The Trump administration is divided over whether to wall off Chinese AI models as corporate adoption accelerates

Chinese AI Draws U.S. Companies. Washington Is Divided.

Airbnb and Cursor are deploying AI models built by Chinese developers. Inside the Trump administration, one faction wants to block them. Another argues the market has already decided.

Moonshot AI released its Kimi K3 model in July, and Trump administration officials revived proposals to restrict U.S. companies from using Chinese-developed AI models, Axios reports 1. The proposals cite national security risks from models built by DeepSeek and Moonshot AI 2.

The fight has two camps with incompatible premises. A national security faction argues that Chinese AI systems embedded in American corporate infrastructure create unacceptable risk and wants to wall off models from DeepSeek and Moonshot AI 1. A pro-business contingent sees companies adopting cheaper Chinese alternatives and argues for leaving the decision to the market 1.

The corporate pull is economic, not ideological. Chinese models like DeepSeek's and Moonshot's Kimi K3 compete on price, and firms that run inference at scale feel every cent of that difference 1. The difference between a domestic model and a cheaper Chinese alternative is measured in margins, and margins are what determine which products survive.

Meanwhile, the decoupling runs in both directions. The New York Times reports that both Washington and Beijing are weighing new barriers around their respective technology sectors 3. China is restricting what American technology can operate within its borders. The United States is debating what Chinese technology can enter its own.

AI models do not behave like the goods Washington has spent decades learning to control. Semiconductors, telecom equipment, and weapons components move through supply chains with customs checks, shipping manifests, and export licenses. An open-weight AI model moves through a download. A company can pull model weights from a public server, fine-tune them internally, and deploy the result behind its own infrastructure without any transaction that resembles an import.

If the national security faction cannot intercept the download, its remaining options narrow. Prohibiting U.S. companies from deploying Chinese models after the fact would require a compliance regime: auditing corporate AI infrastructure, requiring firms to disclose which models they run, and verifying they are not using restricted ones. But open-weight models can be copied, renamed, and fine-tuned into derivatives that bear little resemblance to their original release. The more restrictions the government imposes, the harder compliance becomes to verify, because the thing being regulated transforms the moment a developer touches the weights. The practical endgame is that the pro-business faction wins by default. The cost of building an enforcement apparatus for downloadable software exceeds the political will to impose one, and every month without restrictions is another month of corporate adoption that becomes harder to unwind.

The New York Times frames the broader picture as mutual fortification, with both governments building walls around their technology ecosystems simultaneously 3. But the walls have different architectures. Beijing restricts American technology within its borders by controlling the domestic network. Washington has to convince or compel American companies to stop adopting technology that is cheaper, capable, and already accessible.

Airbnb and Cursor will keep deploying the models that deliver the best value. Moonshot and DeepSeek will keep releasing new ones. And the distance between what Washington wants to restrict and what it can actually restrict will keep widening.