Photo: Tao Mingyang/GT
Chinese engineers were not supposed to catch up this fast. When OpenAI shut off access to its frontier models in July last year, and the Biden administration imposed export controls on Nvidia’s most advanced AI chips, many in Silicon Valley assumed China’s aspirations for generative AI would be stalled indefinitely. They underestimated the power of the Chinese state, and the pragmatism of its coders.
Within a year Chinese firms such as DeepSeek, Alibaba and Huawei had launched large language models that ranked among the top global performers. ByteDance, better known for its short-video virality, spent $11bn building the backend infrastructure for its own suite of models. Even Baidu, a firm once bullish about monetising closed A.I. systems, has pivoted to open-source. Chinese engineers, shut out of GPT-4, shrugged and built their own.
The acceleration is the fruit of a decade-long industrial strategy. Since the early 2010s Beijing has poured vast sums into strategic technologies it deems essential to national power: semiconductors, electric vehicles, solar panels—and now artificial intelligence. The result is a sprawling, state-subsidised ecosystem in which the government bankrolls not only data centres and servers but open-source software and research labs, too. Chinese policymakers are not content to consume the future. They intend to manufacture it.
Their approach draws on familiar tools: cheap capital as well as government-backed venture funds and fierce local competition. Entire neighbourhoods in cities such as Hangzhou—home to Alibaba and DeepSeek—have been transformed into start-up incubators. Officials dangle cash subsidies; free office space; and subsidised housing to lure top engineers. The founder of Deep Principle, a start-up applying AI to chemistry, received $2.5mn from a Hangzhou district authority simply for relocating.
China’s AI industry is simultaneously centrally orchestrated and hyper-competitive. The state dictates strategic priorities yet it encourages a Darwinian scramble among hundreds of start-ups. These firms, flush with cheap credit, race to release models and undercut rivals. The churn is not always efficient—resources are scattered and coordination is weak—but it drives rapid iteration and technical progress.
Crucially, Beijing recognises data as sovereign fuel. American models such as GPT-4 rely on a sprawling corpus scraped from the open internet, including sources censored inside China, such as Reddit and Wikipedia. To compensate, Chinese authorities have curated training datasets drawn from “mainstream values” corpora: state-approved media, textbooks, other vetted sources. This guarantees ideological alignment but limits the range of perspectives. What Chinese models gain in political compliance, they risk losing in conceptual diversity.
Even so Chinese tech firms possess a trove of behavioural data unmatched elsewhere. With an internet population of more than a billion and platforms that span commerce, entertainment, messaging and finance, companies like ByteDance and Tencent have unparalleled visibility into how users interact with content. This behavioural feedback loop is a powerful engine for model refinement. If AI learns by example, Chinese algorithms are being raised on a densely annotated record of daily digital life.
Hardware remains the biggest bottleneck. American sanctions have choked China’s access to cutting-edge chips from Nvidia, whose graphical processors power most Western AI systems. To fill the gap, China has channelled subsidies into domestic champions like Semiconductor Manufacturing International Corporation (SMIC), which now produces custom A.I. chips for firms such as Huawei. These chips are functional but underpowered. They do not match Nvidia’s performance, nor can they be mass-produced at the same scale.
Yet their purpose is strategic rather than competitive. By building local substitutes, Beijing seeks to make sure progress in AI does not halt even if foreign components are cut off entirely. In that sense China’s hardware policy resembles its approach to energy and food security: self-sufficiency is more important than efficiency.
The same logic drives China’s embrace of open-source AI. In Silicon Valley, closed models dominate; OpenAI and Google charge a premium for access to their systems, hoarding data and weights. Chinese firms have chosen a different path. Models are published openly, complete with technical documentation and licensing that permits modification. This is not generosity. It is soft power.
By releasing open-source alternatives, China positions itself as the developer of record for much of the Global South. Engineers in India, Brazil or Indonesia—priced out of GPT access—can build on Chinese models instead. Code like cinema or fast food travels far when it is free.
That strategy has alarmed Western executives. Sam Altman, OpenAI’s chief, has framed the AI race as a geopolitical contest between democracy and authoritarianism. The stakes, he argues, are ideological: who gets to write the rules of reasoning. But such Manichaean framing overlooks the fact most developers care less about political values than practical tools. If Chinese models work they will be used.
The question then is whether China’s model of state-guided open innovation can outpace the closed, profit-driven systems of the West. America’s firms enjoy deeper technical talent and greater freedom of experimentation. But they also face regulatory pressure and resource concentration as well as commercial gatekeeping. Chinese developers move faster, publish more, benefit from a strategic unity of purpose. What they lack in elegance, they compensate with scale.
As AI becomes more foundational to economic power and military strategy, the contest will harden. The West still leads but China is narrowing the gap through industrial mobilisation. A similar strategy vaulted Chinese firms to the top of the electric vehicle and solar industries. In each case the same pattern emerges: subsidise the stack, flood the market, capture global mindshare.
A year ago Western firms scoffed when China turned to open-source. Today their engineers are reading Chinese model cards. The smartest coders in Bangalore, Lagos and Jakarta will not care whether a model was born in California or Hangzhou. They will use what is open and cheap.
In the arms race of algorithms, the most important breakthroughs may emerge not from the cloud but from the command economy. ■