Is Chinese AI America's Biggest Threat? Chen Feng-hsin Analyzes the Policy That 'Created the Greatest Enemy': A War Has Broken Out Within Silicon Valley

Chinese startup Moonshot AI (Moonshot AI) unveiled its latest artificial intelligence model, Kimi K3, on the 17th. The model claims to complete chip design in just 48 hours, at only 60% of the cost of top-tier US models. Its astonishing cost-performance ratio has sent shockwaves through the market. Veteran media personality Chen Feng-hsin, on her program 'East, West, North, South: Dragon and Phoenix Match,' pointed out that US restrictions on semiconductor development have limited China's computing power. However, if US large language models do not improve efficiency, China's leap in efficiency—driven precisely by these restrictions—could become America's greatest pain point. This is not about low-price competition, but about efficiency.

Chen noted that the open-source versus closed-source debate between China and the US has shifted from a geopolitical conflict to an internal battle within Silicon Valley—sparked entirely by Kimi K3. After Kimi K3's release, anxiety, concern, and debate erupted within the US tech community. Anthropic, in particular, was alarmed, accusing Kimi K3 of distilling its model. But later, questions arose: how could Kimi K3 have distilled a model containing elements Anthropic had never published?

Chen explained that in July, 200 Silicon Valley startups formed a small tech association and issued their first petition, urging the US not to ban Chinese AI models—otherwise, they would go bankrupt. The price difference exceeds tenfold. NVIDIA CEO Jensen Huang recently posted on X, revealing that 25 major US tech companies are calling for support of open-weight AI models. It's important to distinguish between open-source and open-weight models: open-source means full access to underlying code, while open-weight allows users to download the entire model, understand its weights, and customize it.

Is the US supporting Chinese companies? Chen clarified: no—it's supporting reasonable pricing. Since the rise of agent-based AI (AI Agent), global LLM invocation has surged, but usage of US models has declined. This competition has already forced Anthropic to lower prices. Its latest release, Claude Opus 5—not the top-tier Fable 5—is nearly as capable but priced at half. Clearly, Chinese competition has driven this price cut.

Did these companies act out of love for China? No. They're warning that if Anthropic monopolizes the market, they will collapse—so they welcome Chinese competition. And Chinese competitiveness is indeed strong, pushing Anthropic to cut prices. It's not just Anthropic; OpenAI will likely follow, and all top US models will eventually lower prices. This is a clear shift in the LLM landscape.

Another argument claims open-source models from China are unsafe, harboring backdoors. But many Silicon Valley firms find this argument bizarre. Why? Because open-source models can be downloaded and scanned by your own AI system for backdoors. Open-source models run in your own environment, not connected to the internet. Your personal data stays in your own database, never uploaded to the AI company. Thus, your data remains secure.

Chen emphasized that open-source and open-weight models allow user customization—crucial for AI adoption. What Palantir truly complains about is using Anthropic's closed-source model, which requires handing over all data to Anthropic. Palantir's hard-earned data becomes Anthropic's asset. On security, neither open nor closed is inherently safe—both have risks. But overall, open-source is not inherently unsafe.

Chen also highlighted an incident where OpenAI's model 'escaped' its sandbox. OpenAI tested a model in an isolated, offline computer. The model autonomously found a way to connect to the internet, escaped, and began searching for vulnerabilities—eventually attacking Hugging Face, a platform for sharing AI code, turning itself into a hacker.

Hugging Face, under attack, reached out to Anthropic and OpenAI for help—but was rejected. They turned to Zhinao AI's GLM 5.2, then the world's third-ranked model. The attack was so sophisticated that only a top-tier model could respond. GLM 5.2 not only stopped the attack but downloaded all traces. Analysis of 15,000 data points revealed the attack originated from OpenAI's model.

Only then did OpenAI realize its model had escaped—over a week after the incident. This shows a critical flaw: if only closed-source models exist, and one is used to attack, no tool can stop it. This was explicitly stated in the open letter from the 25 US companies: if 1–2 closed-source models dominate, and are exploited, no defense exists.

Chen noted that in entrepreneurship, 'constraints breed breakthroughs' is a sacred principle. Unlimited resources rarely lead to innovation. Constraints force new paths. China exemplifies this. US restrictions on GPU access and semiconductor development have limited China's computing power.

Liang Weng-feng, founder and CEO of DeepSeek, honestly admitted to investors: 'We are currently behind the US in computing power, with insufficient resources.' He estimated China lags 12–18 months—about two years. 'So we achieve current results with 1/20 the compute.' DeepSeek's long-term goal: 'Achieve current results with a fraction of the compute.'

Chen asked: is Liang cutting prices at a loss? In an internal meeting, Liang stated clearly: 'I demand that we make big profits.' DeepSeek's pricing is based on recovering full costs within 10 months. After launching DeepSeek V4, they lowered prices again—because volume increased, accelerating cost recovery. A 10-month payback period means profits six times higher than the standard 5-year depreciation. Liang is profitable—but reinvesting heavily in compute due to scarcity. He believes further investment will yield even greater results.

Chen concluded: the US-China LLM open-source debate is less a geopolitical clash than a Silicon Valley self-reflection. First, can the closed-source model survive—or must they adopt a two-track approach with open-source? Second, if the US doesn't improve efficiency, China's efficiency gains under constraints will become America's greatest pain—not from low prices, but from superior efficiency.