The United States currently has a lead over the People’s Republic of China (PRC) in generative artificial intelligence (GenAI) stemming from U.S. industry’s edge in the underlying technology and a more permissive approach to speech and content as an output of GenAI. Nonetheless, the United States could risk losing its advantage if the technology is not effectively integrated across the innovation ecosystem and broader economy. Both countries are expanding domestic partnerships and investments to develop new GenAI applications and current projections of GenAI’s worldwide impact estimate that it could add up to $4.4 trillion annually to the global economy.1 While the U.S. ecosystem includes a mix of general purpose and specialized GenAI innovation, the PRC ecosystem appears more focused on industry-specific AI applications, perhaps due to the Chinese Communist Party’s (CCP’s) discomfort with domestic proliferation of and experimentation with general purpose AI models. However, as more GenAI use cases emerge, sustained PRC government involvement could enable the technology to be deployed faster across many sectors.
| Area of Comparison | People’s Republic of China | United States |
|---|---|---|
| Government Players | While the PRC central government leads a number of national-level AI development initiatives, its involvement in the actual development of GenAI applications is centered around its government-backed academic institutions, including: – Beijing Academy of Artificial Intelligence – Tsinghua University – Zhejiang Lab – Shenzhen’s Peng Cheng Lab | Recent U.S. government efforts demonstrate growing recognition and strategic focus on AI, to include GenAI, but lack the enduring institutional and resource-marshaling power of PRC counterparts. To date, the government has played the role of convener, bringing together industry leaders to guide government thinking via organizations including: – National Security Commission on Artificial Intelligence (NSCAI) – National Artificial Intelligence Initiative Office (NAIIO) – National Artificial Intelligence Advisory Committee (NAIAC) U.S. departments and agencies, however, are launching initiatives to examine how best to leverage GenAI in government operations, such as the Department of Defense’s recently announced Task Force Lima.2 |
| Government Regulation | The PRC has already iterated on and is beginning to implement GenAI regulations that focus on industry, innovation, and asserting government control over its direction. The PRC’s rapid dissemination — and subsequent narrowing — of regulatory controls on GenAI reflect its attempt to balance between its national goals and domestic political considerations, and ultimately determine the right length leash for industry.3 | Current efforts on Capitol Hill to regulate GenAI via statutory changes are in preliminary stages, as the United States takes a more industry-led approach to developing GenAI norms and standards.4 Major U.S. tech players in the GenAI space, at the urging of the White House, have adopted a voluntary framework on how to address future safety and transparency of GenAI applications.5 Leading GenAI tech companies are also forming the Frontier Model Forum6 to coordinate future AI development. |
| Private Investment | A smaller and less sophisticated capital market, as well as regulatory ambiguities, have resulted in significantly less private investment in GenAI in China, compared to the United States. – PRC AI startups raised $13.5 billion worth of private capital in 2022.7 – Baidu launched a GenAI venture fund for domestic startups worth $145 million.8 | U.S. firms benefit from easy access to the world’s largest and most dynamic capital markets. While current demand for venture capital in the United States outweighs supply,9 GenAI startups are still benefiting from massive investment as well as a positive acquisition trend. – U.S. AI startups raised $47 billion billion worth of private capital in 2022.10 – Inflection, Anthropic, Cohere, Runway, and Mistral AI have all raised megadeals in 2023, with Inflection raising $1.3 billion.11 – Snowflake paid an estimated $150 million for Neeva, while Databricks preempted a VC round for MosaicML with a $1.3 billion acquisition, both in 2023.12 |
| Major Tech Players & Frontier Models | Major PRC tech players are inclined to focus on the industrial applications of GenAI.13 The CCP’s regulations and controls mean that these players are limited to conforming with China’s national goals, posing potential constraints on their future innovative capabilities. Currently, China’s domestic major tech players have been producing inferior frontier LLMs compared to global counterparts, but have shown promise in other types of transformer-based AI models.14 Major Tech Players & Their Frontier Models: – Baidu [Ernie] – Huawei [Pangu] – Alibaba DAMO Academy [M6-10T] – Beijing Academy of Artificial Intelligence (BAAI) [WuDao 2.0] – Tsinghua University [BaGuaLu, WuDao, M6, CogView] | The private sector dominates most GenAI innovation in the United States, driven largely by commercial incentives. The current disjuncture between GenAI goals within the public and private sectors means that major tech players’ goals may not always align with the government’s national security concerns. In spite of this, major domestic tech players continue to demonstrate globally-superior GenAI frontier models. Major Tech Players and Their Frontier Models: – OpenAI [GPT-4, ChatGPT series] – Google Deepmind [PaLM 2] – Anthropic [Claude 2] – Microsoft Research [GitHub CoPilot] – Meta [LLaMA-2] |
