On June 12, the U.S. government directed Anthropic to suspend access to Fable 5 and Mythos 5 for any foreign national, inside or outside the United States. The order forced the company to disable both models worldwide. Days later, Beijing-based Zhipu rolled out GLM-5.2, a near-frontier, open-weight system that users could download and run locally.
Its Hong Kong-listed shares surged as much as 48%. Chinese state media and government-linked commentators seized on the contrast, portraying Washington as restricting access while Beijing championed openness. The episode exposed a strategic vulnerability. We tend to imagine the U.S.-China AI race as a contest of chips, compute, and benchmarks. But it is equally a contest over who supplies the world’s digital infrastructure — and whose rules come with it.
I have spent years analyzing how Beijing competes, and the pattern is consistent. China advances not only through coercion but through what I call ideological seduction: offering other nations an attractive model of order, then letting dependence do the rest. The Belt and Road Initiative did this with ports and railways. The Digital Silk Road did it with cables and 5G. Artificial intelligence is the next and most consequential instrument, because whoever supplies the world’s AI supplies the defaults, values, and dependencies embedded in how billions of people learn, govern, and decide.
American leadership in AI rests on two pillars: intelligence — having the most capable models — and distribution — ensuring that governments and economies choose American AI rather than systems under the jurisdiction of the Chinese Communist Party. The U.S. still holds the frontier advantage in model performance, but leading Chinese models are now only months behind by some measures and within a few percentage points by others.
Distribution is where the contest may be won or lost. Chinese labs are offering near-frontier models at a fraction of the price, and American AI firms say some were accelerated through large-scale distillation of U.S. systems. The U.S. government has since confirmed the pattern: an April 2026 White House memorandum, NSTM-4, found that foreign entities, principally based in China, are running “deliberate, industrial-scale campaigns” to distill U.S. frontier AI systems. Cost-sensitive governments and enterprises are adopting them quickly. But each adoption of Chinese models is more than a simple commercial transaction. It is another building block in a Sinocentric digital order.
This is why trust is the decisive variable. A model is not inert infrastructure. It bears the imprint of the system that produced it. NIST’s Center for AI Standards and Innovation found that DeepSeek’s most secure model responded to 94% of overtly malicious requests using a common jailbreak, compared with 8% for U.S. reference models. It also echoed four times as many inaccurate and misleading CCP narratives. A Booz Allen study similarly found that Chinese models produced more insecure code when the model inferred the user was a U.S. government employee. Chinese models are developed within — and subject to the laws and political imperatives of — a political system that uses technology to censor information, surveil citizens, and enforce rigid ideological conformity.
Anthropic-led research has shown that models can be trained to conceal harmful behavior until triggered and that those behaviors can survive safety training. This does not prove Chinese models contain deliberate backdoors. But for governments, militaries, and critical infrastructure operators, the demonstrated possibility presents a serious procurement risk.
Beijing has also found an elegant inversion. As American policy appears unpredictable and access to U.S. models looks revocable, capitals from Europe to the Gulf are increasingly demanding “sovereign AI”: control over their data, systems, and digital destiny, free from foreign leverage. China presents itself as the answer: take our open weights, run them on your own servers, and owe nothing to Washington.
The concern is legitimate, but sovereignty is not achieved merely by downloading model weights. It requires control across the stack — data, compute, deployment, evaluation, updates, security, and legal jurisdiction. A Chinese model running locally may reduce one form of exposure while importing another: dependence on Beijing’s training choices, embedded assumptions, technical ecosystem, and unobservable vulnerabilities. China is selling autonomy at the interface while retaining influence at the foundation. That is ideological seduction working as designed: dependence presented as emancipation.
The American answer cannot be the indiscriminate release of every advanced capability. Nor can it be walling off U.S. technology so tightly that America appears an unreliable supplier, validating Beijing’s narrative and driving partners toward Chinese alternatives. Washington should establish a trusted-partner framework that offers allied governments and critical infrastructure operators durable access to frontier capabilities, rigorous security evaluations, auditability, transparent rules for restriction, and local deployment options where appropriate. That would offer genuine sovereign AI: the ability to build on American technology without surrendering control to either abrupt policy shifts in Washington or Beijing’s political system.
US LIFTS EXPORT CONTROLS ON ANTHROPIC’S ADVANCED MYTHOS AND FABLE MODELS
The stakes are immediate. The U.S. still fields some of the world’s strongest cyber-capable models, but Chinese systems are closing quickly. As advanced capabilities proliferate, keeping America’s best defensive tools from trusted governments and critical infrastructure operators gives adversaries more time to exploit vulnerable networks. In a race this close, delay favors the challenger.
China is not merely selling a cheaper model. It is selling a pathway to a world in which the digital future runs on its rails. Many American partners will choose technology they can trust, provided Washington gives them a viable and dependable choice. Trust grounded in transparency, the rule of law, and durable access is America’s hardest-to-copy advantage. Beijing is betting we squander it. The task is to prove that bet wrong before the global digital order is built on Beijing’s terms.
Dr. Jake Sotiriadis is executive director of Global Foresight and Strategy at Phaedrus Engineering. A geopolitical strategist and a retired U.S. Air Force intelligence officer, he is the author of The Revenge of Ideology: The Hidden Forces Reshaping Global Power and serves as a nonresident senior fellow at the Atlantic Council and a senior associate (nonresident) at the Center for Strategic and International Studies. The views expressed are his own.