Shanghai, 17 July. Xi Jinping takes the stage at the World AI Conference — his first time there in person since the conference began in 2018. He says AI development should not be a solo performance by any single country, but a symphony of international cooperation. He warns of new historical injustices when access to the technology is unevenly distributed. He promises to help the Global South build its own capacity. The same week, three Chinese labs each release a world-class model, and two of them pledge to publish the weights freely.
It is tempting to dismiss it as theatre. That would be a mistake. The question is not whether it is a performance — it is one. The question is what is being staged, and for whom.
The ladder, again
In an earlier essay I described the ladder: the seventy-year path along which poor countries climbed from agriculture to the factory to the office and grew rich — and how AI is pulling that ladder up, on both rungs at once. I also described who owns the machine: that access to compute is the new border, and that the border is being drawn with export rules that divide the world into tiers — allies without a ceiling, most others with one, a few shut out entirely.
What I did not write, because it had not yet happened, was what happens when one of those excluded countries is the world's second-largest economy.
Export controls pulled up the hardware ladder for China. China's answer, delivered from a stage in Shanghai, is to lower a different ladder: the software one. Frontier-class models, freely downloadable, at a fraction of American prices. The West controls the rungs made of silicon; China is making the rungs of code free.
And notice the language. “New historical injustices” — that is the ladder's story, word for word. The thesis that the development path is being pulled away from those who have not yet had time to climb is no longer an analysis in the margin. It is Chinese foreign policy. Beijing is selling the story of the ladder the West pulled up — and offering its own.
The week when everything dropped
Facts first, because they deserve it. On 16 July, Moonshot AI released Kimi K3: 2.8 trillion parameters, the largest open model ever announced, with a context window of one million tokens. The weights are promised for 27 July under an open license — and honesty requires the caveat that as of writing they are not yet online. “Open” is a promise until the files are actually there.
Two days later, at the same conference, Alibaba previewed its Qwen3.8-Max: 2.4 trillion parameters, multimodal, beaten by only one model in the world according to the company — and with a promise of open weights that breaks Alibaba's own pattern, since its top tier has until now always been kept closed. DeepSeek is reported to have updated the same week. Three releases, one stage, one week. This is not a coincidence. It is a demonstration.
Almost as good
How close is “almost as good as the best”? Honest answer: closer than ever, but not all the way — and exactly where the line falls matters less than one might think.
Vendor numbers should be read for what they are. Alibaba's claim about Qwen 3.8 still lacks a model card, benchmark tables and independent tests; what exists is a preview. Moonshot is more becomingly honest — its own technical blog concedes that K3 sits behind the two strongest closed systems overall. But K3 now also has independent support: an established third-party evaluator ranks the model fourth in the world. Fourth. A freely downloadable model, if the promise holds, among the four best that exist.
And then there is a stronger proof than any benchmark: behaviour. On the largest neutral routing platform for AI traffic, Chinese models now account for 46 percent of all tokens, versus 36 percent for American ones. A year ago, the average was eleven percent; in the first half of 2025 it was under five. The crossover itself happened in February. The market votes with its feet, and it does not wait for the benchmarks to be verified. The driver is not ideology but price: the open Chinese models are 60 to 90 percent cheaper to run.
“Almost as good” does not need to win the leaderboard to win the world. For most tasks, almost is enough — and almost has become a commodity.
The price of intelligence
Now to the numbers that shook the market. K3 costs three dollars per million input tokens and fifteen per million output tokens, thirty cents on the dollar on cache hit, the same rate across the entire context window. And here is the point many miss: this is not the bargain story. It is more than three times more expensive than Moonshot's own predecessor. K3 is priced as a flagship — in line with the American mid-tier, below the American top tier, roughly a third of the most expensive rate on the market. Honesty also requires the caveat here: the model always reasons at full stretch and therefore burns more tokens per task, so measured per completed task it lands, according to independent estimates, at around half the price of the closed top models. Not a tenth.
But under the flagship lies the real pressure: the commodity layer. Moonshot's previous generation costs under a dollar per million tokens. DeepSeek's cheapest variant costs pennies. That is two layers of pressure at the same time — near-frontier at mid-tier prices, on top of a floor that is sinking toward zero.
The market responded on release day: the Nasdaq fell one and a half percent, and commentators reached for the same words as in January 2025 — a new DeepSeek shock. But the most telling signal came from inside the American stack itself. GitHub added Moonshot's code model to Copilot on 1 July. And Microsoft, the American frontier labs' largest customer, is reported to be evaluating whether K3 could take over parts of Copilot, with savings of up to six hundred million dollars. The report is unconfirmed. The direction is not: pricing pressure is no longer coming only from the outside. It is coming from inside the Western supply chain.