| Compute Infrastructure | China’s AI development is likely to remain dependent on U.S.-made hardware and software, absent a major disruption to the algorithmic architecture driving or infrastructure underlying the field. China’s AI sector predominantly depends on U.S.-developed Nvidia processors for model training. China also has a strong state-funded supercomputing ecosystem to include at least one world-leading supercomputer based on domestic hardware that is dedicated to AI applications.15 Due to U.S. export controls, China’s commercial AI firms have limited access to Nvidia GPUs, which are a core component of their efforts to build advanced compute clusters for their LLM development.16 China’s AI developers largely used GitHub-based open-source software as a foundation for their own model development. Domestically-developed open-source versions have yet to gain broad traction, but include: – PaddlePaddle (Baidu) – Atlas (Huawei) | U.S.-developed AI hardware and software has been foundational to the transformer model paradigm, and will likely remain so barring a disruption to the algorithmic architecture driving or infrastructure underlying the field. Leading U.S. AI firms are building increasingly large GPU-based high performance computers (HPCs) for their LLM model development needs; current and under development HPCs include Nvidia’s Selene, Meta’s AI Research SuperCluster, Tesla’s DOJO, and Google’s TPU v4 cluster dedicated to LLM training. Other approaches for LLM training include the use of specialized AI processors like Groq and Cerebras and/or accessing compute clusters via cloud providers such as Nvidia, Microsoft, Cerebras, and AWS.17 U.S.-developed software has been foundational in the global development of AI models and include ML platforms like: – TensorFlow (Google) – PyTorch (Meta/Linux) – CUDA (Nvidia) – CNTK (Microsoft) |
| Data | As the world’s second-most-populous country, China generates massive amounts of data and will likely continue to do so.18 While PRC firms are developing English-language LLMs, when it comes to training Chinese-language LLMs specifically, the PRC faces some constraints as less than two percent of the Internet is in the Chinese language.19 – Generated 7.6 zetabytes (ZB) of data (23 percent of total data generated globally) in 201820 – Estimated to generate 48.6ZB (27.8 percent of the world’s data) in 202521 | Despite the United States’ relative lag in data generation compared to China, U.S. AI developers have a massive advantage when training LLMs due to nearly 60 percent of the Internet using the English language.22 Additionally, U.S. tech giants maintain a broad global presence which facilitates even greater data flows into the U.S, evening the lag in data generation between the U.S. & PRC. – Generated 6.9ZB of data (20 percent of global data volume) in 201823 – Estimated to generate 30.6ZB (17.4 percent global data volume) in 202524 |
| Talent | Many of the world’s top AI researchers are based in the PRC, however the country is currently experiencing brain drain to other countries with an estimated 400,000 PRC scientists working abroad.25 However, it is taking steps to mitigate this through national programs such as the Thousand Talents program.26 Additionally, the CCP has increased domestic efforts to create university-level AI programs, training academies, and other initiatives in the hopes of retaining domestic AI talent.27 | Many of the top AI specialists and academics are based within the United States; however, the immigration system has made it more difficult for foreign researchers to stay and work in the U.S. in recent years.28 Additionally, some speculate that the industry is siphoning talent from academia, potentially leading to long-term issues in AI education capabilities.29 Domestic AI education remains focused more on university-level education, instead of a comprehensive approach including primary and secondary levels of education (although some state-led programs are working to change that).30 |
Endnotes
- Michael Chui, et al., The Economic Potential of Generative AI: The Next Productivity Frontier, McKinsey & Company (2023).
- This task force will analyze and employ generative AI applications across the Department of Defense to streamline various processes. See DOD Announces Establishment of Generative AI Task Force, U.S. Department of Defense (2023).
