"Stop Stealing from Us!" Claude Explodes... How China's 'Distillation' Technique Upended the US

<1> Chinese AI Upends U.S. Stock Market

On Friday last week, major semiconductor stocks in the United States, including Nasdaq, Nvidia, and Intel, plummeted sharply. Technology stocks in Japan and Taiwan also plunged in tandem. While multiple negative factors combined to drive the drop, Chinese startup Moonshot AI's open-weight model 'Kimi K3,' unveiled right before the sell-off, was identified as the catalyst that poured fuel on the fire. Kimi K3 is the world's largest open-weight AI model, featuring 2.8 trillion parameters and a processing capacity of 1 million tokens.
It excels in long-duration coding and complex tasks, ranking second among 37 models in a comprehensive evaluation by independent assessment organization Vals AI. The gap with Anthropic's latest model in first place was a mere 0.4 points. Although Moonshot itself acknowledged that "overall performance still falls short of top-tier models," it managed to take first place on the frontend coding leaderboard, surpassing the latest U.S. models.

In particular, its price shocked the market: $3 per 1 million input tokens and $15 per output tokens. Based on official pricing, it is 40 to 50% cheaper than GPT-5.6 Sol and about 70% cheaper than Anthropic Fable 5. Moonshot AI plans to release the model's weights by the end of this month, enabling companies and governments with sufficient hardware to install it on their own servers and modify it for their needs. While U.S. companies have poured astronomical amounts of money into developing top-tier models, China has succeeded in delivering a significantly cheaper alternative with a narrow performance gap. Until now, U.S. semiconductor stocks have rallied on expectations that massive amounts of advanced chips would be required for AI development. However, with the emergence of the cost-effective Kimi K3, the market has begun to question whether the astronomical AI investments in the U.S. will truly yield proportionate returns. Kimi K3 touched upon the most sensitive nerve in an overheating AI market.

<2> How Did China Catch Up This Far?

In May, Moonshot AI raised approximately $2 billion in funding—about one-thirtieth of the $65 billion secured by Anthropic in the same month. This clearly demonstrates how smaller Chinese startups are narrowing the performance gap with the U.S. using relatively less capital. Following the 'DeepSeek Shock' triggered by DeepSeek R1 last year, Zhipu AI's GLM-5.2 has recently closed in on U.S. models in coding and AI agent capabilities. Professor Graham Webster of Stanford University, who researches China's open-source ecosystem, assessed that estimates of the lead held by top U.S. models over Chinese counterparts are "converging around six months," adding that "it is not a huge lead." Behind this rapid technological surge, analysts suggest that U.S. semiconductor sanctions paradoxically forced Chinese firms into extreme compute efficiency. For instance, Kimi K3 operates only 16 out of 896 specialized domains required for query processing, reducing compute load and power consumption. In addition, an ecosystem where open technologies are shared, along with support from the Chinese government and tech giants, accelerated the pace of catch-up.

However, the U.S. is convinced there was another shortcut: "adversarial distillation," in which Chinese companies extracted vast amounts of responses from U.S. AI models to train their own. Distillation is a technique that uses a superior large model as a "teacher" to repeatedly train another model on its outputs. By having a U.S. model solve millions of problems, Chinese developers use those questions and answers as textbooks for their own models. While it does not copy the teacher's entire technology, it allows a model to learn a condensed form of problem-solving methods and expression patterns built through immense investment of money and time. In February, Anthropic made these allegations public, citing specific names and figures. It revealed that DeepSeek, Moonshot AI, and MiniMax mobilized about 24,000 fake accounts to conduct over 16 million Q&A exchanges with Claude to siphon off its capabilities. MiniMax accounted for the most with 13 million exchanges, followed by Moonshot AI with 3.4 million—yes, the very company behind Kimi K3, which recently stunned the world.

Anthropic identified Moonshot's targets as agentic reasoning, tool use, coding, and computer control capabilities—coincidentally overlapping heavily with the core strengths touted by K3. The methods were meticulous. One proxy network operated more than 20,000 accounts simultaneously, blending stolen traffic with normal user traffic to evade tracking. Anthropic stated that for DeepSeek accounts, it traced queries back to specific researchers, highlighting evidence that answers evading sensitive topics related to dissidents or party leaders were extracted, raising concerns about censorship learning. In the same month, OpenAI also warned the U.S. Congress that DeepSeek was targeting its models. The individual who raised these allegations personally took the stage at the Aspen Security Forum last week.

