US data centers filled with American chips are serving Chinese AI models, analysts say
Three analysts from Georgetown University's Center for Security and Emerging Technology (CSET) argue in a Fortune commentary that data centers built with American chips are helping Chinese AI models reach users around the world. The piece was published on November 21, 2026.
The authors, Jacob Feldgoise, Hanna Dohmen, and Sam Bresnick, say this creates a contradiction in the Trump administration's plan to dominate artificial intelligence. The US exports advanced chips and builds computing infrastructure abroad, but that same infrastructure can be used to deliver Chinese models to global customers.
The main points from the commentary
- The US Commerce Department issued a rule on Monday, November 19, 2026, that closes a gap in export controls. It addresses Chinese companies remotely accessing AI chips located outside China.
- The rule followed attention on what the authors call the Nvidia-Megaspeed saga, which highlighted that Chinese firms could use American chips hosted abroad.
- US-owned data centers rent capacity to inference providers, the companies that run AI models for users. Some of those providers use the capacity to serve Chinese large language models (LLMs), which are AI systems trained to understand and generate text.
- Nvidia provides resources to help companies deploy open models on its hardware, including several Chinese models.
- DeepSeek's latest model is offered by at least 16 inference providers headquartered outside China, including Amazon Bedrock, according to data the authors cite from Artificial Analysis.
- Moonshot AI reportedly asked Microsoft, Amazon, and Google for 30% of revenue from serving its Kimi K3 model. Bloomberg reported the company requires providers exceeding $20 million in annual sales from the model to negotiate a revenue-sharing license.
Why the authors say this undercuts US goals
The commentary states that the Trump administration wants the US to lead AI at every level, from chips to models. US-designed chips from companies like Nvidia are being placed in data centers across Latin America, the Middle East, and Southeast Asia.
But the authors argue that American dominance in hardware and cloud services does not automatically translate into dominance in AI models. They note that US companies design nearly all deployed data-center AI chips, and that Amazon, Microsoft, and Google together control more than half of the cloud market. Chinese companies, meanwhile, cannot produce enough advanced AI chips to meet even their own domestic demand, according to the authors.
The piece draws a line between two kinds of models. Proprietary models, such as Anthropic's Claude or OpenAI's GPT, are controlled by their makers and mostly benefit American developers. Open-weight models, which companies can download, customize, and run on their own systems, are currently led by Chinese developers in quality, the authors claim. So infrastructure serving open-weight models tends to benefit China.
The authors say wider adoption of Chinese models generates revenue for Chinese AI labs and supports Beijing's stated goal of leading in AI, especially in the Global South.
What the analysts recommend instead of a ban
The authors argue against banning US cloud providers from letting non-Chinese companies serve Chinese models. They say such a ban would cost too much, would hinder companies in the US and allied countries that want to adopt AI, might not deliver enough national security benefit, and could invite retaliation from China.
Instead, they recommend competing. They point to US companies like Meta and Thinking Machines that already build open-weight models. If those models become more cost-efficient than Chinese ones, more US infrastructure would serve American models, cutting Chinese developers' overseas revenue and reducing the soft power China gains from global model adoption.
They also recommend that the administration direct federal procurement toward American open-weight models for suitable government uses, and design any future AI regulation carefully so it does not slow US open-weight development. They cite the administration's own AI Action Plan from July 2025, which encourages competitive US open-weight models.
What is confirmed and what is opinion
The Commerce Department rule of November 19, 2026, is confirmed through the Federal Register document cited in the piece. Revenue figures and requests from Moonshot AI come from Bloomberg and Reuters reporting cited by the authors. The claim about DeepSeek's providers comes from Artificial Analysis data cited in the commentary.
The broader argument, that US infrastructure is undermining US AI goals, is the opinion of the three CSET authors. Fortune notes that commentary pieces reflect the views of their authors, not the publication.