NVIDIA — A major loophole in U.S. export controls comes to light: AI chips located outside of China are being remotely accessed by Chinese companies. On Monday, in response to this scandal, the U.S. Department of Commerce issued a rule to close this loophole.
However, despite this belated measure helping the US AI laboratories to further widen the gap with their Chinese competitors in terms of models, there is still a paradox at the core of Washington's AI strategy: the infrastructure that the US has established globally is actually helping Chinese models gain market recognition.
The Trump administration hopes that the United States will dominate everything from chips to models AI. One of the keys to the president's vision is to export high-end, American-designed chips to countries around the world. An increasing number of these chips – designed by companies such as NVIDIA – are being deployed in data centers in Latin America, the Middle East, and Southeast Asia.
However, the United States' leadership in computing power infrastructure not only benefits American AI model developers but also helps their Chinese competitors.
Today, data centers that are filled with American chips and owned by the United States still rent out their computing power to inference service providers, who in turn use this capacity to offer services related to China's large language models ( LLM ) to users around the world.
Chinese models are running on American chips, and ironically, the construction of America's AI infrastructure is actually helping Chinese model developers to compete. This poses a threat to America's leading position in the field of AI, as this leadership relies on global consumers and businesses choosing American-developed LLM over Chinese LLM.
The contradiction in the authoritarian dreams of the Trump administration AI lies in the fact that success at one level of the technology stack may undermine its goals at another level.
When people hear about the growth in users of Chinese models, they usually assume that these models run on Chinese-made chips. Although this is becoming more and more common within China, the reality is that China is not able to produce enough advanced AI chips to meet the needs of domestic enterprises, let alone those of overseas markets.
At the same time, the United States has extremely strong control over the infrastructure of AI. American companies such as NVIDIA, Google, AMD, and Amazon have designed almost all the chips used in deployed data centers. In the field of cloud services, Amazon, Microsoft, and Google together control more than half of the market, and a growing group of American "neoclouds" firms have further expanded this leading advantage.
However, the same infrastructure that supports the United States' advantages in the fields of chips and cloud services also provides a platform for the dissemination of China's models. NVIDIA itself provides resources to help companies deploy open-source models on its hardware, including several Chinese models. Model inference service providers based in the United States, such as Together AI, Baseten, and Fireworks, also offer Chinese models to users on infrastructure driven by American chips. For example, DeepSeek's latest models are being provided by at least 16 inference service providers based outside of China, including Amazon Bedrock.
When American infrastructure is used to provide China's LLM, this will accelerate the adoption of these models globally, thereby enhancing China's soft power and generating revenue for its AI laboratories. Beijing has openly stated its goal of becoming a leading provider of transformative AI capabilities, especially in the global South. The wider adoption of China's LLM will help to advance this goal.
The deployment of China's LLM will also bring direct income to Chinese model developers. For example, Moonshot AI requires companies that provide its successful models Kimi K3 and have annual sales exceeding 20 million US dollars to negotiate revenue sharing licenses; it is reported that Moonshot has requested Microsoft, Amazon, and Google to pay a 30% split of the revenue from Kimi K3.
Today, who benefits from the computing infrastructure in the United States depends on whether it serves proprietary models or open-source models. On one side are proprietary models such as those developed by Anthropic, Claude, or OpenAI. Providing infrastructure for these proprietary models largely benefits American developers, as they create the best models. On the other hand, there are open-source models that companies can customize for specific tasks and run locally. Providing infrastructure for open-source models benefits Chinese developers, as users tend to choose the best models available, and in the open-source model market, these models often come from China.
However, the Trump administration should not go so far as to prohibit American cloud service providers from allowing non-Chinese companies to provide certain services to China. The cost of doing so is too high—it will unreasonably hinder American and its allies' enterprises from adopting certain technologies—and it will not bring sufficient benefits to national security. Moreover, it may also provoke retaliation from China.
On the contrary, the United States should respond through competition. American companies such as Meta and Thinking Machines are already developing open-source weight models. If their cost efficiency is higher than that of their Chinese competitors, more American computing power infrastructure will be used to support American models. This shift could limit the income that Chinese developers earn from overseas deployments and also weaken China's soft power gained through the global dissemination of these models.
Therefore, it is correct for the Trump administration to encourage the development of competitive American open-source weighting models. The government should identify suitable government uses for these open-source weighting models and direct federal procurement towards these applications. Lastly, as stronger AI regulations become necessary, the government should carefully design relevant rules to avoid stifling the development of American open-source weighting models.
Washington cannot assume that exporting American infrastructure will automatically lead to the export of the American model. If it wants to succeed in both areas, it needs a clear strategy targeting open-source weight models.
The views expressed in the articles commented on by Fortune.com represent only the author's own opinions and do not necessarily reflect the views and positions of Fortune.
Jacob Feldgoise is a Senior Data Research Analyst at the Center for Security and Emerging Technologies at Georgetown University ( CSET ). Hanna Dohmen is a Senior Research Analyst at CSET. Sam Bresnick is a researcher at CSET and Andrew W. Marshall is also a researcher.