And yet — a counterweight, because it belongs to honesty. Volume is not revenue. On the same platform where Chinese models carry nearly half of the traffic, the largest American vendor takes twelve percent of the traffic but almost half of the revenue. Two markets are forming: intelligence as commodity and intelligence as premium. The question is not whether the premium market disappears. The question is how broad the commodity market becomes — and who can only afford to live in it.
Walls or price war
The West has two moves, and they pull in different directions. One is to compete: lower prices, release its own open models, defend the premium tier with quality. The other is to wall off: tighten export controls, restrict the use of Chinese weights, pull up more ladders.
The paradox is that walling has already been tried, and that is why we are here. The chip controls were meant to slow China's path to the frontier. They have made the path more expensive — China's compute shortage is real, and Moonshot's own infrastructure buckled under demand in the days after release — but they have not closed it. Instead, China moved the fight to a layer the controls cannot reach: software, which copies freely and travels in a download.
Now the American labs are lobbying Washington against open weights. The dry observation of Mozilla's CTO deserves to be quoted: they would not care if the threat were not real. And the strongest argument on the wall-building side also deserves to be taken seriously, because it is true: open weights cannot be recalled. An API can be shut down, patched, fitted with guardrails. A model that has been downloaded by a thousand organisations is beyond any vendor's reach. That is a different risk profile, genuinely.
And Europe? The EU AI Act was written with exemptions for open source — but the exemptions do not apply to models above the systemic-risk threshold, and trillion-parameter models are above it by a wide margin. On paper, therefore, the obligations apply regardless of license. In practice: what is an obligation worth when the provider sits in Beijing and the weights already sit on a thousand European servers? Brussels wrote the rules for a world in which open source meant small European models. Now openness is changing sender, and the rules have not caught up.
The ladder of silicon
And how did the wall actually work out? The scorecard after four years deserves its own chapter, because it is the textbook case of second-order effects.
The controls were supposed to starve Chinese AI of chips. The first-order effect arrived as planned: Nvidia's most powerful processors disappeared from the Chinese market, and the company's China revenue fell toward zero. The second-order effect was the one that counted. China responded by building the dependence away. Huawei is aiming this year for roughly six hundred thousand units of its top chip, manufactured at SMIC, and has published a road map three generations ahead. The latest chip is designed to do without the West at every link in the chain: its own high-bandwidth memory instead of the Korean HBM the controls can reach, a design that avoids Taiwan's packaging technology. A Chinese model with seven hundred billion parameters has been trained on more than six thousand domestic processors. When DeepSeek released its latest model in the spring, it ran from day one on four Chinese chip brands. And the most telling turn of all: when Washington finally reversed course and again allowed export of advanced Nvidia chips, in exchange for a fee, it was Beijing that said no. China's state procurement list of approved AI hardware names Huawei and Cambricon. Nvidia is not on it. The wall that the United States built is now maintained from the inside.
Honesty requires the counter-image. Per chip, the gap is real — Huawei's best delivers roughly a third of Nvidia's flagship, and advocates of continued controls argue that the distance is widening rather than narrowing, up to seventeen times within a year. The bottleneck is in memory, where three foreign firms still dominate, and in lithography, where the most advanced machines cannot be sold. China compensates horizontally: more chips in larger clusters instead of better chips. That costs power — and power is the one input China is building faster than anyone else.
But there is a third floor in this building, and it is often forgotten. The United States controls the top of the silicon stack: the design tools, the lithography, the most advanced fabs. China controls the bottom: the polysilicon everything is made from, the gallium, the germanium, the rare earths, the specialty chemicals without which the fabs do not work. And China has started to pull on its end. Controls on rare earths were tightened in two rounds during 2025; by the spring, Japan was in effect cut off from the heavy metals and gallium; in January of this year, export controls were placed on semiconductor-grade silicon and the technology to produce it. Two hands are choking the same ladder from opposite ends — one at the top, one at the foot.
And there, in the middle, sits the asymmetry that explains the whole game. China has the world's largest chip market and dominates the cheap layers of the chain — but the profit sits at the top: the combined annual profit of Chinese semiconductor firms is less than a twentieth of Nvidia's alone. That is the order Beijing is trying to rewrite, with open models in one hand and raw-material controls in the other. The ladder is not being lowered out of generosity. It is being lowered because whoever owns the bottom of a ladder wants to own more of it.
The Silica Peace
Washington's answer to the symphony is already here, and it carries a name worth translating. Pax Silica — the silica peace. The U.S. State Department launched the initiative in December 2025: a coalition around exactly the stack the previous chapter mapped, from rare earths and energy to fabs, chips and AI infrastructure. Its architect, Under Secretary of State for Economic Growth Jacob Helberg, calls it “an economic security coalition built on the reality that our security is inseparable from our technological edge.”
