Xi Jinping had never spoken at the World Artificial Intelligence Conference before. WAIC has run in Shanghai since 2018, and in eight years the top of the Chinese state had left it to premiers, ministers, and the usual parade of lab founders. On Friday, July 17, that changed. Xi delivered the opening keynote himself, and the fact of his presence was the first half of the message. The second half was the content: China, he said, would treat artificial intelligence as a shared inheritance rather than a national weapon, and would build the institutions to prove it.
The line that traveled was the musical one. AI, in the official translation, “should not be a solo performance by a single country, but a symphony of international cooperation.” Around it Xi hung a specific program — open source as a “global public good,” a pledge to help developing nations build their own AI capacity, and the announcement of a new World AI Cooperation Organization, a governance body that Chinese outlets said would be headquartered in Shanghai. The New York Times read it as Beijing positioning itself as the AI leader for the parts of the world Washington was not courting. That reading is correct as far as it goes. It just undersells what happened.
Two weeks ago we argued that China’s open-weight strategy was a market move: while the United States built a gate around its frontier models, Chinese labs gave theirs away, because a model you can download beats a model you have to be cleared to use. That was a commercial logic. On July 17 it stopped being only commercial. Xi took the giveaway and made it doctrine — an official account of what AI is for and who China intends to be in it. The strategy did not change. Its author did.
The primary text, and its careful silences
The temptation with a speech like this is to summarize it into whatever confirms your priors. The discipline is to read the words. Both Xinhua and CGTN published full English texts the same day, and they are worth quoting rather than paraphrasing, because the gap between the phrasing and the Western gloss is where the real story lives.
Xi’s framing rested on a familiar phrase in Chinese AI policy: 开源开放, rendered as “open source and openness.” He called on countries to “uphold open source and openness,” to reject “technological monopoly” and “exclusive small circles,” and to make sure AI does not become “a game of the wealthy nations.” On the Global South specifically, the pitch was explicit — China would offer training, infrastructure, and shared resources so developing countries could “ride the express train” of AI rather than watch it leave the station. Channel News Asia captured the sharpest version of the frame: Xi warned that leaving poorer nations behind would create a “new historical injustice,” language that deliberately echoes older grievances about colonialism and the digital divide.
Now the caveat, and it is not a small one. 开源开放 is a governance posture, not a licensing commitment. “Open source and openness” is a slogan that predates this fortnight by years; it describes a general disposition toward sharing technology, cooperation, and non-monopoly. It does not, on its own text, promise that any particular model’s weights will stay downloadable, or under what license. Kyle Chan, who studies Chinese industrial policy, made this point directly in the days after the speech:
Yes, I think there’s been some confusion about China and open source AI. For China, open source and open development of AI are related but not the same thing. Xi’s speech at WAIC grouped a bunch of these terms together: open source 开源, open 开放, win-win 共赢, cooperation…
— @kyleichan July 21, 2026
This is the correct first-order skepticism, and the analyst discourse landed on it fast. A political slogan is not a software license. Beijing’s claim — AI as public good, freely shared — is not the same as a verifiable commitment, and it would be sloppy to treat the keynote as if Xi had personally licensed a model under Apache 2.0. The Jamestown Foundation’s analysis read the whole exercise as governance leadership as much as generosity: the World AI Cooperation Organization is, among other things, an instrument for setting the rules in rooms where the United States is absent, which is a bid for standing, not charity.
So the accurate summary is narrow: Xi elevated an existing strategy to the level of stated national policy and attached institutions to it. Whether the openness survives contact with China’s own security instincts is a separate question — one the same week promptly complicated.
The contradiction that showed up within days
Doctrine is easiest to puncture when the state contradicts itself, and this one did almost immediately. While Xi was in Shanghai calling for openness, Reuters and the Financial Times reported that China’s Ministry of Commerce — the country’s export regulator — had spent the previous month meeting with Alibaba, Moonshot, and other AI companies about restricting how much of their technology flows abroad. Mark Witzke summarized the collision on X:
FT confirms Reuters reporting: “MofCom talked to AI companies … on limiting the transfer of key data … overseas, as well as allowing their model weights to be downloaded by foreign users.” This conflicts with Xi recently affirming open source approach at WAIC.
— @mkwitzke July 21, 2026
Hold both facts at once, because that is the honest position. The head of state champions open weights as a global public good; the export ministry quietly studies limits on foreign weight downloads. This is not necessarily hypocrisy — it may be a state working out where openness serves it and where it does not, which is exactly what the United States has been doing from the opposite direction. It is a useful corrective to any triumphalist reading. The doctrine has an asterisk, and the asterisk is that Beijing reserves the right to close what it has opened. That is precisely the discretion we flagged inside Washington’s own voluntary gate: both capitals now treat model access as a lever of state, and both keep the lever’s tension deliberately vague.
The cleanest way to test a doctrine is to line each claim up against what has actually happened since. Read that way, the WAIC program is neither pure slogan nor delivered promise; it is a ledger with entries in both columns.
| Pledge or claim (WAIC / the speech) | Concrete action to date | Reads as |
|---|---|---|
| Open source as a “global public good” | Kimi K3 open weights shipped July 16, a day before the keynote; a new Qwen, LongCat, and a GLM update landed in the same fortnight | Doctrine with a downloadable referent, not only a phrase |
| Help the Global South “ride the express train” through training and shared infrastructure | The models are free to pull from Hugging Face; no funded delivery mechanism for the training and infrastructure pledge announced yet | A real artifact wrapped around an aspirational program |
| A new World AI Cooperation Organization, headquartered in Shanghai | Body announced at the ceremony; membership terms and rule-setting mandate not yet published | A standing and governance bid, named before it is built |
| ”Uphold open source and openness,” reject “technological monopoly” | The same week, Reuters and the FT report MofCom studying limits on foreign weight downloads | Openness held as a state lever, not a binding license |
The pattern is that every row where the action column is full points back to a shipped model, and the one row that reads as pure posture is the one MofCom is quietly hedging. That is the whole tension in a single frame: the doctrine is real exactly where the weights are.
