WEF Work | Why China's AI Breakthroughs Should Come as No Surprise

For most of the last 25 years, I’ve worked as a writer for the World Economic Forum at various events. Recently, some of my work has been bylined, including this piece I wrote during the recent Annual Meeting of the New Champions in Tianjin. The link to the original piece is here.

---

The past year has seen a striking series of developments in China’s generative AI landscape.

Open-source models like DeepSeek’s R1 and V3 have posted performance numbers on a par with and in some cases surpassing their American counterparts - despite constraints imposed by US export controls. Others, including Alibaba’s Qwen3, which ranks near the top of global leaderboards in reasoning and language tasks, and the more recent MiniMax M1, trained on just 512 Nvidia H800 GPUs, have competed successfully against leading Western models.

These breakthroughs have prompted astonishment in US tech and policy communities, sparking something of a “Sputnik moment.”

That surprise, however, is in itself surprising. A more clear-eyed assessment of China’s structural position in the global AI ecosystem — the scale of its STEM education, the density of its computing infrastructure, the maturity of its applied research base and the ability of its governance system to align investment, policy, and talent — would have made it obvious that this level of performance was within reach.

Much of my own perspective on this derives from years spent inside the Chinese technology sector. This included a period at Baidu during the early surge of interest in AI; from about 2013 to 2016, I worked closely with machine-learning pioneer Andrew Ng, and observed Baidu co-founder and CEO Robin Li’s personal and organisational commitment to making AI central to the company’s future. In the years since, through conversations with researchers, entrepreneurs and policymakers which I share on the Sinica Podcast, I’ve revisited and deepened that understanding.

It’s clear to me that these developments did not emerge from a vacuum. Here are some of the factors that made such rapid progress all but inevitable.

How China quietly built a world-class AI ecosystem

By the time DeepSeek and Qwen startled Western observers in late 2024 and early 2025, the groundwork for China’s generative AI surge had already been laid over the better part of a decade. A 2017 State Council directive — the “Next Generation Artificial Intelligence Development Plan” — established AI as a national strategic priority. It was followed by cascading provincial and municipal implementation blueprints, generous funding from state-backed venture vehicles and regulatory sandboxing that gave AI startups considerable room to manoeuvre.

Chinese firms responded with vigor. By 2022, China was already filing four times as many AI-related patents as the U.S. and closing the gap in top-tier research output too. By 2024, models like DeepSeek-V3 were outperforming Meta’s Llama 3.1 and Anthropic’s Claude 3.5 Sonnet on common language and reasoning benchmarks, thanks to architectural innovations such as Mixture-of-Experts and Multi-Head Latent Attention.

The emergence of new players like MiniMax has underscored the rapid maturing of China’s domestic AI ecosystem. Meanwhile incumbents like Baidu, Tencent, and Zhipu AI have continued to iterate on their own large language models, expanding model capability, task generality and deployment at scale.

One prevailing narrative assumes that China’s AI sector would falter without access to cutting-edge chips, foreign capital or open research ecosystems. The announcement of strict export controls, including bans on the involvement of US persons in critical sectors of the Chinese semiconductor industry, prompted predictions of annihilation across China’s advanced tech sector.

But the steady accumulation of institutional and industrial capabilities tells a different story. China did not leap forward overnight, as many have assessed. It climbed steadily, through years of disciplined policy and wide-ranging public-private alignment. China has fostered a society primed to regard AI not with suspicion but enthusiasm.

Navigating export controls with innovation

For much of the last five years, Washington has sought to put in place export controls, particularly when it comes to exporting to China to protect AI advancement in the US. These measures began in earnest with the addition of Huawei and other firms to the Commerce Department’s “Entity List”. They accelerated with sweeping October 2022 restrictions on the export of advanced AI chips such as Nvidia’s A100 and H100, along with tools for semiconductor manufacturing. Even customised lower-performance variants like Nvidia’s H800 and A800 were eventually banned in October 2023.

While these controls significantly widened the compute gap between US and Chinese AI firms in the short term, they also prompted a burst of improvisation and efficiency innovation across China’s AI sector. DeepSeek, for instance, stunned global observers by releasing its R1 model in January 2025 — trained for what the company claimed was just $5.6 million on approximately 2,000 Nvidia H800 GPUs — at a time when comparable models in the West required significantly more compute and capital.

Several Chinese firms, including High-Flyer AI, which created DeepSeek, had preemptively stockpiled tens of thousands of A100 chips before trade restrictions came into effect. Others reportedly sourced GPUs via intermediaries in Southeast Asia. Despite these logistical workarounds, the longer-term response has been structural: growing investment in domestic chip design, led by Huawei’s Ascend series, and in homegrown semiconductor fabrication, with SMIC reportedly producing chips using deep ultraviolet (DUV) lithography instead of the more advanced extreme ultraviolet (EUV) process monopolized by Dutch company ASML — which fell under US export controls.

