China Power: State-Spun, Not State-Run
China’s new white paper sets out its bid to lead the world. But its lead in open-weight AI was built by firms that took their orders from the market.
Grüezi!
In January 2025 investors wiped $589bn off Nvidia in a single session – the largest one-day loss in US market history, on Bloomberg’s reckoning – all because of a free Chinese model.
Airbnb now uses a model from Alibaba to answer its customers; the maker of Cursor built its coding model on one from a Beijing lab.
And last week Apple asked Washington for clearance to buy memory from a Chinese firm the Pentagon listed.
China has published one of those earnest papers on global governance that set the China hawks worrying that the world’s future is being mapped out in a meeting room in Zhongnanhai.
Yu Jie, at Chatham House, made the case for reading it more carefully. The paper, she argues, bundles Beijing’s long-standing themes – a multipolar order, the UN at its centre, a louder voice for the global South – into a single bid to reshape how the world is governed, whilst quietly exposing the gap between that ambition and the money China has to spend on it. She is right on both – it’s not a Marshall Plan, and it is a serious document.
But like almost everyone, she is also reading China’s strategy off its paperwork, and that isn’t where the strategy lives. It lives in the choices engineers and firms make — which models to build on, which suppliers to buy from – not in government policy documents.
And so Beijing is benefitting from its firms’ success through mechanisms it never engineered, and is credited with a foresight its planners never had.
1 The wrong units
China is rising into the space the United States is leaving, and the white paper is yet another sign of it.
The people paid to track that rise watch China’s announcements – the chips it buys, the budgets it sets, and – of course – the declarations it issues.
But the real rise is in the adoptions: in defaults, toolchains and bills of materials, one download at a time. Repositories, not Confucius Institutes, are what’s growing China’s reach.
2 The default
A year ago DeepSeek, a Hangzhou firm, released a model close to the frontier which, on its telling, had been trained for a few million dollars. It turned out that figure covered only the final run, but that wasn’t really the point.
The point was the licence. The weights were free to download, run and alter. Nothing was asked in return, and so the model spread like wildfire through a cost-conscious community of students, developers, and ultimately big enterprises.
The results show up in two numbers from Hugging Face, the field’s main repository. Over the past year, Chinese models passed American ones in downloads, and Alibaba’s Qwen is the starting point for roughly two in every five new models built on top of an existing one.
Qwen, DeepSeek, Zhipu’s GLM and Moonshot’s Kimi are the open families developers reach for first. America’s closed labs still hold the frontier, but in AI everything open is increasingly Chinese.
It has reached real products. Airbnb leans heavily on Qwen to answer its customers – its chief executive, who had found OpenAI’s tools not yet up to the job, called it fast and cheap.
Cursor’s own coding model, used by a great many developers, was built by training further on Moonshot’s Kimi.
These are Chinese models doing daily work inside Western firms. The question is – how tightly does that bind them to China?
3 State-spun, not state-run
None of these were state labs. None were quite private, either.
DeepSeek began as a side hustle inside a quant hedge fund, High-Flyer, run by its founder, Liang Wenfeng.
Qwen came out of China’s answer to Amazon, Alibaba.
Zhipu spun out of Tsinghua University, took money from Tencent, and sat close enough to the state that the US Commerce Department put it on the Entity List – its trade blacklist – in January 2025 for, it claimed, helping China’s military.
For China hawks, somewhere behind all of them stood the government, offering cheap capital, subsidised computing power, and naming open models a national priority.
People who have no faith in government to achieve anything in the West, are quick to attribute progress to state intervention in China.
But none of that tipped the balance in making models spun out of a hedge fund’s experiment and a university lab the base that the rest of the world fine-tunes on.
People downloading weights decided that, and they didn’t ask Beijing for permission.
The Party didn’t start it.
4 What walking away costs
China dependencies run deep in a global economy.
China makes almost all the world’s polysilicon, ingots and wafers – near 95 per cent, by the IEA’s reckoning – so a plant in Texas assembling its own panels still buys its wafers from China, because at that scale and price there is nowhere else to go. The only real alternative to Chinese solar is not building solar.
Huawei is also hard to dislodge. It is still the world’s biggest maker of telecoms equipment – 31 per cent of the market in the first half of 2025. Ericsson and Nokia could supply the alternative, so there is a possible Western option. It just costs a fortune to take.
When Washington made small rural phone companies tear out the Huawei kit they had installed, the bill went from an estimate of $700m to almost $5bn, and seven years on, the FCC reported back in June, only 53 of 126 projects were complete.
