The most widely adopted open AI models in the world are running on American chips inside American-built data centers, and Washington's trillion-dollar push to own AI infrastructure is quietly handing its biggest rival the distribution it wants.
Nearly every data-center AI chip in service anywhere was designed by Nvidia, Google, AMD or Amazon, and Amazon, Microsoft and Google control more than half of the cloud market, with a growing roster of U.S. neoclouds stretching that lead further.
The catch is what runs on top of that stack. In the open-weight market, where customers download a model and host it themselves, the strongest options are Chinese, and users tend to pick whatever works best. DeepSeek's newest model is offered by at least 16 inference providers headquartered outside China, Amazon Bedrock among them. Together AI, Baseten and Fireworks all serve Chinese models on American silicon, and Nvidia itself hands out tools and engineering help to get open models running on its hardware.
Every one of those deployments does Beijing two favors at once. It pushes Chinese models into the Global South, where China is bidding to become the default supplier of transformative AI, and it sends money back to Chinese labs. Moonshot AI requires companies serving its Kimi K3 model above 20 million dollars in annual sales to negotiate a revenue-sharing license, and has reportedly asked Microsoft, Amazon and Google for 30 percent of their Kimi K3 revenue.
The dividing line is the license, not the hardware. Infrastructure running closed systems such as OpenAI's GPT or Anthropic's Claude benefits American developers almost exclusively, because they build the best closed models. Infrastructure running open weights flows mostly to Chinese labs. Washington cannot expect exported chips to automatically mean exported American models.
The fix is competition, not prohibition. Barring U.S. cloud providers from hosting Chinese open-weight models would raise costs for American and allied companies, invite retaliation, and buy little security in return. Meta and Thinking Machines are already building competitive open-weight models, and if American versions become cheaper to run, more of the world's compute will serve American software by default. Federal procurement directed toward domestic open-weight models for suitable government work, plus regulation careful enough not to smother them, would push the same way.
Dominance at one layer of the stack is currently eroding the goal at another, and only a deliberate open-weight strategy closes that gap.