[Tarun Chhabra / Head of National Security Policy at Anthropic: "The problem of model distillation is real. I believe it is narrowing America's technological edge. We are currently blocking accounts that facilitate this distillation on a scale of millions every week."]

Chhabra estimated that while the current U.S. AI lead over China is 6 to 9 months, it would have been 1 year to 18 months without distillation. In effect, stolen answer sheets erased half the gap. The CEO of News Corp, who shared the same stage, offered a pointed joke.

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

The White House Office of Science and Technology Policy issued a memo stating that distillation poses a national security threat, and response measures such as sanction listings and exclusion from government procurement began to be discussed. The White House acknowledged that serious internal discussions are underway.

[Karoline Leavitt / White House Press Secretary (on July 24): "Relevant discussions are certainly ongoing. To reiterate, this issue is vital not only for national security but also for maintaining America's continued leadership in AI."]

Anthropic also issued a more fundamental warning: models trained through theft lack the safety guardrails of the originals, and if such models are released as open source, no government will be able to control them.

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

Having narrowed the tech gap through any means necessary, China is now aiming for leadership in the global AI market. China's message is clear: "The U.S. controls technology, but China will open the doors of AI to more nations."

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

It is not just rhetoric. 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 countries over the next five years. In practice, even if Chinese models lag slightly in certain benchmarks, they offer lower prices and higher openness. Once weights are released, they can be customized to match local languages and industries. For emerging economies and small and medium-sized enterprises that struggle to build proprietary AI, Chinese models offering affordability and accessibility may prove more attractive than top-tier performance. By exploiting this niche, Beijing aims to draw global developers and businesses into its ecosystem and expand Chinese technology standards—a strategy dubbed the 'AI Belt and Road.'

<4> U.S. Facing Immediate Threat—A Sudden Shift in Strategy?

To slow China's AI progress, the U.S. has enforced export controls on advanced semiconductors and manufacturing equipment. However, recent developments demonstrate once again that blocking chips alone cannot stop the chase. As a result, the U.S. government has recently begun demanding that countries signing trade agreements adopt export control and investment screening mechanisms similar to its own, complete with monitoring systems to filter out indirect exports. Indicating how explicit these demands are, Jamieson Greer, the United States Trade Representative, publicly expressed frustration with South Korea, Japan, and the EU for hesitating due to sovereignty concerns, while noting that Malaysia, Indonesia, Argentina, and Ecuador complied with U.S. requests.

[Jamieson Greer / U.S. Trade Representative (USTR): "It is harder dealing with Japan, South Korea, and the European Union. Let me be clear. Surprisingly, they are hesitating on this issue. They ask, 'What about our sovereignty?' but I don't really care for that excuse."]

Although Washington claims it is not asking nations to sever all trade ties with Beijing, it is undeniably a demand that access to the U.S. market and frontier tech requires alignment with U.S. economic security standards. Pressure is mounting on Seoul to align its policies with Washington regarding which equipment exports to restrict to China and how far to review Chinese investments in high-tech firms.

<5> South Korea's Choice in a Complex Matrix

South Korea faces three overlapping interests simultaneously on this matter: As a key security and technology ally of the U.S., South Korea is deeply integrated into America's advanced semiconductor supply chains and export control regimes. At the same time, companies like Samsung Electronics and SK Hynix maintain production facilities and major customers in China, meaning tighter controls could heighten business uncertainty in the Chinese market. Furthermore, affordable, self-hosted Chinese open-weight models present an attractive option for South Korean businesses. For South Korea, relying solely on U.S. models brings rising costs and technological dependence, whereas choosing Chinese models raises security and censorship concerns. Adding the U.S. warning that stolen models lack safety guardrails makes the calculus even more complex. Without securing its own proprietary AI models, semiconductor capabilities, and cloud ecosystems, South Korea risks being reduced to merely choosing whose technology to borrow between two superpowers. How far to join U.S. controls, to what extent to utilize low-cost Chinese AI, and how to safeguard its domestic AI and chip industries—South Korea now faces the daunting challenge of resolving these three dilemmas at once.

(Reported and Written by Kim Minjeong | Video by Park Woo-jin | VJ & Video Editing by Kim Hye-joo | Graphics by Yook Do-hyun | Produced by SBS Digital News)
※ Please note: This article was translated by AI and may contain errors.