Apple’s chip problem is easier but also illustrative. For a month Apple has been pressing Washington to let it buy memory from ChangXin, a firm on the Pentagon’s list of Chinese military companies flagged for a possible Commerce blacklist.
But Samsung and SK Hynix make the same memory, and Apple has so far bought none of ChangXin’s. This is price pressure, not dependence: a glimpse of where a Chinese supplier might one day be hard to refuse. And also where the costs of shutting down globalisation are exactly that – costs.
The AI cases aren’t yet as embedded, so leaving is easier. Cursor’s coding model was trained on top of Kimi, so to change bases it would have to retrain its own product – costly, but possible, and a Western open model might do as the replacement.
Qwen answers Airbnb’s customers today, but it is one hosted model among several the firm could use.
But blocking open source will raise costs for US firms forced to shop at home.
5 What it buys, and for whom
So what does China gain from this shallow AI adoption?
Alibaba gets reach, and no fee – the weights are free. Broader Chinese industry gets something more durable: as more software is built on Qwen and Kimi, they become the defaults everyone else works to. That’s the power of standards-setting, and it accrues with a kind of invisible inevitability.
A firm that has built deeply on a Chinese input acquires a stake in the status quo, and turns, without meaning to, into another voice against rules that would make it change. Apple, lobbying for its ChangXin licence, is one minor instance – although it’s asking to be “let in” to dependency.
The more the world runs on what China controls, the less America’s export controls will bite. That provides a shield rather than a sword for Beijing. It makes China harder to sanction without – yet – giving it the kind of control it can use to throw its weight around.
With free weights taking the price of a model to zero, all that remains is running it. That means chips and electricity — and China has cheap answers for both. Its chips still lag, but if they catch up, near-zero power costs would let frontier-grade intelligence run at a fraction of what American suppliers could charge.
Lower cost, not low cunning, is changing the global AI arena, one price decision at a time.
6 The limits of the off-switch
American controls work where the engineering is hardest – the fast memory that feeds AI chips, the latest production nodes, the machines that print the finest detail – which is why the memory Apple wants from China is the commoditised kind and not cutting edge.
If the frontier is where AGI will be discovered and frontier capability decides who gets there first, then the open models everyone else is building on are a sideshow.
But what if it isn’t?
And also that openness cuts both ways. What’s out there now can’t be called back: Washington tried to keep it out and couldn’t, Beijing built it and can’t reclaim it.
Beijing’s trap is the sharper one – models are taken up because they’re cheap and come without political baggage. The moment they are instrumentalised by the Chinese state they cease to be either. Beijing would risk doing what the old Chinese parable warns against: pulling at the seedling to make it grow faster, and tearing it dead from the soil.
Washington has the opposite trouble. This month it pulled up the drawbridge on the newest US frontier models – shutting out friends and allies. A reminder to the rest of the West that there will be no place in the citadel when – or if – the AI storm comes.
And elsewhere, each restriction on a closed American model nudges the next builder towards an open Chinese one that so far no government can recall.
7 What follows
China’s global strategy may be in its white paper, but its global power is building in the dependencies it is creating.
The questions are plain: for each Chinese model a firm runs, where does it run, who hosts it, who sees the prompts, who ships the updates, what would moving cost, and what is the substitute if the answer turns bad.
Few firms can answer those questions today. Congress has started asking by subpoena – in April two House committees demanded that Airbnb and Cursor account for their use of Chinese models – but a subpoena is a blunt tool for a question that belongs in every company’s risk report.
Businesses have already run the audit in reverse. Coinbase swapped Claude and GPT for open models from Zhipu and DeepSeek, and cut its AI bill by roughly half. It’s not a Chinese lock-in — the models are swappable, which is the whole point — but proof that the price floor has dropped out from under the American labs. Anthropic could find itself the Audi of AI: an up-market name undercut by cheaper Chinese models that get the job done.
Running the same enterprise workload through Anthropic’s Claude was nine times more expensive than through Zhipu’s GLM 5.2, for the same results.
Every company doing a year-end budget review is looking at the same numbers. Weighed against the security risk, the saving will win nearly every time.
China is inside the world’s systems at a depth that changes case by case: entrenched in the ones with no substitute, barely dug in to the ones with many. And the white paper is a poor guide by which to judge where that power is headed.
The more honest guide to China’s global reach is the more boring one – what runs on what, what would leaving cost – and in AI that is America’s to draw, company by company.
Beyond America it is also a simpler question for the so-called middle powers. In the AI future, would you rather be eaten by Godzilla or ripped apart by King Kong?
So far no one wants to come up with an answer to that.
Thanks for reading!
Bis bald,
Adrian