China-US Large Model Competition: Widening Gap or Neck and Neck
-
In 2023, the semiconductor industry faced multiple challenges including macroeconomic and geopolitical factors. Welcoming 2024, Jiwei Network launched a retrospective and outlook series, inviting representative companies in the industry to summarize the past year's industrial chain developments and hot topics, while also looking ahead to the future. Through this series, it aims to provide in-depth references for the semiconductor industry, helping companies better adapt to new development trends.
Jiwei Network reported (by Chen Xinghua) With the accelerated arrival of a new global wave of industrial and technological transformation, artificial intelligence, as the core driving force in this wave, has increasingly garnered high attention from countries worldwide. Against the backdrop of evolving international geopolitical environments and rising trends of industrial chain deglobalization, China and the United States are striving to become the main leaders in the global AI large model field, attracting worldwide attention. This is not only about exploring and leading a new high ground in technological and industrial transformation but also concerns the future of human socio-economic development.
Currently, government agencies, industry think tanks, media scholars, and others have conducted research on the artificial intelligence strategies of major technological powers. However, there is relatively little comparative analysis of the industrial policies, technological innovations, academic research, industry applications, and market financing of AI large models between China and the United States. In light of this, this article compiles and extracts key points from reports by relevant authoritative institutions and think tanks, as well as column articles and rankings, with the aim of providing references and promoting the development of China's AI large model industry.
Fortune's 50 AI Innovators: US Companies Dominate, Only Baidu from China
On November 21, 2023, Fortune magazine unveiled its inaugural "50 AI Innovators" list. The vast majority of these 50 companies are American, including prominent AI and tech firms like OpenAI, Anthropic, NVIDIA, Google, Microsoft, Meta, IBM, and Adobe. Emerging players such as Midjourney, Hugging Face, Adept, Runway, Pinecone, Exscientia, and ARC also made the cut.
In contrast, only one Chinese company - Baidu - was included in the list. In its commentary on Baidu's selection, Fortune noted that in October 2023, Baidu launched ERNIE Bot 4.0, claiming this model outperforms OpenAI's chatbot in handling many Chinese-specific tasks and rivals ChatGPT in complexity and functionality. Additionally, beyond using machine learning in its search engine, cloud computing division, and other products, Baidu is developing autonomous driving algorithms and operates a fleet of driverless "robot taxis."
Fortune states that the AI 50 list is an important guide designed to help the industry understand companies that are creating technologies defining the future. The list was curated through in-depth discussions and screening by Fortune in collaboration with venture capitalists, industry analysts, and Fortune's AI expert team to identify companies at the forefront of the industry. While the AI field is rapidly evolving, one thing is certain: the work of these companies will not only shape the future of AI but also influence the world in which people live.
Industry analysis points out that, based on the number of companies listed, there remains a significant gap in comprehensive AI innovation between China and the U.S. This disparity may stem from listed companies focusing more on research direction and practicality, whereas most domestic AI companies in China merely follow market trends, lacking core advantages. They require more investment and R&D to further enhance innovation capabilities and competitiveness.
"Comparative Study of AI Large Model Applications in China and the U.S.": Mixed Results in Competition, Domestic Investment Remains Cautious
On September 9, 2023, the Titanium Media International Think Tank released the influential report 2023 Comparative Study on AI Large Model Applications Between China and the US. The report evaluates four dimensions—industry penetration, acceptance, application maturity, and financing landscape—focusing on mainstream AI application sectors: AI + Office, AI + Finance, AI + Healthcare, AI + Entertainment, AI + Education, AI + Transportation, and AI + Energy. It provides a multidimensional comparison of the current state, characteristics, strengths, and weaknesses of AI large model applications in China and the US.
The report highlights that large models currently have the highest penetration rate in the financial sector (78%) globally, with office applications also showing significant adoption. However, penetration remains low in energy and construction industries. In the AI + Office sector, US giants lead the trend, while Chinese companies are catching up. In AI + Finance, the US has a more mature ecosystem, though China has also entered the application phase. For AI + Healthcare, China faces slow adoption due to data constraints, whereas the US benefits from data advantages and focuses on R&D.