In parallel, architectural innovation took on new urgency. Facing hard constraints on compute, Chinese model developers began optimising aggressively for efficiency, developing techniques that minimised training costs without sacrificing performance. DeepSeek’s use of sparse Mixture-of-Experts models and memory-efficient inference pipelines is one result. The company’s public embrace of open-source release strategies, including models under permissive licenses, appears to have bolstered domestic and international collaboration. According to several Chinese industry observers I’ve spoken to, its success has also reinforced state enthusiasm for open-source development as a hedge against platform dependency and geopolitical risk.

What export controls did, in effect, was sharpen the incentives for AI engineers, founders, and funders in China to do more with less — to optimize performance per watt, per dollar, per GPU. Instead of derailing China’s AI momentum, these constraints made it more resilient and self-reliant.

China’s advantage in technology diffusion

A critical yet underappreciated dimension of China’s AI strength lies not in “frontier innovation” alone, but in its ability to diffuse general-purpose technologies such as large language models quickly and broadly across industrial, governmental, and consumer applications. China’s ability to scale has been explored in depth by scholars like MIT’s Yasheng Huang in his book The Rise and Fall of the EAST. But the distinction between frontier innovation and diffusion has been made persuasively by scholars like Jeffrey Ding, who argues that national power in transformative technologies stems less from their initial invention than from the institutional capacity to embed them at scale across the economy. Ding concludes that China is actually comparatively weak in this regard. But his measure is fairly narrow; he tallied up the institutions in a given country where at least one researcher has placed a paper at one of three leading AI conferences. I'd argue that while this is true, China nevertheless holds a decisive structural advantage when it comes to tech diffusion.

The country’s record of rapid deployment in other critical technologies — from mobile payments to high-speed rail to industrial robotics — has established a template for translating emergent technologies into widely adopted infrastructure. That same playbook is now being applied to generative AI. Whether in customer service bots, document translation engines, AI-powered medical triage, or educational tutoring apps, LLMs are already being deployed at scale within China’s domestic market. Hundreds of models have been submitted for regulatory clearance, a process now publicly trackable through China’s generative AI algorithm registration database.

In cities like Shenzhen, the fusion of software, hardware and supply chains into tightly integrated ecosystems enables rapid prototyping and real-time iteration that is difficult to replicate elsewhere. Model development is not separated from deployment; research and application converge spatially and institutionally. The proximity of model developers to domain-specific implementers — whether in industrial design, logistics, or consumer electronics — accelerates feedback loops and shortens the time between proof of concept and market impact.

This capacity for scale isn’t merely a function of demand. It reflects an organisational logic embedded across Chinese tech sectors. Companies like Alibaba, Tencent and ByteDance have matured in a competitive environment where successful tools are rolled out to hundreds of millions of users with minimal friction. Even less prominent LLM developers can test populations at scales that would be unthinkable elsewhere.

Moreover, China's vast language-specific dataset — drawn from social media, e-commerce platforms, state media and academic corpora — confers unique advantages in training models that perform exceptionally well on Chinese-language tasks and increasingly on bilingual or multilingual tasks. This linguistic scale, paired with institutional coordination and market reach, makes the diffusion of AI models not merely a technical exercise but a systemic national capability.

Star Trek vs. Black Mirror

If institutional alignment, infrastructure and talent form the skeleton of China’s AI capacity, its cultural orientation toward technology provides much of the muscle. A broadly held and socially reinforced belief in the power of technology to drive personal and national progress remains one of China’s most important — if often overlooked — advantages.

In contrast to the techno-anxiety that dominates much of the discourse in the US and Europe — where fears of surveillance capitalism, AI-driven disinformation, and existential risk shape regulatory debates — Chinese society has remained largely in what might be called its “Star Trek” phase: a techno-optimistic orientation where innovation is associated with improved material well-being, state capacity and national rejuvenation. As I’ve argued elsewhere, this stands in marked contrast to the “Black Mirror” mindset increasingly common in the West.

That contrast should not be misunderstood. To be sure, untempered techno-optimism can be reckless, especially when powerful models are rolled out at scale without adequate safeguards. But in China’s case, optimism does not mean the absence of regulation. As Kendra Schaefer pointed out in a recent interview, China’s AI regulatory infrastructure is, in many respects, ahead of the EU and the US. The generative AI algorithm registration database maintained by the Cyberspace Administration of China (CAC), which now includes hundreds of publicly listed LLMs, offers a level of transparency that no Western regulatory regime currently matches. The process mandates disclosures about model architecture, training data sources and safety mechanisms and is complemented by technical guidance on alignment and content controls.

