Intelligence on the cheap: China's open-source AI strategy against the capex bubble

On 17 July 2026, China’s Moonshot released Kimi K3, a 2.8-trillion-parameter model billed as the world’s largest open-source model, with weights set to open by month’s end. In the same run, Alibaba’s Qwen family passed a billion cumulative downloads on Hugging Face, faster than any other model lineage, and DeepSeek keeps shipping its models under a permissive licence. Seen from Silicon Valley, this is a technology rivalry. Seen from l0g, it is something else: an economic weapon. By making frontier-class models free, China is collapsing the price of intelligence at the precise moment US giants commit hundreds of billions of dollars whose repayment assumes that intelligence stays expensive. The analysis that follows does not judge the models’ quality, it reads the price shock they cause.

The strategy: commoditise the model layer

China’s dominance in open-source models is no accident of timing, it is a choice. Alibaba has released more than a hundred models under the Apache 2.0 licence, DeepSeek published its own under an MIT licence together with a paper describing its training method, and Moonshot is now opening its largest model. Giving away the model layer means destroying its market value for everyone, oneself included, for the sole purpose of denying competitors their pricing power.

This mechanism is familiar to us. It is the exported deflation we described in the world’s factory selling off, transposed from the solar panel to the language model: flood the world with free capacity to suffocate the other side’s margin. Where Chinese industrial overcapacity drives down the price of goods, open-source generosity drives down the price of inference. In both cases, China exports a price drop its rivals absorb.

The target: pricing power

The model layer was supposed to be the moat, the scarce asset that justified the margins. It is becoming a commodity. A Chinese open-source model such as DeepSeek charges for its output inference around $0.28 per million tokens, against some thirty dollars for a leading US frontier model, a ratio near a hundred to one. On reasoning models, the gap stays an order of magnitude. The market rule is merciless: when someone offers comparable quality at one-hundredth of the price, the margin evaporates.

The point is not that closed US models are technically outdated, they are not necessarily. The point is that their edge no longer monetises at the token level. If raw intelligence trends toward free, the rent must lodge elsewhere, in the application, distribution, proprietary data. Yet it is precisely the token level that was meant to repay the infrastructure.

The point of impact: $725 billion of capex

Here it is, where the Chinese strategy meets the American balance sheet. The four largest hyperscalers, Amazon, Alphabet, Meta and Microsoft, plan to commit about $725 billion of capital expenditure in 2026, up 77% year on year; CreditSights puts the top five, Oracle included, between $700 and $900 billion. Nearly all of it goes into AI infrastructure: processor clusters, proprietary silicon, data centres. And a growing share of that bill is debt-financed, a mechanism we dissected in the debt behind AI, circular financing and the residual-value guarantee of infrastructure credit.

The return on these $725 billion assumes one thing: that the model layer keeps enough pricing power to yield the margin that will repay the debt. China’s open-source strategy attacks that assumption directly. If the price of inference keeps trending toward zero, the gap between committed capex and the revenue it generates, already scrutinised by markets that are repricing, becomes a solvency problem, not just a profitability one. It is the fragility we flagged in the AI boom and financial fragility and in the bubble within the bubble.

The geopolitical lever

To the financial dimension is added a power dimension, just as concrete. You cannot sanction a weights file once it has been downloaded. Where US export controls on chips aim to slow China, the diffusion of open-source models bypasses the symmetric lever: it installs the Chinese stack as the default among developers in the Global South, in universities and administrations that have neither the budget nor the access to closed US APIs. Setting the free standard means capturing the ecosystem and future dependence, a long-game logic that revenue tables alone do not capture.

The opposite reading

Fairness demands setting out the counter-argument, because it is serious and could overturn the conclusion. The first point is the Jevons paradox: cheaper intelligence widens the market, and a fall in the token price can blow up usage volumes, hence demand for compute, hence justify the capex rather than doom it. On this reading, value does not disappear, it migrates toward compute, that is toward chipmakers and inference operators, and toward the application layer that US labs monetise with enterprises. The second point is that open source carries hidden costs, compliance, security, no support, that keep many large accounts on closed providers. The third, finally, is that Chinese generosity is not pure strategic altruism: constrained by chip restrictions, unable to easily monetise closed models globally, China makes open source a rational second-best as much as a weapon.

These objections shift the question without cancelling it. Even if the capex is justified by volumes, it changes beneficiary: the silicon maker and the inference operator collect, while the closed lab that financed its moat on credit sees its margin thesis crumble. The deflation of intelligence is real; it does not destroy value, it redistributes it, and that redistribution does not follow the map of the debt.

Where the value goes

The moat shifts from the model to the layer above. That observation, banal on the surface, carries a heavy financial consequence: whoever financed the model layer’s margin with debt has tied its repayment to an asset whose price trends toward zero. The likely winners are compute and the application layer; the relative losers, the closed labs whose token rent evaporates and the creditors who backed them. China’s open-source strategy is therefore not an episode of the model wars, it is a risk factor for the heaviest capex cycle in the sector’s history.

The signals to watch fit in a few lines. The gap between the hyperscalers’ capex and revenue, which markets are starting to punish at every earnings call. The pace of Chinese releases, of which Kimi K3 is only the latest. The adoption of open-source models by large enterprises, the only arbiter of the migration. And the token price, the most direct thermometer of the margin left to defend. Reading AI through the lens of debt means seeing that the real question is not which model wins, but who repays when intelligence no longer sells.

---

Data and sources: MIT Technology Review, the future of Chinese open-source AI (Kimi K3, Qwen, DeepSeek); large-model API price comparison, 2026 (token price collapse, DeepSeek vs frontier models); 2026 hyperscaler capital expenditure, about $725 billion and CreditSights estimates. Kimi K3’s claimed performance figures are vendor announcements, without independent validation at this stage; inference prices and capex guidance move, the levels cited are those of mid-2026.

To go further: our analysis of the world’s factory and Chinese deflation; our AI-and-debt cluster, with the debt behind AI, circular financing, the residual-value guarantee, the AI boom and financial fragility and the bubble within the bubble; and our critical look at AI productivity gains.