Skip to content
LIVE
The Executives BriefThe Executives Brief

The U.S. built the AI infrastructure. Chinese models are using it to win.

Commerce just closed the remote-access loophole, but U.S. data centers packed with Nvidia chips still rent capacity to providers serving DeepSeek and Kimi K3.

ByAbdullah Al-OtaibiBusiness Desk, The Executives Brief
·3 min read
The U.S. built the AI infrastructure. Chinese models are using it to win.
Executive summary

The U.S. Commerce Department issued a rule Monday closing the loophole that let Chinese companies remotely tap AI chips located outside China, the gap exposed by the Nvidia-Megaspeed saga. The catch for decision-makers: U.S.-owned data centers full of U.S. chips still rent capacity to inference providers that serve Chinese LLMs, so compute dominance is quietly spreading Chinese models rather than American ones.

Washington just closed one door and left another wide open. On Monday, largely in response to the Nvidia-Megaspeed scandal, the U.S. Commerce Department issued a rule shutting the loophole that let Chinese companies remotely access AI chips sitting outside China. That is a real fix for a real gap. But it does not touch the paradox underneath: the same U.S. infrastructure that is supposed to lock in American AI dominance is also the platform Chinese models are using to reach the world. U.S.-owned data centers, filled with U.S. chips, still rent capacity to inference providers, and those providers serve Chinese large language models to users everywhere.

Start with how lopsided the hardware layer is. China cannot produce enough advanced AI chips to satisfy its own companies, let alone meet overseas demand. America's grip is close to total: Nvidia, Google, AMD, and Amazon design nearly all deployed data-center AI chips, and in cloud services Amazon, Microsoft, and Google together control more than half the market, with a growing cohort of U.S. neoclouds extending that lead further. The Trump administration's plan leans hard on this advantage, exporting high-end U.S.-designed chips to data centers in Latin America, the Middle East, and Southeast Asia. The bet is that owning the compute layer means owning the models layer. The evidence so far says otherwise.

Nvidia itself provides resources to help companies deploy open models, including several Chinese ones, 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 is being offered by at least 16 inference providers headquartered outside China, Amazon Bedrock among them. Every one of those deployments is a Chinese model getting faster, cheaper, and more familiar to global users, running on American silicon. When people hear that Chinese models are gaining users, they usually picture Chinese chips. Inside China, that is increasingly true. Everywhere else, it often is not.

That matters for two reasons: soft power and money. 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 directly. It also pays. Moonshot AI requires companies that serve its 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. U.S. compute, in other words, can end up funding Chinese AI labs and extending Chinese influence.

The split comes down to model type. Infrastructure serving proprietary LLMs such as Anthropic's Claude or OpenAI's GPT almost exclusively benefits American developers, because they build the best ones. Infrastructure serving open-weight LLMs benefits Chinese developers, because users tend to pick the best available model, and in the open-weight market those are Chinese. That is the contradiction at the heart of the Trump administration's AI strategy: success at the chip and cloud layer is undermining its goals one layer up.

The authors argue the answer is not a ban. Barring U.S. cloud providers from letting non-Chinese companies serve Chinese LLMs would cost too much, unjustifiably hindering U.S. and allied companies trying to adopt AI, while failing to deliver enough national security benefit, and it could invite Chinese retaliation. Their prescription is to compete. Meta and Thinking Machines are already building open-weight models; if those become more cost-efficient than their Chinese rivals, more U.S. computing infrastructure would serve U.S. models, crimping both the revenue Chinese developers earn overseas and the soft power China gains from global reach.

That leads to concrete policy asks. The administration should encourage competitive U.S. open-weight models, identify government uses well-suited to them, and direct federal procurement toward American open-weight models for those uses. As stronger AI regulation becomes necessary, it should be designed carefully so it does not stifle U.S. open-weight development. The closing warning is the one worth taping to the wall: Washington cannot assume that exporting U.S. infrastructure will automatically export U.S. models.

For operators and boards, the takeaway is that the AI stack is not a single bet. Owning the chips and the clouds does not guarantee owning the demand. Any company building on open-weight models should know where its inference is actually running and whose license terms it is signing, because revenue-share clauses like Moonshot's turn a procurement decision into a geopolitical one. And any U.S. firm betting its strategy on infrastructure dominance should stress-test the assumption that the models riding on that infrastructure will be American.

Executive ActionsLocked

This story's Key Insights and Take-aways are locked.

Create a free account to unlock Executive Actions for one credit.

Register to Unlock

Always free for Executives Club members. Join the Club

More in Business