Fig 1. Chinese National Day Parade during the Cultural Revolution of the late 1960s.

Introduction

In June 2023 at a conference in India, Sam Altman, CEO of OpenAI, remarked that any startup’s attempt to compete against his company with just a few million dollars would be “pretty hopeless.”¹ In the same month, Ilya Sutskever, a chief scientist behind ChatGPT, predicted an eternal performance gap between open-source models and private ones². This regard for compute resources as their “moat” for innovation, along with the glamorization of their proprietary architectures, had been among the biggest narratives in AI development commonly sold by renowned faces of AI in Silicon Valley. It wasn’t until January 20th, 2025, that these predictions were disproved by an innovation from a lesser-known research lab. This innovation disrupted conventional understanding of how intelligence can be achieved in foundation models, but what truly flustered Western tech communities was the fact that the breakthrough came from China, which was believed to be lagging behind by a few years in AI development.

How has China’s AI development evolved from being a distant follower to challenging US dominance? Indeed, the purpose of this essay is to provide nuanced insights on the state of AI in China through different periods of history so that we understand what has actually evolved in China’s AI and how. To provide these nuances effectively, I will focus only on three key ideas in the AI history of China and examine them in depth, while briefly covering other milestones in the entire historical timeline just on a high level.

Early AI History in China

The narrative of early AI history in China is often presented in parallel with the reception of cybernetics in the Soviet Union. It is in the form of contrasting opinions on cybernetics and AI between the PRC and the Soviet Union alongside their ideological divergence after the death of Stalin in 1953. While the Soviet Union made a U-turn into embracing cybernetics, Chinese scientists are portrayed to criticize this transformation by the Soviet as “revisionism” and cybernetics itself as “capitalist pseudo-sciences”, rejecting cybernetics in general altogether. It wasn’t until 1978 when the Eight-Year Plan was established at the National Science Conference in Beijing after the Cultural Revolution that the activities of science and technology, including cybernetics and AI, are regarded as starting to revive. This Dengist narrative as termed by Bo An³, resonates well with the conventional narratives from the West, which marks this opening-up period after the Cultural Revolution as the “proper” beginning of AI development in China with the adoption of “proper” science from the West. Although this framing is helpful in grasping the high-level idea of early AI development in China, by perceiving it as the truth, we can not only lose sight of potentially valuable alternative approaches to AI that existed in China before 1978 but also ignore the nuances in its historical reception of cybernetics. Since this piece is not intended as a comprehensive counter-argument of this flattened early history AI in China, I will attempt to cover the history on a high level, albeit indicating where over-simplication often occurs in most narratives.

The Sino-Soviet Treaty of Friendship and Alliance in February 1950 marked the beginning of what unfolded in science and technology in China in the 1950s. China, being a country recovering from massive damage caused by WWII, received a large amount of scientific and technological assistance from the Soviet Union. Thousands of Soviet scientists, experts, teachers, and workers were sent to China to help China with heavy industries, agriculture, education, and scientific research, which implies the significant influence of the Soviet Union on the science and technologies policies in China⁴. From the late 1940s and early 1950s under Sterlin, cybernetics was regarded as pseudoscience and the study thereof was deemed subservient to the influence of the West⁵. Therefore, the topic of machine intelligence was at least a common interest of Chinese scientists if not regarded with pessimistic views that were reflected in the Soviet Union.

With the de-Stalinization process after Stalin’s death in 1953, the Soviet’s attitude towards cybernetics was completely changed. However, de-Stalinization alongside some other proposals made by Khrushchev, Salin’s successor, such as “peaceful coexistence,” and “peaceful transition to socialism” were scorned by the CPC as revisionism⁶. This indicates one of the ideological divergences between the two nations, which is where the arguments for China’s rejection of cybernetics came from. However, looking at the Party newspaper and various public opinion pieces, one can see that the potential of the machine and cybernetics were relatively optimistic. For instance, in 1965, a piece titled “Cybernetics and Its Achievement” appreciated the subjects of contemporary AI research such as pattern recognition and self-learning machine (essentially machine learning) as “outstanding achievements” of cybernetics³. However, it did attack those who “make the mistake of applying cybernetics to human social issues, eradicating class struggle, and the essential distinction between man and machine”⁷. To see the context, while the Soviet Union was generally described to embrace cybernetics, what Soviet scholars focused more on was the compatibility of the philosophy of cybernetics and their political ideology since they intended to use it as their ideological tool more so than a practical tool for engineering applications, industries, manufacturing, and so on. This very attempt to apply cybernetics in the social realms is also the only view that Mao Zedong, the founding leader of the PRC, expressively rejected. Essentially, he belittled the idea of machines having agency and consciousness especially in political and social space, albeit appreciating machines existence as tools of productivity.