Why the “it’s just a slogan” read is the comfortable mistake
Here is where the consensus skepticism, correct in its first move, walks into its second error. Yes, 开源开放 is a slogan. Yes, MofCom’s talks expose the limits. But the move from “it’s a slogan” to “therefore it’s cheap talk” skips the one variable that decides whether statecraft has teeth: is there anything behind the words?
There is. The keynote did not land in a vacuum. It landed in the middle of the densest fortnight of open-weight releases the industry has seen. In the days bracketing the speech, Chinese labs shipped or previewed a wall of near-frontier models with weights attached — Kimi K3 from Moonshot, a new Qwen from Alibaba, LongCat, and a GLM update. The specifics of that cluster deserve their own accounting, and we will map them properly. What matters here is the timing. Xi said “open source is a global public good” on July 17. Moonshot had released Kimi K3 on July 16. The doctrine and the artifact arrived within a day of each other, and that sequencing is not an accident — it is the entire argument.
Henry Gao, whose post was the most-engaged take in the WAIC discourse we tracked — 219 likes, 178 bookmarks, and north of 72,000 views — made the case that the speech rewards a closer read than the diplomatic register suggests:
The speech may sound like standard diplomatic boilerplate at first glance, but it is actually a revealing statement.
— @henrysgao July 18, 2026
He is right, and the reason he is right is downloadable. A pledge to help the Global South “ride the express train” is empty if the train does not exist. But if a finance ministry in Jakarta or a research group in Nairobi can pull a competitive open-weight model off Hugging Face for free, run it on modest hardware, and fine-tune it on local data without asking anyone’s permission, then the pledge has a referent. The government joining the model-release process is a phenomenon we’ve now watched in both hemispheres — but the two governments joined it holding opposite tools. Washington’s instrument is the clearance list; Beijing’s is the release itself.
That asymmetry is the whole game. A gated model is a favor the United States can grant or withhold, which makes every recipient a supplicant. A downloadable model is a fact the recipient already controls, which makes the giver look generous whether or not it intended to be. This is what we meant in July when we described open weights as physical infrastructure that ships without a permission step — an artifact, once released, cannot be un-released by any review. Openness-as-statecraft works not because the speech is sincere but because the models are real. Sincerity is unfalsifiable. A 2-trillion-parameter checkpoint under a permissive license is not.
What the room said
The discourse split the way you would expect, and the shape of the split is itself a signal. The China-watchers and policy analysts — Gao, Chan, Alex Colville — spent their energy on precision: what 开源开放 does and does not commit to, how open source differs from open development in Chinese usage, whether the World AI Cooperation Organization is a governance play or a generosity play. That’s the substantive lane, and it out-thought the cheerleaders even where it didn’t out-travel them.
The higher-reach posts were flatter. The single largest audience in the window belonged to Lord Bebo’s record-size-conference framing (over 800,000 followers, 42,000 views on the post), which is essentially scale theater — WAIC as the biggest AI conference in history, 1,100-plus companies, 70-plus countries. True, and beside the point. The most analytically useful engagement clustered on the smaller, sharper accounts. Gao’s “revealing statement, not boilerplate” was the most-bookmarked serious take we saw, and bookmarks are the tell that matters here: people were saving the argument to think about, not just liking a headline.
The skeptics landed one clean hit and one weak one. The clean hit was the MofCom contradiction — genuinely load-bearing, because it comes from the export regulator’s own reported conduct. The weak one was the reflexive “it’s propaganda,” which several accounts reached for and none managed to reconcile with the release calendar. Notably, the distillation-and-doubt crowd that dominated earlier Chinese-model discourse was quieter this round. When the models are shipping on schedule and topping arenas, “it’s fake” is a harder sell than “it’s a slogan,” and the discourse migrated accordingly.
The doctrine is only as good as the next checkpoint
The Deep Feed’s position: the “it’s just propaganda” read is the comfortable mistake, and the “it’s binding policy” read is the naive one. What Beijing actually did on July 17 is convert a release calendar into a foreign policy — one that works precisely because it is not a law. We would judge it on exactly one metric and ignore the rhetoric entirely: whether the flagship checkpoints keep shipping open past the point MofCom’s caution starts to bite. The slogan is unfalsifiable; the next download is not.
The right way to hold July 17 is as a hypothesis Beijing has now committed to in public. The hypothesis is that giving away frontier-adjacent models buys more global standing than hoarding them does — that in a world where the United States gates access and calls it safety, the country that opens access and calls it justice wins the developing world by default.
That hypothesis is not proven by a keynote. It is proven, or falsified, by whether the weights keep coming after the speech, at the licenses the slogan implies, past the point where MofCom’s caution starts to bite. If a year from now the flagship Chinese models are quietly API-only and the “open source” is confined to last year’s checkpoints, the doctrine will have been what its critics said — a slogan with a conference attached. If the releases keep landing the way they landed this fortnight, the July 4 market move will have matured into something Washington has no clean answer to: a foreign policy you cannot sanction, because the product is already on every laptop that wanted it. Xi did not ask the world to trust China with AI. He handed the world the models and let the download do the arguing.