A week ago in Shanghai, Xi Jinping told the world that AI should not be walled off. "We should seize this rare, historic opportunity to encourage open source, openness, collaboration, and sharing," he said, warning against countries "overstretching the national security concept." In Washington, the official line rhymes: the Trump administration, its Treasury secretary said this week, "supports open source models." Listen to the speeches and you would think both superpowers had just discovered the virtues of giving AI away. Watch what they are actually doing, and you see two governments reaching for the same set of locks.
The American move came Tuesday. Treasury Secretary Scott Bessent said the United States would examine popular Chinese open-weight models for signs of intellectual property theft and could sanction the companies behind them. The technical term is distillation, using one model's outputs to train another, but Bessent preferred a blunter word: "you and I would call it theft." He claimed investigators are "finding watermarks of our US large language models on many of the Chinese models." The trigger is no mystery. Moonshot's Kimi K3, which releases its full weights on 27 July, has been matching or beating OpenAI's GPT-5.6 Sol and Anthropic's Fable 5 on some coding and agent benchmarks, at a fraction of the price. Anthropic has accused Alibaba, Moonshot and DeepSeek of siphoning its models; the administration now signals it may turn those accusations into policy.
The charge is not as clean as it sounds. Distillation is a routine industry technique, and even the American labs do it. Microsoft's Satya Nadella recently called it "ironic" that companies claim fair use to train on the public internet, then treat learning from their outputs as a crime. Hugging Face's Clement Delangue argues distillation is a minor factor, and that China's labs are simply doing very good, very open research. There is a glass-house quality to the whole dispute: Anthropic is paying $1.5 billion to settle claims that it trained on pirated books, and OpenAI is still fighting The New York Times over the same kind of borrowing.
Here is the twist that makes this more than a familiar Washington complaint. At the very moment the United States is thinking about how to keep Chinese models out, China is reportedly thinking about how to keep them in. According to the Financial Times, Beijing's Ministry of Commerce has been consulting Alibaba, ByteDance and Zhipu about restricting foreign downloads of their model weights, limiting the transfer of training data abroad, and even barring firms like TSMC from fabricating chips based on Chinese designs. Regulators are also weighing tighter rules on foreign takeovers of strategic AI startups, in part to prevent another deal like Meta's $2 billion purchase of Manus, which Beijing later ordered undone.
Put the two together and the rhetoric collapses. The feature that makes Chinese models so disruptive, that anyone can download, inspect and run them for almost nothing, is exactly what both capitals are now moving to fence off. America fears the open model because it threatens the lead of its closed champions. China fears giving away the one thing it has finally done better than anyone. The two sides will meet for their first formal AI talks in September, with Bessent leading the American delegation. They may find they have less to argue about than it seems. On the value of openness for everyone else, the world's rivals are quietly converging: it was a fine idea, right up until it started to work.