Mao’s pragmatic views on cybernetics steered machine intelligence research away from distractions of philosophical questions. These favorable views towards practical applications seemed healthy in retrospect and could have theoretically benefited the development of AI. However, two infamous turmoils in the history of China, the Great Leap Forward (1958-1962) and Cultural Revolution (1966-1976), devastated China’s scientific capacity through famine, economic collapse, and persecution of intellectuals, which crippled technological development for over a decade⁸⁹. Fortunately, after the Cultural Revolution, AI research in China started to take off with consequential conferences such as the National Science Conference in Beijing in 1978 and the establishment of Chinese Association for Artificial Intelligence (CAAI) in 1981. Thus, we can see the making of AI, albeit in the name of cybernetics and perhaps other domains, even before the opening-period in the 1980s, which is widely, yet misleadingly, remarked as the “foundational” memoment of AI in China.

China’s Advantages in the Age of Implementation

China has produced volumes of AI publications since 1987, the year of China’s first research publication on AI from Tsinghua University¹⁰. However, it has been undeniably the US, Canada, and the UK who took the lead in emerging topics in AI such as expert systems, symbolic AI, machine learning, and lastly deep learning¹¹. Meanwhile, it was not easy for China to close this technology gap until the mid-2000s when the internet became readily accessible. Back then, the study of AI in China was largely concentrated among a few elite scholars at research institutes. It means that it was not accessible to broader Chinese audiences and was not yet developed as an engineering discipline focused on practical applications. Concretely, the only window Chinese students had into the state of global AI research was outdated textbooks that were poorly translated and occasional lectures from a visiting scholar. Internet access in schools was a scarce commodity, and studying abroad was difficult without a full scholarship¹¹. This is not to disparage the Chinese government’s endeavors in AI nor several outstanding achievements in AI by Chinese scholars at the international level in the 2000s, but rather to highlight, on a high level, how China is lagging behind the US when it comes to innovation or pushing the field forward.

Looking at the global state of AI, what has pushed the field forward is arguably the rise of deep learning, a machine learning paradigm that enables machines to automatically learn complex patterns from vast amounts of data without requiring hand-crafted rules or features. Although the foundation for deep learning has been laid slowly and steadily for decades, it was not until the success of AlexNet in the ImageNet challenge in 2012 that deep learning was found to work. This marked the beginning of being able to translate previous academic achievements into real-world use-cases. However, as Kai Fu Lee put it, “The great majority of China’s technology community didn’t properly wake up to the deep-learning revolution until its Sputnik Moment in 2016”, referring to AlphaGo’s stunning victory over the world champion in the ancient game of Go¹². Unlike IBM’s Deep Blue that defeated the world chess champion in 1997 by evaluating multitudes of possible positions for each move with “brute force”, AlphaGo ran on deep learning which allowed it to develop intuitive pattern recognition, much like how human players learn to ‘feel’ the right moves instead of calculating every possibility. Less than two months after this miracle of deep learning, the CCP released its 13th five-year plan, which set clear benchmarks for progress by 2020 and 2025 were set and projected to become the center of global innovation in AI by 2030¹³.

It is important to note that the fact that deep learning has opened up a plethora of potential applications has tilted the AI playing field twoard China. It signifies the transition from “the age of discovery” to “the age of implementation”, which means while the progress in AI was previously driven by a handful of elite scholars mostly from the West who did much of the difficult and abstract work of AI research, the progress started to become driven by AI entrepreneurs and engineers who can translate the ideas of deep learning into real-world applications for business, healthcare, manufacturing and so on. Concretely, four private tech giants, Alibaba, Baidu, iFlytek, and Tenchent were selected to form the first AI “National Team”, and each was assigned to take the lead in a specific field of AI application: Alibaba for smart cities, Tencent for medical imaging, Baidu for autonomous vehicles, and iFlytek for voice recognition¹⁴. This is what is meant by the age of implementation. The progresses made in these fields do not equal pushing the boundary of science of AI itself or making the progress towards AGI (Artifical General Intellgience), and yet, they represent massive economic value creation and have profound implications for transforming industries and improving people’s daily lives. The reason why this transformation from discovery to implementation favors China is that it trivializes one of China’s greatest weak points, “outside-the-box approaches to research questions” and leverages the country’s most significant strength, “scrappy entrepreneurs with sharp instincts for building robust businesses.”¹¹

