Data Centers Full of U.S. Chips Are Serving Chinese AI Models, Undercutting Washington's Tech Ambitions
Key Takeaways
- •The U.S. Commerce Department issued a rule on Monday closing an export-control loophole that permitted Chinese companies to remotely access advanced AI chips located outside China.
- •U.S.-owned data centers rent capacity to inference providers that serve Chinese large language models, meaning Chinese models often run on American-made chips.
- •China cannot produce enough advanced AI chips to meet domestic demand, so its models increasingly depend on U.S.-controlled computing infrastructure.
- •DeepSeek's latest model is available through at least 16 inference providers headquartered outside China, including Amazon Bedrock.
- •The commentary recommends that Washington avoid banning Chinese models on U.S. clouds and instead support competitive American open-weight models, including through federal procurement.

The Nvidia-Megaspeed saga has drawn attention to a significant gap in U.S. export controls: the ability of Chinese companies to remotely access AI chips located outside of China—a route that bypassed any physical shipment of hardware into China. On Monday, largely in response to the scandal, the U.S. Commerce Department issued a rule closing that loophole. While the long-overdue action can help U.S. AI labs widen the gap between their models and those of their Chinese competitors, a paradox sits at the heart of Washington's AI strategy: the American infrastructure spreading around the world is helping Chinese models gain traction.
The Trump administration wants the United States to dominate AI from chips to models. Central to the president's vision is exporting high-end, U.S.-designed chips to countries around the world. Increasingly, these chips—designed by companies such as Nvidia—are being deployed in data centers across Latin America, the Middle East, and Southeast Asia.
But U.S. leadership in compute infrastructure does not only help American AI model developers. It also helps their Chinese competitors. U.S.-owned data centers filled with U.S. chips still rent capacity out to inference providers—companies that host and run AI models on demand—which use that capability to serve Chinese large language models (LLMs) to users around the world.
In other words, Chinese models are running on U.S. chips, and the American AI infrastructure build-out is counterintuitively helping Chinese model developers compete. That poses a problem for U.S. AI preeminence, which will depend on global consumers and enterprises choosing U.S.-developed LLMs over Chinese ones. The contradiction at the heart of the Trump administration's AI dominance ambitions is that success at one layer of the technology stack is undermining its goals at another.
China's chip bottleneck
When people hear that Chinese models are gaining users, they generally assume those models are running on Chinese chips. That is increasingly the case within China, but the country cannot produce enough advanced AI chips to satisfy its own companies, let alone meet overseas demand.
America's grip on AI infrastructure, meanwhile, is formidable. U.S. companies—Nvidia, Google, AMD, and Amazon—design nearly all deployed data-center AI chips. In cloud services, Amazon, Microsoft, and Google together control more than half of the market, with a growing cohort of U.S. “neoclouds”—specialized providers focused on renting out AI computing capacity—further extending that lead.
Yet the same infrastructure that secures America's advantage in chips and cloud services gives Chinese AI models a platform to spread. Nvidia itself provides resources to help companies deploy open models—including several Chinese models—on its hardware. U.S.-based inference providers such as Together AI, Baseten, and Fireworks serve Chinese models to users on infrastructure powered by U.S. chips. DeepSeek's latest model, for example, is being offered by at least 16 inference providers headquartered outside China, including Amazon Bedrock.
When U.S. infrastructure serves Chinese LLMs, it accelerates those models' global adoption, which benefits China's soft power and generates revenue for its AI labs. Beijing's stated ambition is to be the leading provider of transformative AI capabilities, especially in the Global South, and wider adoption of Chinese LLMs advances that objective.
Chinese deployments also generate direct revenue for Chinese model developers. Moonshot AI requires companies that serve its successful Kimi K3 model and exceed $20 million in annual sales to negotiate a revenue-sharing license, and Moonshot has reportedly asked Microsoft, Amazon, and Google for 30% of their Kimi K3 revenue.
Proprietary vs. open-weight
Who benefits from U.S. computing infrastructure today depends largely on whether it serves proprietary or open-weight models. Infrastructure serving proprietary LLMs—such as Anthropic's Claude or OpenAI's GPT—almost exclusively benefits American developers, because they make the best ones. Infrastructure serving open-weight LLMs, which companies can customize for specialized tasks and run locally, tends to benefit Chinese developers: users gravitate toward the best available models, and in the open-weight market, those are Chinese.
Compete rather than ban
The commentary argues the Trump administration should not go so far as to ban U.S. cloud providers from allowing non-Chinese companies to serve Chinese LLMs. Doing so would incur too high a cost—unjustifiably hindering companies in the United States and allied countries seeking to adopt AI—while failing to deliver sufficient national security benefits. It could also invite Chinese retaliation.
Instead, the United States should compete. U.S. companies such as Meta and Thinking Machines are already developing open-weight models. If those models become more cost-efficient than their Chinese competitors, more U.S. computing infrastructure would serve U.S. models. That shift could limit both the revenue Chinese developers earn from overseas deployments and the soft power China gains from its models' global reach.
The administration is therefore right, in the commentary's view, to encourage the development of competitive U.S. open-weight models. It should identify government uses well-suited to open-weight models and direct federal procurement toward American open-weight offerings for those uses. Finally, as stronger AI regulation becomes necessary, the administration should design those rules carefully to avoid stifling U.S. open-weight model development.
Washington cannot assume that exporting U.S. infrastructure will automatically lead to the export of U.S. models. If it wants to succeed on both fronts, it needs a deliberate strategy for open-weight models—and two concrete items to follow are how the new remote-access rule is enforced, and whether the revenue-sharing terms Moonshot has reportedly requested from Microsoft, Amazon, and Google are ultimately agreed.
The opinions expressed in Fortune.com commentary pieces are solely the views of their authors and do not necessarily reflect the opinions and beliefs of Fortune. This story was originally featured on Fortune.com.