China’s Accelerated Rise in AI Power and Global Competition

When the State Council issued its New Generation Artificial Intelligence Development Plan in the summer of 2017, it set a headline goal that read, at the time, like a distant ambition: a core AI industry worth one trillion yuan by 2030. That target was cleared five years early, with China's core AI industry surpassing 1.2 trillion yuan in 2025. It would be easy to read the achievement as a straightforward story of a race being won, but the more revealing story is about how it was won. China has not out-resourced its principal rival on the terms that rival prefers. It has instead pursued a deliberate substitution, trading state coordination, algorithmic efficiency, and sheer scale of deployment for the compute capacity and private capital it cannot yet match. The result is a campaign that has consistently outrun its own timelines while remaining hostage to a single, stubborn bottleneck.

The 2017 plan was candid about the country's starting weaknesses, acknowledging significant gaps in basic theory, core algorithms, and high-end chips. The years since have been an exercise in closing them through escalating state commitment. The Fourteenth Five-Year Plan of 2021 ranked new-generation AI first among strategic technologies; Premier Li Qiang's 2024 "AI Plus" initiative, modeled on the earlier Internet Plus campaign, pushed for deep integration of AI across the economy; and by August 2025 the State Council had translated that ambition into concrete adoption targets stretching to 2035. The Fifteenth Five-Year Plan, adopted in early 2026, invoked AI four times as often as its predecessor and called for a national integrated computing-power network. Money followed direction. Precise figures remain elusive given the opacity of Chinese budgeting, but the structural contrast with the United States is unmistakable: where American AI spending is dominated by private capital — some 109 billion dollars in 2024, roughly twelve times China's private funding — China mobilizes through the state, channeling hundreds of billions of yuan through government guidance funds, a National AI Industry Investment Fund, and local-government venture vehicles. The two countries are not financing AI in remotely the same way.

That coordinating logic extends to the companies. China's technology giants have become de facto instruments of national policy, and their most consequential move has been to embrace open weights. Alibaba's Qwen family became the world's most popular open-weight model series in 2025, with more than forty million downloads across over a hundred released models. Baidu's Ernie Bot passed 200 million users; ByteDance's Doubao became the country's most-used assistant while triggering an industry price war by cutting its model to a fraction of prevailing rates; Tencent folded its Hunyuan model into WeChat and more than two hundred other services. Huawei, though never formally part of the "national team," anchors the whole edifice through its Ascend chips and software stack. By making capable models freely available, Chinese firms have converted a relative weakness in frontier monetization into a lever of global influence, seeding the open ecosystem that developers elsewhere increasingly build upon.

No single event captured the strategy better than the emergence of DeepSeek. Founded in 2023 as a spinoff of a quantitative hedge fund and led by Liang Wenfeng, the company released its R1 reasoning model in January 2025, matching or beating OpenAI's o1 on key benchmarks while reportedly costing only 5.6 million dollars in its final training run — a fraction of the outlay assumed necessary for frontier performance. The market response was seismic. Nvidia shed 589 billion dollars in a single session, the largest one-day loss in US stock-market history, as investors reconsidered whether massive compute budgets were truly indispensable. Chinese analysts called it evidence that export restrictions had accidentally triggered reverse innovation; Western commentators reached for the phrase "AI Sputnik moment." By early 2026, nine of the world's top ten open-weight models originated with Chinese developers. DeepSeek was the clearest proof that cleverness in algorithms could partially stand in for hardware China could not buy.

Partially, but not wholly — and there lies the constraint the entire strategy is built to work around. The hardware gap remains severe. The RAND Corporation estimated in 2025 that the United States held a tenfold advantage in total AI compute, controlling roughly three-quarters of the world's capacity against China's fifteen percent. SMIC, the country's most advanced foundry, can produce seven-nanometer chips but cannot scale them without the extreme-ultraviolet lithography it is barred from importing, leaving yields far below commercial norms. Domestic accelerators still depend on foreign high-bandwidth memory, and Nvidia's CUDA software ecosystem remains a moat that homegrown alternatives have yet to cross. Huawei's most advanced chip reaches perhaps eighty percent of an H100's performance and compensates through clustering — its systems can match Nvidia's on some workloads only by using more than twice the chips and considerably more power. Even DeepSeek has said its single biggest limitation is access to compute. The bottleneck is real, and much of what surrounds it is an attempt to route around it.

That bottleneck is largely a policy artifact. Since October 2022, successive rounds of US export controls have banned advanced AI chips, closed the loopholes that let modified versions slip through, restricted the equipment needed to manufacture chips domestically, and by early 2025 amounted to a near-total ban on China's access to leading accelerators. The controls have imposed genuine costs, constraining domestic production and forcing reliance on architecturally inferior alternatives. Yet their record is double-edged. They have not meaningfully stopped China from training cutting-edge models, as DeepSeek demonstrated, and they have hardened the country's resolve toward self-sufficiency, spurring domestic chip design, retaliatory bans on critical minerals such as gallium and germanium, and enormous state investment across the semiconductor supply chain. Restriction and stimulus have proven difficult to separate.

The competitive picture that results defies the arms-race framing it usually receives. On the technical merits, the frontier has converged with remarkable speed: gaps between top US and Chinese models on major benchmarks that ran to double digits in 2023 had narrowed to low single digits, in some cases near parity, by the end of 2024. But convergence at the frontier masks a set of durable asymmetries. The United States retains decisive leads in compute, private investment, the most-cited research, and commercial monetization; China leads in AI patent filings — roughly seventy percent of the global total — in industrial deployment, and in the proliferation of open models. As several analysts have put it, the two are less racing on one track than running in different lanes, the United States pushing frontier capability while China prioritizes mass integration. A quieter shift compounds the point: a reverse brain drain has begun, with dozens of prominent Chinese scientists leaving US institutions amid tightening visa scrutiny and research cuts.

Around this technological contest China has built one of the world's most active regulatory regimes, and here too the approach is distinctive. Rather than a single comprehensive statute — a draft was floated and then quietly dropped from the 2025 legislative schedule — Beijing has issued a stream of modular rules governing recommendation algorithms, synthetic media, and generative services, backed by a public algorithm registry that by 2025 listed thousands of tools, the only registry of its kind anywhere. The rules bind private developers tightly, require generated content to uphold core socialist values, and coexist with a surveillance apparatus that operates under far looser constraints, producing a dual-track system that pairs stringent data protection for companies with expansive latitude for the state. Abroad, China has moved to shape the rules others follow, launching a Global AI Governance Initiative, sponsoring a widely co-signed United Nations resolution on AI capacity building, and exporting AI infrastructure across the developing world while positioning itself as a champion of Global South interests.

Taken together, these threads describe a country that has moved from aspirational planning to substantive execution faster than most observers expected, and that has done so by playing a different game than the one its rival is winning. China has substituted the coordinating hand of the state, efficiency in algorithms, and scale of deployment for the compute and capital it still lacks, beating its own targets while remaining tethered to a chip bottleneck that no degree of coordination has yet dissolved. The era in which AI leadership could be captured by a single metric or awarded to a single nation is over. What has replaced it is a slower, stranger, and more consequential competition, spread across technology, governance, economics, and influence, in which each side leads on the terms it has chosen and neither can claim the whole field.