This embrace of regulatory oversight has not, however, diluted enthusiasm for technological development. The public stature of engineers, computer scientists and AI researchers remains unusually high in China. The success of technologists is not viewed with suspicion or unease, but rather as evidence of national vitality. Startups and state-owned enterprises alike promote AI achievements as integral to China’s long-term modernization. Schoolchildren are taught coding at earlier ages; local governments run AI literacy campaigns; LLM-powered applications proliferate across e-commerce, health care and logistics.

In short, China’s techno-optimism is not naïve. It coexists with a growing state capacity to shape and steer AI development. But that foundational-cultural orientation — a belief that smarter machines will produce better lives — continues to shape how talent is recruited, how models are used and how innovation is received by the public.

Strategic alignment: When State, market, and academia row in the same direction

One of the less visible but profoundly consequential enablers of China’s rapid advance in generative AI is the unusually tight coordination among its public sector institutions, academic research bodies and private firms. While this kind of alignment is sometimes viewed with suspicion outside of China, particularly when it is framed in terms of state-led industrial policy or state-backed enterprise, the practical effect has been to lower barriers between research and application, to accelerate funding decisions and to unify long-term technological goals across domains.

Consider how China’s most capable research universities — Tsinghua, Peking University, Shanghai Jiaotong, Zhejiang University — serve not only as training grounds for AI talent, but as intellectual incubators for commercial ventures. Many of the leading generative AI firms in China, including Zhipu AI and Baichuan, emerged directly from university research labs, often with seed funding from state-affiliated venture arms and built-in partnerships with municipal development zones or digital economy clusters.

State guidance funds, particularly those aligned with the “New Infrastructure” initiatives launched in the late 2010s, have prioritised compute infrastructure, AI chips and cloud services. These funds offer long-horizon capital to projects that would likely struggle to gain equivalent traction in private markets, particularly during periods of economic tightening or when returns on investment are uncertain. Yet at the same time, the market incentive remains intact. Leading Chinese AI startups face intense domestic competition from rivals like DeepSeek, MiniMax, Moonshot AI and iFlyTek, all of which operate in a fast-moving environment that rewards iterative gains and rapid deployment.

The effects of this alignment are perhaps most visible in the open-source strategy adopted by firms like DeepSeek and Alibaba. DeepSeek’s decision to release models such as R1 under the permissive MIT license not only widened its international visibility but also helped catalyse a wave of downstream innovation and adaptation across China’s AI developer community. According to several observers, this approach has drawn favorable attention from Chinese policymakers, who increasingly view open-source models as a strategic counterbalance to Western platform dominance — and a way to internationalize Chinese-developed frameworks without becoming entangled in dependency relationships with U.S. cloud providers or model APIs.

Academic-industry linkages further reinforce this alignment. Research published in top conferences such as NeurIPS or ACL frequently involves co-authorship by scholars affiliated with industry labs. And China’s model regulatory regime, though top-down in structure, allows for extensive private-sector participation in refining content guidelines, interface design, and usage monitoring. The net effect is a pipeline from basic research to consumer product that is faster, more vertically integrated and often better capitalised than in more fragmented innovation ecosystems.

This alignment is not without tradeoffs. It can encourage groupthink, and in some domains — particularly those involving politically sensitive content — places firm limits on expression and experimentation. But when it comes to generative AI as a technical and economic domain, the coherence of China’s approach allows it to translate breakthroughs into societal-scale tools with exceptional speed.

Recognizing the inevitable

None of this should have come as a surprise.

That China would rise quickly in generative AI was not a geopolitical plot twist but the logical outcome of decisions made years earlier: to prioritise education in technical fields, to direct capital toward high-compute infrastructure, to mobilise talent at scale, to align state strategy with academic research and private-sector innovation and to promote a regulatory climate that, while distinct from Western norms, has proven remarkably proactive and adaptive. The headlines may have focused on chip bans and export controls but the enduring foundations put in place - scale, coordination, optimism and commitment - have been far more important.

Those who reduced China’s prospects to its access — or lack thereof — to Nvidia GPUs missed the bigger picture. They failed to see how China’s whole-of-society approach to emerging technologies — its capacity not just to invent but to diffuse and scale — would allow it to close the gap with remarkable speed. They underestimated a country whose regulatory sophistication in AI, as seen in algorithm registration regimes and content oversight, now surpasses that of the EU. They misjudged a culture where technologists are still viewed with largely unalloyed admiration, and not with suspicion, dismissed as “tech bros” with sinister motives. China is still a place where the idea of a machine improving the world has not yet curdled into cynicism.

As generative AI continues to evolve, the global conversation would benefit from fewer narratives of shock and more careful attention to the systems — educational, industrial, ideological, even cultural — that make progress possible. China’s model is not frictionless, and its political constraints remain real. But its ascent in AI reflects not only its resilience but perhaps even more importantly, strategic design.