Additionally, in AI + Entertainment, the US faces challenges, while China has potential to overtake. In AI + Education, the US emphasizes teacher assistance, while China prioritizes exam-oriented applications, each with distinct strengths. In AI + Transportation, China holds an advantage in traffic applications, while both countries compete in autonomous driving. For AI + Energy, both nations exhibit low penetration and remain in the exploratory phase.
The report concludes that China and the US lead other countries in the direction of AI large models, yet their competition has become overt, with each side demonstrating varying strengths. Tech giants like Google, representing the US, have long been dedicated to foundational theoretical research, enabling the US to lead the trend in artificial intelligence. Leveraging a robust engineering culture, the US maintains a leading edge in foundational large models.
In the primary market, the US is also proactive in investing in AI large models. Tech giants such as NVIDIA and Microsoft have emerged as the most significant "unicorn hunters," fueling the development of US artificial intelligence through substantial investments. However, due to the constraints of large models on talent, capital, and technology, domestic primary market investments in related projects are not as fervent as in the US. Investment institutions act more cautiously and tend to favor startups that leverage open-source models from leading companies for practical applications.
According to incomplete statistics, in the first half of 2023, the US AIGC primary market saw Silicon Valley complete 42 financing rounds in the AI sector, totaling approximately $14 billion, accounting for 55% of global financing. The average financing per round was $330 million, nearly 13 times the average level. During the same period, China's AI sector recorded 161 investment and financing events, a decrease of 153 compared to the same period in 2022, marking a 49% decline. Additionally, the total investment and financing in China's AI sector amounted to 6.174 billion yuan, a reduction of 9.9 billion yuan year-on-year, down by 62%.
On the other hand, the overall informatization level across various industries in the United States is relatively high, having accumulated abundant structured data, which lays a solid foundation for the implementation of large models in the US. This is particularly evident in areas like medical research and development, where the US continues to maintain a first-mover advantage. In contrast, China's large models have largely played the role of a follower. However, the gap between China and the US in foundational models is not significant. The research philosophy of 'emphasizing application over foundation' and the vast downstream demand in China have led domestic institutions to focus more on applied research for implementation.
Therefore, in terms of the application of AI large models, China can be said to have an edge. On one hand, the primary market in China sees more entrepreneurial projects centered around large model applications. On the other hand, mature enterprises across various industries leverage their deep expertise in their respective fields, either by integrating foundational models externally or by developing their own models using open-source frameworks, to incorporate AI large models into their industries. Nevertheless, many industries in China have yet to complete their informatization process, and the lack of foundational data makes it challenging for AI large models to gain traction in certain sectors. Additionally, computational constraints have become a sharp blade in the AI competition between China and the US, making the effort to overcome these constraints a top priority in the development of AI.
Jiwei Consulting: China Leads in Natural Language Processing Academic Achievements, While the US Excels in Research Capabilities
The AIGC frenzy sparked by ChatGPT worldwide is essentially an application of large language models based on Natural Language Processing (NLP) technology. With the deepening development of artificial intelligence, the demand for natural language processing continues to rise, and both the market size and application fields are expanding, while also presenting more challenges. In this regard, academic research achievements in natural language processing can provide direction for industrial technology choices and guide enterprises in technological innovation.
In August 2023, JW Insights proudly launched the report "Behind ChatGPT's New Wave of AIGC: Analyzing Natural Language Processing Technology from an Academic Perspective." This report conducts a comprehensive analysis of global top-tier journal academic papers in the NLP field over the past two decades, examining publication trends, source countries/regions, publishing institutions, key scholars, funding agencies, publications, and hot topics. It provides the industry with corresponding interpretations of academic research achievements in this field.
The report shows that as of the end of April 2023, there were 74,992 NLP papers published globally. In terms of publication trends, the number of global NLP papers grew relatively steadily between 2002 and 2012 but began to increase more rapidly from 2013 onward. In 2017, the publication of the generative pre-training model Transformer sparked a new wave of academic research in the NLP industry, causing the number of papers to rise sharply.