To understand entrepreneurship in China, it is helpful to have a look at its distinct cultures in the internet ecosystem. While copycating another company’s idea is stigmatized in Sillicon Valley, it is culturally acceptable in China. Most famously, Wang Xing, a Chinese entrepreneur even called “The Cloner”, copied the business model of America’s hottest startups as well as even the UI of their sites and served them to Chinese users. As a case in point, Meituan, the group-buying site he found, became even more successful than the American company he cloned from, Groupon, and once reached the fourth most valuable startup in the world. However, the success didn’t come from simply bringing the group buying business to China. Over five thousand companies did the exact same thing after all. What made the difference was how the company executed the business model among the fierce domestic competition. This widespread cloning created an internet battle ground where thousands of mimicking competitors forced each other to better iterate their products, control costs, and raise money at exaggerated valuations. It is this very competitive landscape with ruthless copycats that breeds a generation of the most determined and versatile entrepreneurs on earth. When the wave of business models surrounding AI arrived in about 2016, there were already these experienced entrepreneurs, who were survivors of fierce competitions since the inception of the Internet. They were willing to dive into this promising field of AI and capable of creating numerous AI applications in China. On top of that, Chinese government VC funds and private funds not only provide financial support to AI startups but also empower them in strategic planning, R&D, and market expansion¹⁵.

In the age of Implementation, data matters as much as expertise in the field of AI. Arguably, the biggest advantage China has is in the ability to provide its firms and researchers abudance of data. The Chinese government manages to collect data primarily through an extensive network of surveillance, security, and traffic cameras. Here, Chinese private companies benefit tremendously from partnership with the government, where they gain access to this valuable government data and yet provide their services to the state in return. For instance, a company like SenseTime provides facial recognition technology to help with public security¹⁶. In return, they get access to millions of facial images and behavioral data from the government’s camera network which is impossible to collect privately¹⁷. This creates a unique advantage for Chinese AI firms that simply isn’t available in most Western countries due to privacy laws and different government-private sector relationships. What’s unique about AI businesses is the virtuous circle of data where more high-quality data leads to better products, attracting more users, which in turn generate more data that further improve the product.

On balance, China may not have as strong AI research labs as the United Kingdom, Canada and the US. However, with its venture-capital ecosystem, large user base, and access to a wealth of data, China is gradually reshaping the world order of AI in the age of implementation.

Taking a Lead in Innovation

China has been by far the leader in number of AI patents over the last decade with three times as many granted patents as their U.S. counterparts (35,310 vs 12,080) in 2022 for example¹⁷. However, using the number of patents produced by a nation as a metric for innovation is misleading for they vary greatly in quality. Indeed, only 4% of AI patents first filed in China were also filed in another jurisdiction, compared with 32% of patents first filed in the U.S, indicating the superiority of U.S patents in quality¹⁸. Filing the same patent in another country’s patent office is typically time consuming and expensive, and therefore, it is carried out only when the inventor thinks the patent is valuable enough to protect internationally. So, although China has been indeed outstanding in creating top-notch AI applications such as hyper-accurate facial recognition technology, it wouldn’t be wrong to remark that Chinese companies didn’t participate in real AI innovation until recent few years. But, not all companies in China must be solely priortizing commercialization of AI products and profit maximization. There must also be Chinese entrepreneurs who are genuinely driven by curiosity and desire to create rather than freeriding the innovation brought from the West. Indeed, Liang Wenfen, CEO of DeepSeek, was such an entrepreneur.

Deepseek is fully funded by High-Flyer, a top quantitative hedge fund which was Liang’s previous main venture, and has no plans to fundraise¹⁹. It was officially founded in July 2023 even after the release of GPT4 from OpenAI. In an interview with Liang, Deepseek’s ambition for innovation is clear in his answer to the question of why DeepSeek has chosen to focus on only research and exploration unlike most Chinese companies who choose to have both models and applications. He answered, “Because we believe the most important thing now is to participate in the global innovation wave. For many years, Chinese companies are used to others doing technological innovation, while we (Chinese companies) focus on application monetization.”