- The measure defines Generative AI, as well as GAI service providers. It aims to regulate GAI when it has public applications, thus, service providers within academia, industry/organizations do not have to conform to the regulatory measures. It also notes that outside GAI providers working or deploying products within China must conform to these regulations. These regulations particularly target the GAI providers role in deploying the technology, such as the providers are responsible for monitoring GAI content, marking generated content, and protecting user information. See Jenny (Jia) Sheng, et al., China Finalizes Its First Administrative Measures Governing Generative AI, Pillsbury (2023).
- Andrew Solender & Ashley Gold, Scoop: Schumer Lays Groundwork for Congress to Regulate AI, Axios (2023).
- Ryan Heath & Ashley Gold, White House Gets AI Firms to Take Safety Pledge, Axios (2023).
- Dan Milmo, Google, Microsoft, OpenAI and Startup Form Body to Regulate AI Development, The Guardian (2023).
- Just How Good Can China Get at Generative AI?, The Economist (2023).
- Connie Loizos, As Part of AI Push, Chinese Tech Giant Baidu Is Now Rolling Out An AI Venture Fund, TechCrunch (2023).
- Kyle Stanford, et al., Accounting for the Overcapitalization of VC, Pitchbook (2023).
- Artificial Intelligence Index Report 2023 Chapter 4: The Economy, Stanford University (2023).
- Brendan Burke, et al., Artificial Intelligence & Machine Learning Report Q2 2023, VC Trends and Emerging Opportunities, Pitchbook (2023).
- Brendan Burke, et al., Artificial Intelligence & Machine Learning Report Q2 2023, VC Trends and Emerging Opportunities, Pitchbook (2023).
- Josh Ye, China’s Slow AI Roll-Out Points to its Tech Sector’s New Regulatory Reality, Reuters (2023).
- Zeyi Yang, The Bearable Mediocrity of Baidu’s ChatGPT Competitor, MIT Technology Review (2023); Melissa Heikkilä, New AI Systems Could Speed Up Our Ability to Create Weather Forecasts, MIT Technology Review (2023).
- Sebastian Moss, China May Be Planning 10 Exascale Computers by 2025, Data Center Dynamics (2022); Nicole Hemsoth, New Chinese Exascale Supercomputer Runs ‘Brain Scale AI’, The Register (2022).
- Justin Hair, Bytedance (TikTok) Orders $1 Billion of NVIDIA GPUs: Boosting AI Capabilities Amidst Export Ban Concerns, Medium (2023).
- Agam Shah, Nvidia Launches AI Factories in DGX Cloud Amid the GPU Squeeze, HPCWire (2023).
- China is building one of the world’s strictest data governance regimes to manage its massive data holdings. See Eva Xiao, China Passes One of the World’s Strictest Data Privacy Laws, Wall Street Journal (2023).
- Languages Most Frequently Used for Web Content as of January 2023, By Share of Websites, Statista (2023).
- Saheli Roy Choudhury, As Information Increasingly Drives Economies, China is Set to Overtake the US in Race for Data, CNBC (2019).
- Saheli Roy Choudhury, As Information Increasingly Drives Economies, China is Set to Overtake the US in Race for Data, CNBC (2019).
- Languages Most Frequently Used for Web Content as of January 2023, By Share of Websites, Statista (2023).
- Saheli Roy Choudhury, As Information Increasingly Drives Economies, China is Set to Overtake the US in Race for Data, CNBC (2019).
- Saheli Roy Choudhury, As Information Increasingly Drives Economies, China is Set to Overtake the US in Race for Data, CNBC (2019).
- Paul Scharre, To Stay Ahead of China in AI, the U.S. Needs to Work with China, TIME (2023); The Global AI Talent Tracker, Macro Polo (last accessed 2023).
- Alison Snyder, China Talent Program Increased Young Scientists’ Productivity, Study Says, Axios (2023).
- Liu Caiyu, School Students in East China’s Zhejiang to Study AI as Compulsory Course, Global Times (2023); Zhejiang To Make AI Courses Compulsory in Primary, Secondary Schools, China Daily (2023).
- Kaveh Waddell, Most Top AI Researchers in the U.S. Are Foreign, Axios (2019).
- The Catch-22 Facing Academia and the Tech Sector in the War for Top Talent, Toptal (last accessed 2023).
- Diana Gehlhaus, Why States Need an AI Education Agenda–Now!, Council on Foreign Relations (2022).