang event. It is not. It has been moving layer by layer since the Huawei restrictions: telecom equipment, semiconductor tools, AI chips, AI systems, and now optical links, standards, and software stacks.

The second misreading is that decoupling is uniformly bad for Korea. It is bad over a long horizon if standards fragment and China substitutes. But ITIF’s model shows that in the short run South Korean firms can absorb a meaningful share of U.S. firms’ lost sales.

The third misreading is that China’s rejection of H200 proves self-sufficiency. It does not. China still lags at the chip level. What it shows is a willingness to accept inefficiency in exchange for domestic control.

The fourth misreading is that the summit means decoupling is easing. The summit reduced escalation risk; it did not dismantle the strategic conflict.

7. The Next Catalyst: September and October

The next major window is a potential Xi visit in September and the October anniversary of the prior truce framework. The base case is truce extension: both sides want stability. A tougher scenario is a renewed escalation around export controls, rare earths, or Taiwan language. A true grand bargain remains lower probability.

For investors, the playbook is simple. If the truce is extended, existing Korean semi exposure can remain. If escalation returns, near-term drawdowns can create better entries in memory and AI component names, while optical-interconnect names may become more strategically relevant. If a grand bargain triggers a sharp relief rally, that may be a chance to trim overheated first-order winners.

8. Bottom Line

The Beijing summit did not end the U.S.-China technology conflict. It confirmed that the conflict is being managed.

The H200 episode and CloudMatrix 384 are the real signals. The U.S. can approve exports, but China can refuse them. China may not have the best chip, but it can build a system that is good enough for parts of its domestic market. That is a more expensive equilibrium, not a peaceful one.

For Korea, the answer is not simply bullish or bearish. Decoupling is positive today, mixed tomorrow, and risky later. SK hynix and Samsung Electronics remain the core short-term beneficiaries, but the more interesting medium-term angle may be optical interconnect, AI passive components, substrates, and structural OSAT improvement.

The alpha is not in saying “decoupling is good” or “decoupling is bad.” The alpha is in separating the time horizon.

Sources and Evidence Notes

Fact

• ITIF estimates that in a full decoupling scenario U.S. semiconductor firms could lose about USD 77B of China-related sales in the first year, while South Korean firms could gain about USD 21B.
• Reuters-cited reporting and Tom’s Hardware coverage indicate the U.S. cleared H200 sales to roughly 10 Chinese firms but no chips had shipped as of mid-May.
• Tom’s Hardware and SemiAnalysis-based reporting describe Huawei CloudMatrix 384 as a 384-chip system with roughly 300 PFLOPs BF16 and about 559kW system power.

Inference

• H200 non-shipment is interpreted here as a weakening of the export-license carrot.
• CloudMatrix 384 is interpreted as a system-level workaround rather than a chip-level victory.
• Korea’s exposure is segmented into short-term benefit, medium-term mixed impact, and long-term fragmentation risk.

Main references: ITIF Decoupling Risks report, Reuters H200 licensing report via Investing.com, Tom’s Hardware on China blocking H200 purchases, Tom’s Hardware on CloudMatrix 384, SemiAnalysis CloudMatrix note.

This post is for research and commentary only and is not investment advice. Summit details, H200 shipment status, and CloudMatrix specifications are based on public reporting as of May 17, 2026 KST. ITIF’s numbers are model outputs and may not match actual decoupling outcomes. Scenario analysis and stock implications are the author’s interpretation and may be wrong.