From the overall regional distribution of papers, academic research achievements in the global natural language processing field are mainly concentrated in mainland China and the United States, with both countries publishing over 18,000 papers, occupying the high ground of global natural language processing academic research. Although countries such as India, Germany, the UK, and Japan also have considerable academic research outputs, the gap compared to China and the US is significant.
In terms of institutions publishing papers, among the top ten research institutions globally by paper output, those from China, the US, and France constitute the majority. The Chinese Academy of Sciences ranks first globally with 1,971 papers, far exceeding the second-place University of California by nearly 50%. Tsinghua University and Peking University also appear on the list. In contrast, apart from traditional universities, high-tech companies in the US place great emphasis on fundamental academic research in natural language processing, with cross-industry giants such as Microsoft, IBM, and Google publishing a considerable number of papers. This characteristic of US academic research in this field is unique globally.
Regarding funding for papers, the primary sources are government-level entities in various countries, such as the National Natural Science Foundation of China, the US National Science Foundation, and the European Commission. Among these, the National Natural Science Foundation of China has funded the highest number of papers, at 8,335. In the US, institutions like the National Science Foundation, the Department of Health and Human Services, and the National Institutes of Health have collectively funded the publication of over 6,000 papers. This demonstrates that governments worldwide are strongly supporting fundamental research in natural language processing technology.
Additionally, JW Insights calculated comprehensive scores for academic papers based on multiple indicators including writing quality, journal level, influence, and advancement. From natural language processing academic papers, they selected the 50 most outstanding papers, clearly presenting their titles, countries/regions, affiliated institutions, research fields, and key points to serve as references for corporate research and innovation.
Breaking it down by country and region, the United States leads in natural language processing research strength, with 28 papers in the TOP50, far surpassing other countries and regions—8 of which came from Google. China ranks second, but with only 6 selected papers. The UK ranks third with just 3 papers on the list, while other countries have only sporadic high-value academic achievements. Compared to the US, China's natural language processing academic papers still need further improvement in writing quality, influence, and advancement.
"China's AI Large Model Map Research Report": Forming a Cutting-Edge Large Model Technology Cluster
At the AI Large Model Development Sub-forum of the Zhongguancun Forum held on May 28, the 'China AI Large Model Map Research Report' was officially released. This report was jointly compiled by the China Science and Technology Information Institute, the Ministry of Science and Technology's New Generation AI Development Research Center, and related research institutions. Using visual mapping methods, the report analyzes the developmental characteristics of China's large models, publishes a domestic large model capability ranking, and also reveals existing problems and shortcomings in related developments.
The report states that institutions like Google and OpenAI in the United States continue to lead the technological frontier of large models, while an increasing number of countries including China, Europe, Russia, Israel, and South Korea are investing in large model research and development. From the global distribution of released large models, China and the United States significantly lead, accounting for over 80% of the global total, with the U.S. consistently maintaining the highest number of large models worldwide.
Currently, in the field of AI large models, China is maintaining synchronous growth with the United States. In technical branches such as natural language processing, machine vision, and multimodal applications, China is keeping pace and developing rapidly. A group of influential pre-trained large models such as Pangu, Wudao, Wenxin Yiyan, Tongyi Qianwen, and Xinghuo Renzhi have emerged, forming a large model technology cluster that closely follows global frontiers.
Additionally, the report analyzes 79 publicly released large models in China based on open-source information. The results show that large model R&D teams are distributed across 14 provinces/regions, with Beijing and Guangdong having the highest concentration, indicating relatively strong regional clustering. In terms of domain distribution, natural language processing remains the most active focus area for large model development, followed by multimodal fields. There are relatively fewer large models in computer vision and intelligent speech domains. Meanwhile, various innovation entities including universities, research institutions, and enterprises are actively participating in large model R&D, though collaborative development between academia and industry remains limited.
