US Chip Export Boom Is Quietly Serving Chinese AI Models
A new Commerce Department rule closes the remote-access loophole, but U.S. data centers keep serving Chinese LLMs like DeepSeek and Kimi K3 — and sending revenue back to Beijing's AI labs.
By Grace Kim
3 min read

What's News
- On Monday the U.S. Commerce Department issued a rule closing the remote-access loophole for AI chips located outside China.
- DeepSeek's latest model is offered by at least 16 inference providers headquartered outside China, including Amazon Bedrock.
- Moonshot AI reportedly asked Microsoft, Amazon, and Google for 30% of their Kimi K3 revenue.
- Amazon, Microsoft, and Google control more than half of a cloud market that passed $143 billion in Q2.
- Nvidia, Google, AMD, and Amazon design nearly all deployed data-center AI chips.
At least 16 inference providers headquartered outside China — including Amazon Bedrock — currently serve DeepSeek's latest model on infrastructure built with U.S.-designed chips. That single figure exposes the paradox at the center of Washington's AI strategy: the global build-out of American compute is helping Chinese models spread.
On Monday, largely in response to the Nvidia-Megaspeed scandal, the U.S. Commerce Department issued a rule closing the remote-access loophole in export controls — Chinese companies remotely accessing AI chips located outside China. It was a long-overdue fix. But it addresses only part of a bigger problem.
What does the new rule miss?
The Trump administration wants U.S. dominance across the AI stack, from chips to models. Its plan leans on exporting high-end, U.S.-designed chips — silicon from Nvidia and others — into data centers across Latin America, the Middle East, and Southeast Asia.
Here's the catch: U.S.-owned data centers filled with those chips rent capacity to inference providers, which use it to serve Chinese large language models to users worldwide. Chinese models run on U.S. chips, and the U.S. infrastructure build-out is counterintuitively helping Chinese developers compete.
That matters because U.S. AI preeminence depends on global consumers and enterprises choosing American LLMs over Chinese ones. Success at one layer of the stack is undermining the administration's goals at another.
How dominant is U.S. infrastructure, really?
Very. U.S. companies — Nvidia, Google, AMD, and Amazon — design nearly all deployed data-center AI chips, according to Epoch AI data. In cloud services, Amazon, Microsoft, and Google together control more than half of a market that passed $143 billion in Q2, per Synergy Research, with a growing cohort of U.S. "neoclouds" extending that lead.
Yet the same infrastructure that secures America's chip and cloud advantage gives Chinese models a distribution platform. 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 on U.S.-chip-powered infrastructure.
The result: Chinese LLM deployments generate direct revenue for China's AI labs and advance Beijing's stated ambition to lead in transformative AI, especially in the Global South.
Where does the money flow?
The licensing terms are concrete. Moonshot AI requires companies serving its Kimi K3 model that exceed $20 million in annual sales to negotiate a revenue-sharing license. Moonshot has reportedly asked Microsoft, Amazon, and Google for 30% of their Kimi K3 revenue, according to Bloomberg and Reuters reporting from August and September 2026.
Who benefits from U.S. compute depends on the model type. Infrastructure serving proprietary LLMs — Anthropic's Claude, OpenAI's GPT — almost exclusively benefits American developers, because they make the best ones. Infrastructure serving open-weight LLMs benefits Chinese developers: users pick the best available model, and in the open-weight market, those are Chinese.
Should Washington ban it?
The argument against a ban is straightforward: blocking U.S. cloud providers from serving Chinese LLMs for non-Chinese companies would cost too much, unjustifiably hindering AI adoption by U.S. and allied companies while delivering insufficient national security benefit — and it could invite Chinese retaliation.
The alternative is competition. Meta and Thinking Machines are already developing open-weight models. If U.S. open-weight models become more cost-efficient than Chinese rivals, more U.S. infrastructure would serve U.S. models — cutting both the overseas revenue Chinese developers earn and the soft power China gains from global model reach.
The administration's AI Action Plan already encourages competitive U.S. open-weight development. The next steps, per the argument: identify government use cases suited to open-weight models, direct federal procurement toward American open-weight models for those uses, and design future AI regulation so it doesn't stifle U.S. open-weight development.
Washington can't assume that exporting U.S. infrastructure will automatically export U.S. models. Winning on both fronts takes a deliberate strategy for open weights, not just chips.
Source: fortune.com
More from Grace Kim
Show full bio
Market editor covering people, launches and day-to-day practice in entrepreneurship, startups and venture funding at Business Bearings.
2,582 articles