<1> Chinese AI Shakes Up US Stock Market

Major semiconductor stocks in the United States, including the Nasdaq index, Nvidia, and Intel, fell sharply. Tech stocks in Japan and Taiwan also tumbled alongside them. While multiple negative factors contributed to the decline, the recently unveiled open-weight model "Kimi K3" from China's Moonshot AI was pointed to as the spark that fueled the sell-off. Kimi K3 is the world's largest open-weight AI model, featuring 2.8 trillion parameters and a processing capacity of 1 million tokens.
Exhibiting strong performance in long-duration coding and complex tasks, it ranked second among 37 models in a comprehensive evaluation by independent testing organization Vals AI. The gap between Kimi K3 and the top-ranked latest model from Anthropic was merely 0.4 points. While Moonshot itself acknowledged that "overall performance still falls short of top-tier models," it managed to take the top spot on the front-end coding leaderboard, beating out the latest American models.

Pricing, in particular, sent shockwaves through the market: $3 for input and $15 for output per 1 million tokens. Based on official prices, it is 40 to 50% cheaper than GPT-5.6 Sol and about 70% cheaper than Anthropic's Fable 5. Moonshot AI plans to release the model weights by the end of this month, enabling sufficiently equipped enterprises and governments to host it on their own servers and fine-tune it for their specific needs. While US firms were pouring astronomical sums into developing top-tier models, China succeeded in offering a far more affordable alternative with a minimal performance gap. US semiconductor stocks had previously rallied on expectations that AI development would require vast amounts of advanced chips. However, with the emergence of the cost-effective Kimi K3, the market has begun to question whether the astronomical AI investments in the US will yield proportionate returns. Kimi K3 touched upon the most sensitive nerve in an AI market already embroiled in over-investment debates.

<2> How Did China Catch Up This Far?

Last May, Moonshot AI raised approximately $2 billion in funding—roughly one-thirtieth of the $65 billion secured by Anthropic in the same month. This starkly illustrates how smaller Chinese startups are narrowing the performance gap with the United States using relatively less capital. Following the "DeepSeek Shock" sparked by DeepSeek R1 last year, Zhipu AI's recent GLM-5.2 has also closed in rapidly on US models in coding and AI agent tasks. Professor Graham Webster of Stanford University, who researches China's open-source ecosystem, assessed that estimates of the gap between top-tier US and Chinese models are "converging around six months," adding that "it is not a substantial lead." Analysts suggest that US semiconductor sanctions paradoxically forced Chinese companies to achieve extreme computational efficiency, driving this rapid technical pursuit. For instance, Kimi K3 reduces computational load and power consumption by activating only 16 of its 896 specialized domains required to process a query. An ecosystem that mutually leverages open technology, alongside support from the Chinese government and tech giants, further accelerated this momentum.

However, the US remains convinced that another shortcut was involved: "adversarial distillation," a practice in which Chinese firms allegedly extracted vast volumes of responses from American AI models to train their own. Distillation uses an advanced, large model as a "teacher" to repeatedly train another model on its outputs. By having a US model solve millions of problems, Chinese companies use those questions and answers as textbooks for their own models. While this does not copy the teacher's entire core architecture, it allows them to learn compressed problem-solving techniques and expression patterns that took vast amounts of money and time to develop. Last February, Anthropic made these allegations public, citing real company names and figures. Anthropic asserted that DeepSeek, Moonshot AI, and MiniMax mobilized around 24,000 fake accounts to conduct over 16 million query exchanges with Claude to siphon off its capabilities. MiniMax accounted for the largest share at 13 million exchanges, followed by Moonshot AI at 3.4 million exchanges—the very company that shocked the world with Kimi K3.

Anthropic highlighted that Moonshot targeted agent reasoning, tool usage, coding, and computer control capabilities—areas that overlap significantly with the key strengths advertised for K3. The methods were also sophisticated; one proxy network operated over 20,000 accounts simultaneously, blending stolen traffic into legitimate user traffic to evade detection. Anthropic noted that in the case of DeepSeek accounts, it traced activity back to specific researchers, pointing to evidence that queries were designed to extract answers that evade sensitive topics regarding dissidents or party leaders, suggesting censorship training. In the same month, OpenAI also warned the US Congress that DeepSeek was targeting its models. The executive who raised these concerns spoke directly at the Aspen Security Forum.

[Tarun Chhabra / Anthropic Head of National Security: "The model distillation issue is real. I think it is narrowing the US technological lead. We are currently blocking millions of accounts every week involved in assisting this distillation."]