Furthermore, the report's nationwide survey of computing infrastructure distribution reveals that Beijing, Guangdong, Zhejiang, and Shanghai host the highest number of large models. These four regions also accounted for the most AI server purchases in recent three years, demonstrating significant correlation and providing crucial support for large model development. Simultaneously, local governments are supplementing rapidly growing AI computing demands through public computing resources, offering additional computational support for large model R&D.
Looking deeper, as China's earliest and most professionally recognized comprehensive evaluation benchmark, the Chinese General Large Model Benchmark (SuperCLUE) released its June rankings following this report. The rankings consist of four components: overall leaderboard, basic capabilities list, Chinese characteristics list, and 7-billion-parameter category list, evaluating model capabilities across different dimensions. Professional capabilities include middle school, university, and specialized examinations covering over 50 competencies from mathematics, physics, and geography to social sciences.
Industry analysis indicates that based on the June rankings, GPT-4 excels in both comprehensive and individual capabilities, significantly outperforming the second-place GPT-3.5-turbo. Following closely are GPT-3.5-turbo and Claude, which are on par in terms of scores. However, domestic large models in China are still in a catching-up phase, with performance across various aspects needing improvement. The pace of catching up in the AI model competition with the US needs to accelerate. Notably, the latest November rankings from SuperCLUE show a widening gap in overall scores and individual capabilities between Chinese and US AI models.
The Economist: Chinese Large Models Lag Behind the US by Two to Three Years, May Eventually Catch Up
On May 9, 2023, The Economist published an article titled How Strong is China in Generative AI? The article compares the AI capabilities of China and the US through charts on metrics such as the number of papers, systems, computing power, and chip hardware. It points out that Chinese large models currently trail behind the US by two to three years, attributing this gap to differences in training data, hardware like chips, and tech talent. However, it suggests that the two countries may eventually possess comparable AI capabilities in the future.
The article states that in recent years, China has led the US in certain metrics measuring AI capabilities. In 2021, 26% of global AI research papers originated from China, compared to only 17% from the US. Nine of the top ten institutions by AI paper output are based in China. Additionally, according to a commonly cited benchmark, the top five laboratories in computer vision are all located in China.
However, the US holds a clear advantage in the field of "foundational models" that power generative AI systems. For instance, ChatGPT and its underlying advanced models were developed by US-based startup OpenAI. Other American companies, ranging from smaller firms like Anthropic or StabilityAI to tech giants like Google, Meta, and Microsoft, also possess robust systems. While Baidu has created Ernie Bot as a counterpart to ChatGPT, the industry consensus is that it hasn't yet reached ChatGPT's level of sophistication.
As a result, industry experts conclude that China is approximately two to three years behind the US in developing foundational models. Three primary factors contribute to this gap: The first revolves around data. According to W3Techs, an internet research site, 56% of global websites host English content, whereas only 1.5% feature Chinese. This disparity benefits US model developers. Dr. Fu Yiqin from Stanford University notes that Chinese users primarily interact online via apps like WeChat and Weibo, where most content isn't indexed by search engines, making it challenging for AI models to ingest this data during training.
The second reason is related to hardware. In 2022, the United States imposed export controls on key AI technologies to China, including microprocessors used in cloud computing data centers and chip manufacturing tools that would allow China to produce such semiconductors independently. This has impacted China's large-scale model development. An analysis by the UK think tank Centre for AI Governance of 26 Chinese large models found that over half rely on chips from the American company Nvidia. Meanwhile, China's largest chip manufacturer can only mass-produce chips that Taiwan's TSMC was manufacturing three to four years ago.
The third reason is the difficulty Chinese AI companies face in attracting talent from the United States. Currently, the U.S. remains highly attractive to global tech talent: two-thirds of U.S. AI experts publishing in journals were born abroad, with Chinese engineers making up about a quarter of this top-tier group. Many Chinese AI researchers have studied or worked in the U.S. before returning with their expertise, but factors like heightened U.S.-China tensions have led to a decline in this group.