To appreciate the miracle that DeepSeek pulled off, we need to have a look at the AI landscape in recent years. Since the inception of ChatGPT in late 2022, the center of attention in the global AI community has been large language models (LLM) as an architecture to achieve Artificial General Intelligence (AGI). At their core, LLMs rely on the Transformer architecture, which essentially enables the models to predict the most likely next token in a sequence based on all preceding context after being trained on sufficient amount of text corpus. According to the scaling law²⁰, the performance of a LLM revolves around three main variables, model size (number of parameters), dataset size (number of training tokens), and compute budget (total amount of FLOPS for training). As any one of these three is scaled up while keeping the others constant, performance tends to improve in a predictable, power-law way until it hits a bottleneck caused by one of the other two. Therefore, training a state-of-the-art LLM that can lead in the leaderboards is regarded only in the capacity of big tech giants for being extremely capital expensive.

In 2023 and 2024, even China’s best LLMs are estimated to have a twofold gap in model structure and another twofold in data efficiency compared to the best international levels¹⁹. It means they have to consume four times more computing power to achieve the same performance as the best LLMs from the U.S. This is a huge problem for Chinese innovators since computing power comes from hardware, specifically GPUs, and China does not have comparable capacity in producing GPUs as good as Nvidia’s best GPUs. Simply importing those Nvidia GPUs is difficult because the Biden administration imposed the ban on the export of their chips to China in 2022 once they have found that China had just tested a hypersonic missile system, technology that would allow it to fire a nuclear warhead faster than anyone else, which was built using American microchip technology²¹. However, this export control was, as Jensen Huang put it, a “failure” since it only “gave Chinese companies the spirit, the energy and the government support to accelerate their development.”²² This was exactly what happened to DeepSeek. It revolutionized how LLMs are trained with a series of innovations in model architectures in order to consume less compute, memory and cost without compromising the model’s performance²³. Their novel architectural ideas accumulated over a year culminated with DeepSeek R1 model in January, 2025, disrupting the long-held belief in Sillicon Valley that scaling either of the model size, dataset size, and compute budget is the only way to improve the performance.

Still, Liang acknowledged that bans on shipments of advanced chips remain a significant obstacle to their research. Of course, the more access to better AI chips, the more beneficial it is to research regardless of how much efficiency China can eke out of restricted hardware resources. This challenge has driven China’s strategic pivot toward developing self-sufficient AI capabilities without relying on US chip technologies or semiconductors from Taiwan. Currently, as a case in point, Huawei’s Ascend 910B AI chip is claimed to be capable of up to 80% of the performance of NVIDIA’s A100 GPU when training large language models (LLMs)²⁴. However, Reuters analysts suggest that while Huawei chips are comparable to NVIDIA in terms of raw computing power, they still lag in performance²⁵. Despite these current limitations, the trajectory suggests China is steadily narrowing the technological gap.

Closing Thoughts

Overall, China’s current position in the AI race represents the culmination of a compelling historical journey that began long before the commonly assumed 1978 “opening-up” period. It was not just in some recent year that China managed to catch up with the U.S in the AI race. Since 2016, China has quickly excelled at translating and adapting novel ideas in AI into applications that empower businesses, raise people’s standards of living, advance military capability, and so on due to its venture-capital ecosystem and ability to collect enormous amounts of data. The period of 2012-2022 is when the world is reaping the results of academic advances in deep learning that are made over past decades by building applications that change the fabric of our daily lives. In this age of implementation, China implemented very well although new breakthrough inventions in AI were mostly still coming from the U.S. Indeed, technologies that drastically changed the world’s perception on AI, such as Transformer architectures, came from an American research lab, but Chinese companies, most notably DeepSeek, are now at the forefront of pushing the field forward instead of enjoying being just an excellent implementer of AI innovation like in a few years ago. AI experts are increasingly believing that AGI may not come from language models but rather some architecture that is fundamentally different from Transformers. This shift in perspective aligns with Liang’s philosophy that true competitive advantage in AI lies not in proprietary technologies or computational resources, but in cultivating “an organization and culture capable of innovation”, and China is now observed to have both hard (capital and computes) and soft resources (talents and culture for innovation).

References