The US-China AI Race: Technological Catch-Up Under Export Controls and Global Power Dynamics

The US-China AI Race: Technological Catch-Up Under Export Controls and Global Power DynamicsIn 2017, China unveiled its Next Generation Artificial Intelligence Development Plan, aiming to become a global AI innovation center and leader by 2030. However, the emergence of ChatGPT and subsequent US chip export controls have cast a shadow over this ambition

The US-China AI Race: Technological Catch-Up Under Export Controls and Global Power Dynamics

In 2017, China unveiled its Next Generation Artificial Intelligence Development Plan, aiming to become a global AI innovation center and leader by 2030. However, the emergence of ChatGPT and subsequent US chip export controls have cast a shadow over this ambition. This paper delves into the intense competition between the US and China in the AI field, analyzes the impact of export controls on Chinas AI development, and offers a perspective on the future global AI landscape.

I. The ChatGPT Shock and China's Rapid AI Rise

OpenAI's release of ChatGPT in late 2022 caught global attention, including China, off guard. At that time, Chinese tech companies were navigating a regulatory storm, with the tech sector losing approximately $1 trillion, consequently hindering AI development. The month before ChatGPT's launch, the Biden administration's export controls further restricted China's access to crucial chips needed for training large-scale AI models. These factors significantly increased the difficulty of China achieving its AI leadership goal by 2030.

However, China's AI industry didn't stagnate. In 2024, companies like Alibaba, Tencent, and DeepSeek launched competitive AI products, significantly narrowing the performance gap with OpenAI's models. Alibaba and DeepSeek's inference models approached the performance of OpenAI's GPT-O1 preview in some tests; Tencent's updated HunYuan large model surpassed top US open-source models in several benchmark tests; and DeepSeek-v3 topped prominent online leaderboards in the open-source arena, even competing with top closed-source systems from OpenAI and Anthropic. These breakthroughs demonstrated China's rapid AI development and progress, surprising many observers.

Eric Schmidt, former Google CEO, initially estimated a two-to-three-year US lead in AI. But by November 2024, he acknowledged China's progress particularly Alibaba and Tencent's achievements was rapidly closing the gap, surprising him and suggesting that he had underestimated Chinas ability to overcome the chip restrictions.

II. The US-China AI Competition: A Contest of National Power

The US-China competition in AI transcends national prestige; it profoundly impacts the global balance of power. AI technologies, especially intelligent systems capable of directing weaponry or executing cyberattacks, will provide decisive military advantages. Widespread applications of intelligent agents could also drive economic growth. As many nations must choose between these two AI systems, AI will become a powerful tool for global influence.

China's rapid rise has sparked questions about whether US semiconductor export controls can maintain its dominance.

III. Three Core Elements of AI Development: Data, Algorithms, and Computing Power

Building more powerful AI systems hinges on three core elements: data, innovative algorithms, and computing power.

Large language models like GPT-4 typically train on data from the internet, readily accessible to global developers. Algorithms, the innovative ideas for improving AI systems, also transcend borders through widely disseminated academic papers. Even without academic exchanges, China, with its vast talent pool, cultivates more top AI researchers than the US.

Unlike data and algorithms, however, advanced chips, as physical products, are susceptible to border restrictions. The advanced semiconductor supply chain is predominantly controlled by the US and its allies. US companies NVIDIA and AMD virtually monopolize the GPU market for AI data centers, while the manufacturing of these chips relies on companies like TSMC and ASML.

IV. US Chip Export Controls and Their Impact

To maintain its competitive edge, the US has implemented various restrictions, including the 2022 Biden administration export controls banning the sale of the most advanced chips to China. This policy continued measures from the Trump administration, aiming to curb China's access to chip manufacturing technology. These measures not only limited advanced chip access for China but also stifled the development of its domestic chip industry.

However, the effectiveness of US export controls is questionable. Some Chinese developers stockpiled large quantities of soon-to-be-restricted chips before the policy took effect. For instance, DeepSeek built a computing cluster of 10,000 NVIDIA A100 GPUs a year before the export controls. Furthermore, Chinese companies have acquired restricted chips through overseas shell companies or leased GPU resources from offshore cloud providers.

While US export controls restrict access to the most advanced semiconductors, they still allow the sale of less powerful chips. Determining which chips can be sold and which should be banned remains a complex challenge. NVIDIA introduced compliant chip versions for the Chinese market, which, despite reduced performance, still hold practical value in AI development, prompting further US restrictions in October 2023.

Lennart Heim, director of AI and computation at the RAND Corporation's Technology and National Security Policy Center, believes that export controls haven't comprehensively impacted Chinese AI development, largely because China can acquire nearly equivalent chips and because of the time lag for new chips to integrate into developer infrastructure.

V. Resilience and Innovation in China's AI Development

Despite chip restrictions, China has made remarkable progress in AI. Tencent's HunYuan large model excelled in multiple benchmark tests, even surpassing Meta's Llama 3.1 model, trained using less powerful, unrestricted NVIDIA H20 GPUs. This demonstrates that Chinese developers have optimized software to use hardware resources more efficiently. DeepSeek's DeepSeek-v3 model performed exceptionally well in the open-source arena, requiring surprisingly low computational power for training.

VI. Future Outlook: Continued Competition and the Possibility of Cooperation

While President-elect Trump's specific AI policy direction remains unclear, experts generally predict continued, and possibly expanded, export controls. Chinese companies are again massively stockpiling soon-to-be-restricted chips.

Ritwik Gupta, an advisor to the US Department of Defenses Defense Innovation Unit, argues that the US response strategy needs rethinking, moving beyond a whack-a-mole approach targeting hardware chips, and focusing on preventing the rise of military AI systems. He also acknowledges that US restrictions on other key aspects of the chip supply chain (such as controls on exports of ASML chip manufacturing equipment) have played a significant role in curbing the development of China's domestic chip industry.

Heim suggests that China may now be able to develop products comparable to top US open-source models, but these models still lag behind the strongest US closed-source models by about a year. He emphasizes this doesn't signify the failure of export controls and advocates against a binary view of their effectiveness.

Computational power for training AI models has increased dramatically over the past decade. US companies are building massive supercomputers with hundreds of thousands of GPUs, far exceeding the capacity needed for current leading AI models. Heim predicts that US chip export controls will significantly limit Chinas AI development capabilities, and that the impact will likely deepen over time.

Scott Shane, visiting scholar at the Carnegie Endowment for International Peace's Technology and International Affairs Program, notes Washington's current hesitation regarding whether to engage in AI negotiations with China. He believes dialogue with China on AI is crucial, as even with computational limitations, China's continued technological advancements could lead to the development of potentially dangerous AI systems in the future. As China's models approach parity with US models, the US needs to consider how to engage with China to ensure the safety of both nations' AI systems.

VII. Conclusion

The US-China AI competition is intensifying. While US export controls have impacted China's AI development, they haven't halted its rapid progress. The future will likely see continued competition, but the possibility of cooperation to address the shared challenges posed by AI technology is also worth exploring. This will profoundly influence the global balance of power and the future world order.