Chhabra estimated that while the US AI lead currently stands at 6 to 9 months, it would have been 1 year to 18 months without distillation, meaning stolen answer keys erased roughly half of the gap. The CEO of News Corp, who shared the stage, offered a sharp observation.

[Robert Thomson / News Corp CEO: "Isn't it interesting how the word 'distillation' has somehow become synonymous with 'stealing someone else's property'? It used to mean making bourbon or vodka."]

The White House Office of Science and Technology Policy issued a memorandum classifying model distillation as a national security threat, and potential countermeasures such as inclusion on sanctions lists and exclusion from government procurement began to be discussed. The White House acknowledged that serious discussions are underway within the government.

Anthropic also issued a more fundamental warning: models trained on stolen data lack the safety guardrails of the original, and once such models are released as open source, no government will be able to control them.

<3> China's 'AI Belt and Road'

Having closed the technological gap by any means necessary, China is now eyeing leadership in the global AI market. Its messaging is clear: "While the US controls technology, China will open the doors of AI to more nations."

[Xi Jinping / Chinese President (At the World AI Conference in Shanghai on July 17): "Artificial intelligence development should not be a solo performance by a single country, but a symphony of global cooperation."]

These were not mere words. Around the time of the speech, 29 countries, including Russia and Pakistan, signed an agreement to establish the "World AI Cooperation Organization" headquartered in Shanghai, and China promised 5,000 AI training opportunities for developing nations over the next five years. In practice, while Chinese models may lag slightly in certain areas of performance, they offer lower prices and high open-source accessibility. Once weights are made public, they can be adapted to fit each country's language and local industries. For emerging economies and small to medium-sized enterprises struggling to build proprietary AI, Chinese models offering affordability and accessibility can prove more attractive than top-tier performance. China's strategy is to leverage this opening to draw global developers and businesses into its ecosystem and expand Chinese technology standards—a initiative dubbed the "AI Belt and Road."

<4> Fire Under Their Feet: Is the US Making a Sudden Strategic Shift?

The US has relied on export controls for advanced semiconductors and manufacturing equipment to slow China's AI development. However, this recent breakthrough has proven once again that blocking chips alone cannot stop the chase. Consequently, the US government has recently begun demanding that countries signing trade agreements adopt export controls and investment screening mechanisms similar to those of the US, complete with monitoring systems to filter out indirect exports. Highlighting how explicit these demands have become, US Trade Representative Jamieson Greer stated that while countries like Malaysia, Indonesia, Argentina, and Ecuador accepted US demands, he publicly expressed dissatisfaction with South Korea, Japan, and the EU for hesitating on grounds of sovereignty.

[Jamieson Greer / USTR Representative: "Dealing with Japan, South Korea, and the European Union is harder. Let me be clear: surprisingly, they are hesitating on this issue. Each country asks, 'What about our sovereignty?' but I am not particularly fond of that excuse."]

While Washington stated it is not asking allies to sever all transactions with China, the requirement is clear: participation in US economic security standards is necessary to maintain access to US markets and advanced technology. Pressure is mounting on South Korea to align its policies with the US regarding which equipment exports to restrict to China and how far to screen Chinese capital investments in advanced tech firms.

<5> What Is South Korea's Choice in This Complex Matrix?

South Korea finds itself balancing three concurrent interests in this matter. As a security and tech ally of the US, South Korea is deeply integrated into the US advanced semiconductor supply chain and export control framework. Simultaneously, key Korean companies like Samsung Electronics and SK Hynix maintain production facilities and major client bases in China, meaning tightened controls could escalate operational uncertainty in the Chinese market. Furthermore, affordable Chinese open-weight models that can be run on local servers represent an attractive option for Korean businesses. For South Korea, relying solely on US models increases costs and technological dependence, whereas adopting Chinese models brings security and censorship risks. The math becomes even more complicated when considering US warnings that models built on stolen data lack safety guardrails. Without securing an independent AI model, semiconductor, and cloud ecosystem, South Korea risks being reduced to merely choosing whose technology to borrow between the two superpowers. South Korea now faces the difficult challenge of navigating three issues simultaneously: how far to join US controls, to what extent to utilize low-cost Chinese AI, and how to protect its domestic AI and semiconductor industries.

(Reported and Written by Kim Minjeong | Camera by Park Woo-jin | VJ & Video Editing by Kim Hye-ju | Graphics by Yook Do-hyun | Produced by SBS Digital News)