Xi Jinping endorsed open-source AI Friday and called it a symphony of global collaboration. The Beijing company that released its most capable model hours before the speech called it a good business move. Both things are true, and the tension between them is the entire story.
The scene in Shanghai was choreographed with precision. Nine Nobel and Turing laureates in the audience. António Guterres present. Dignitaries from twenty-nine countries signed up the day before to create a new China-led body for global AI cooperation. Moonshot AI released Kimi K3 — an open-weight model that, per reporting, surprised people in Silicon Valley with its capability — hours before Xi took the stage. Xi warned against “overstretching the concept of national security in the field of AI,” did not name the United States, and said the new body would “answer the call of the Global South.” The choreography tells you what the words are for: this is a pitch for global AI standards written by the Chinese Communist Party, dressed in the language of openness.
And the language is doing real work, because it is aimed at a genuine contradiction in the American position. The Trump administration has blocked access to Anthropic’s Mythos model on national-security grounds. It has spent years restricting exports of advanced Nvidia semiconductors to China. OpenAI and Anthropic build largely proprietary models. The message Xi is delivering to twenty-nine countries — and to every country that wasn’t invited — is that the United States wants to set the rules of AI while keeping the best AI behind a gate, and China will give you the keys. That pitch is working because the contradiction is real, not because the pitch is honest.
Here is what the choreography obscures. MSI reported earlier this month that Beijing had held discussions with domestic AI labs about restricting overseas access to the country’s most advanced models — a potential reversal from China’s own previous strategy of promoting open-source globally. Read the two facts together: China champions open-source AI in a hall full of Nobel laureates on Friday, and on a different day explores export controls on its own frontier models. Publish the weights, control the pipeline. This is not hypocrisy unique to China. OpenAI was founded as a nonprofit to ensure AGI benefits all of humanity, then restructured into a capped-profit entity, then deleted “safely” from its mission statement. Google open-sourced TensorFlow while running a search monopoly that regulators on four continents have spent a decade trying to crack. The pattern is structural: open-source as strategy, not as principle. Cory Doctorow has a term for this — chokepoint capitalism — what happens when powerful intermediaries adopt the language and aesthetic of openness while preserving a closed apparatus, the same apparatus that makes the openness legible in the first place. Every time a model’s weights are published, the question is not whether the download works. The question is who owns the training data, who controls the compute required to run it, and who decides what the next version looks like.
Kai-Fu Lee, the Taiwan-born venture capitalist who formerly ran Microsoft’s and Google’s operations in China, told the conference he was skeptical of Silicon Valley’s “winner-take-all” approach to AGI. The framing is revealing. The problem with Silicon Valley’s approach to AI is not that it is winner-take-all — the problem is that the entire AGI framing is itself a fundraising document dressed as a technical specification. You cannot verify a system whose specification you cannot write down, and nobody has written down what AGI means in terms that would survive a protocol review. What Lee is describing is not a contest between openness and closure. It is a contest between two sets of chokepoint operators, each offering the world a different lock and calling it a key. Xi’s “symphony of global collaboration” is the pitch for lock number two.
Graham Webster, the Stanford research scholar who focuses on Chinese tech policy, told the Journal the new governance body was “more rhetorical than substantive so far” and that governing AI risks globally “is by definition going to need the U.S. and China both involved.” This is the single honest sentence in the coverage, and it is doing more analytical work than the rest of the conference combined. A global AI governance body that excludes the United States is not a governance body. It is a trade bloc with a UN veneer. A U.S. AI policy that excludes China is not a security policy. It is a market-access program with a defense veneer. Neither arrangement produces what Xi’s speech promised and what the American restrictions also promise: AI that is “secure and controllable” and “always under human control.” Those are the right goals. Neither country is pursuing them. Both are pursuing market dominance and calling it safety.
The competing standards play is not new. It is the same playbook that produced VHS versus Betamax, GSM versus CDMA, and the EU’s GDPR as a vehicle for European regulatory influence over a data economy dominated by American firms. In each case, the standard that won was not the technically superior one. It was the one backed by the largest captive market willing to deploy it. China’s twenty-nine-country coalition and the Brics AI bodies Xi announced are infrastructure plays — not the software itself, but the agreement to adopt it, the training seminars, the multilateral frameworks that make Chinese AI models the default in a bloc covering half the world’s population. The U.S. version is the same thing in reverse: Nvidia’s hardware stack as the substrate, American proprietary models as the software layer, export controls as the moat.
There is an open-source tradition that means something. The people who built arXiv, who run Linux kernel development, who fought the DMCA in court — they were not building chokepoints. The question for anyone thinking about AI policy seriously is whether “open-source AI” can mean what open-source software meant, or whether the compute and data requirements of frontier models make the term inapplicable at the scale that matters. A model whose weights you can download but cannot train, fine-tune, or run without access to compute controlled by the same four hyperscalers is open in the way a PDF of a library book is open. You can read it. You cannot build on it. You cannot leave.
The open-weight release of Kimi K3 on the morning of Xi’s speech is the tell. The release was not a contribution to a commons. It was a product launch timed to a political event — the same move Meta makes every time it releases LLaMA weights under a license that gives Meta control over who competes and how, harvesting goodwill from the open-source community while preserving monopoly on the compute and data that matter. Xi’s “symphony” is a business strategy with a policy veneer. So is Washington’s. The countries sitting in the audience, being asked to choose a side, are being asked to choose a vendor.
There is a public consultation period in neither Beijing nor Washington. The question of whether AI governance belongs in the UN framework or in a China-led alternative or in a U.S.-led export-control regime is being answered right now, in choreographed speeches and model releases timed to news cycles, and the answer is being written by the two countries that benefit most from the world treating their chokepoints as its options. The work is to name what is happening. The deadlines are the ones already past.