The Economist argues that while these factors pose obstacles to China's AI development, whether they will hinder China's AI ambitions in the longer term is another matter. Currently, the Chinese government has provided large datasets to modeling institutions and expressed a desire to dismantle the 'walled gardens' of Chinese apps to release more data. At the same time, China is seeking workarounds in hardware and can use open-source models to somewhat mitigate the lack of chips and talent.
Therefore, China and the United States may ultimately possess comparable AI capabilities, even though China will pay additional costs due to U.S. sanctions during this process. However, if the competition between their large AI models remains evenly matched, another advantage of the United States could make it the ultimate winner in AI—its ability to widely apply cutting-edge technologies throughout its economic system.
Comparative Analysis of China-US Policies: The Two Nations' Visions Are Not Incompatible, Harmonious Cooperation Could Create a Better Future
On October 12, 2023, the official WeChat account "Social Science Library" of the Economic and Technological Sociology Research Office at the Institute of Sociology, Chinese Academy of Social Sciences, published an article titled "Comparative Analysis of China-US Artificial Intelligence Policies." This article was co-authored by Emmie Hine, a Master of Social Science of the Internet from the Oxford Internet Institute, and Luciano Floridi, Professor of Philosophy and Ethics of Information at the Oxford Internet Institute and Director of the Digital Ethics Lab.
The article employs a combination of quantitative and qualitative methods to analyze the AI development policies of the United States and China. While outlining the formation of policies in both countries, it focuses on the differences in technological philosophies and historical-cultural traditions behind their policy-making. From a deeper perspective, it dissects the differences in AI policies against the backdrop of great-power technological competition, filling a gap in comparative studies of China-US AI policies.
The article mentions that in terms of policy evaluation, the Biden administration's legislative and social governance initiatives combine the Obama administration's efforts to promote diversity with the Trump administration's focus on American leadership. First, in the Biden administration's vision for a good AI society, the government will continue to emphasize free-market principles in research and development while also taking measures to reshape the industry. Second, the U.S. explicitly views China as a competitor, promoting multilateral initiatives while emphasizing diversity and ethical credibility. The Biden administration has redefined the beneficiaries of AI as all Americans and U.S. allies, excluding China.
Currently, China's central and local planning still prioritizes the practical application of AI technology over fundamental research. According to the "Development Plan," achieving global leadership in AI is China's primary goal, including major technological breakthroughs by around 2025, the formulation of relevant regulations, and the development of world-leading AI by around 2030 to become the "world's leading AI innovation center." To achieve these goals, China's central government is collaborating with local governments and the private sector to advance these initiatives. Numerous specific measures aim to maintain social stability while encouraging innovation and technological progress.
Overall, the U.S. AI strategic vision has undergone significant changes over the past three administrations. Under the Biden administration, the U.S. has adopted a more hands-on (though still market-oriented) approach, emphasizing the value of American leadership and innovation, as well as strengthening close cooperation with allies. Driven by the Protestant work ethic at the individual level, this represents a broader vision of technological excellence, which includes both global cooperation and competition with China.
In contrast, China's external vision for AI development includes gaining global leadership and fostering cooperation based on fundamental human values. Internally, Confucianism continues to profoundly influence China's planning and practice of a 'good AI society'. Therefore, China's cultural traditions and current political system drive the country to seek a development model that balances AI innovation with maintaining social stability.
The article further states that while neither China's nor the U.S.'s development philosophies were formed solely to address international technological competition, their visions are not irreconcilable. Protestant ethics emphasize the individual while Confucianism emphasizes society, yet both aim for the benefit of the public. However, neither domestic nor international conflicts serve this goal, making international cooperation the only path to achieving a 'good AI society'.
Governments should transcend traditional competitive geopolitical dynamics, adopt pluralistic value positions, and engage in effective dialogue while acknowledging multiple governance approaches and models. This would help outline specific standards for shared values in a good global AI society. Although the current competitive dynamic between China and the U.S. makes this vision challenging, if both sides can examine and understand each other's perspectives, AI could become part of humanity's collective project under harmonious cooperation between the two major powers - realizing the vision of a 'good global AI society'.