Read that sentence next to Xi's speech and the mirror is exact. Xi calls on the world to jointly resist stretching the concept of national security to cover AI. Helberg makes the stretching the founding principle. Each side names the other's main instrument as the main threat — Beijing sees the walls, Washington sees the dependence — and both are right, about the other.
Nor is this open versus closed, however much the story wants to be. Trump's AI Action Plan explicitly urges American labs to release their own open-weight models. The difference sits in the distribution logic. China puts the ladder in the town square: free to download. The United States lowers it from the balcony — “the full stack,” from chip to model, as a package deal for trusted partners. Diffusion as commons, versus diffusion as licence.
Hence two different audiences as well. WAICO recruits the excluded; among its signatories stand Russia, Pakistan, Indonesia and Laos. Pax Silica recruits the well-endowed — the coalition addresses itself, in the State Department's own words, to countries that host the world's most advanced technology companies, plus the raw-material states that will supply them. In the Philippines, which has joined, the question is already being asked whether the role in practice becomes that of the subcontractor. One ladder is lowered toward those who never got to climb. The other, toward those already standing at the top.
For Nordic readers there is one more line, and it is not a small one: we have already signed. According to the State Department's own account, Sweden was among the first eleven signatories of the Pax Silica Declaration, and at the June 2026 summit Sweden, Norway, Denmark and Finland signed the bloc's joint AI statement on a growth-friendly regulatory line. When the blocs are counted in the next chapter, the Nordics are not on the sidelines. We are already in one of the teams.
Honesty demands the ironies, in both directions. At the June summit, Helberg told governments they should not approach AI “primarily through the lens of restriction” — said on behalf of the country that invented the chip controls. And China preaches openness through the world's most controlled internet. But behind the ironies lies an admission that weighs more: Pax Silica is the proof that the wall alone did not hold. A superpower confident in its lead would not have needed to build a coalition around polysilicon and gallium. The silica peace is, like every pax that history has named after a power, not the absence of conflict but a way of organising it.
Whose ladder?
So: is the ladder real?
Halfway. Open weights are not open access. A model with 2.8 trillion parameters requires compute, power and engineers that most of the world does not have; running it yourself is not a real cost lever for most. The ladder has been lowered, but the rungs still cost — and the power, the cooling and the competence are sold by the same handful of actors as before.
And the dependence does not disappear. It changes owners. Licenses are conditional and can be rewritten. The tempo is unrelenting: six weeks after K3, Moonshot retires its previous flagship, and whoever built on it has to move. Whoever owns the ladder decides where it leans.
Then the geometry. The day before Xi's speech, 29 countries signed the founding of a world organisation for AI cooperation, headquartered in Shanghai. Pax Silica gathers 35. A single country appears on both lists: Kazakhstan. Read that again — of over sixty countries in the two blocs, one, exactly one, has chosen both.
In the vocabulary of the earlier essay, this is not Sharing. It is Friction: a world splintering into tiers where access decides everything. The United States divides the world into compute tiers. China divides it into governance blocs. Two ladders, two hands holding them, and a majority world that has to choose which hand to trust.
But — and this is the essay's real question — Xi's offer sounds like Sharing. Gains and technology made broadly available. The ladder put back. Capacity building, cooperation centres, open weights as global commons. My earlier essay on convergence closed with a question: whether the gains would be shared, quickly enough, before the fractures cascaded. Xi now stands on a stage and says: we share.
What decides the coming decade is whether it is true. Whether it is Sharing for real — or Friction dressed in the language of Sharing. A ladder put back, or a ladder someone else is holding. The circle broken, or just another turn.
Last time I wrote that the outcome is not a forecast but a choice. That still stands. What is new after the week in Shanghai is that the choice is no longer the West's alone.
Reference to the earlier essay: When the Ladder Was Pulled Up (22 June 2026). Empirical claims about K3 weights, independent rankings, Qwen 3.8 weights, price data, traffic and revenue figures, the EU AI Act, Huawei production numbers and China's raw-material controls draw on open sources as of the publication date. The Microsoft/K3 evaluation report is unconfirmed by Microsoft. The seventeen-times figure is an estimate from the pro-control camp (CFR) and is attributed as such. Information on Pax Silica (launch, membership lists, Helberg's statements, the Nordic signatures) draws on published U.S. State Department material and independent reporting as of 22 July 2026. That Kazakhstan is the only country on both lists should be read with the caveat that the blocs expanded in June 2026.