The U.S. wants to contain China’s AI. Silicon Valley keeps using it

The U.S. government and major American tech firms are increasingly concerned about the rapid advancement of artificial intelligence in China, fearing that Beijing is using advanced techniques such as model distillation to bridge the gap with Western counterparts. However, this apprehension appears to be at odds with the reality of the global AI landscape, where collaboration and knowledge sharing transcend geopolitical divides. Recent developments highlight both the strategic concerns of American officials and the complex interplay of innovation across borders. In late September, Tarun Chhabra, Anthropic’s chief national security officer and a former architect of U.S. export controls under President Joe Biden, warned that Chinese AI developers were leveraging American models to enhance their own capabilities. He specifically named Zhipu, a prominent Chinese AI firm, as one of several entities suspected of engaging in model distillation, a process where a smaller model learns from a larger, more powerful one. This practice, Chhabra argued, poses a national security risk, as it allows Chinese researchers to bypass the need for expensive and time-consuming independent development. Yet just hours later, the narrative took an unexpected turn. Mira Murati, the former chief technology officer of OpenAI, launched her own venture called Thinking Machines, which announced that its first foundation model had been developed using components from Chinese models. Specifically, the company cited DeepSeek-V3 and Moonshot AI’s Kimi K2 as key influences. This admission underscores a growing trend in the AI industry: the open-weight model community, which favors transparency and accessibility, is building upon the work of others, regardless of geographic origin. This situation creates a paradox. On one hand, American authorities are raising alarms over the potential misuse of U.S.-developed AI technologies by Chinese actors. On the other, Silicon Valley startups are actively integrating insights and tools from China’s expanding open-source ecosystem. The contrast highlights the evolving nature of AI development, where innovation often flows in multiple directions rather than being strictly controlled by any single nation. Model distillation, while controversial, is not a novel concept. It has long been a standard practice in machine learning, where a large, complex model serves as a “teacher,” generating outputs that train a smaller, more efficient “student” model. Over the past few years, however, the scope of distillation has broadened significantly. Today, it encompasses not only model compression but also the use of synthetic data generated by one system to refine another. This means that advancements in one lab can quickly influence the trajectory of research elsewhere. Leading labs around the globe, from OpenAI and Anthropic in the United States to DeepMind and Google in Britain, and from Alibaba and Tencent in China to Moonshot and DeepSeek in Asia, are all engaged in similar practices. Each publishes research detailing methods of synthetic data generation, reasoning distillation, and other related techniques. What sets them apart is not whether they employ these strategies, but who holds the dominant position in the teacher-student relationship. As AI models grow more sophisticated, the value of their outputs increases exponentially. This has created a financial and technical incentive for companies to access high-quality training data, even if it means tapping into external resources. As a result, firms like OpenAI and Anthropic have taken steps to secure their models, implementing stricter API controls, behavioral monitoring, and detection mechanisms aimed at preventing large-scale output harvesting. Despite these efforts, the recent revelations suggest that the lines between innovation and imitation are becoming increasingly blurred. American firms, despite their public stance against unauthorized use of their models, are quietly adopting elements from China’s open-weight ecosystem. This dynamic reflects a broader shift in the AI industry, one where competition and cooperation coexist, and where the pace of progress continues to outstrip traditional notions of intellectual property and national sovereignty.