- The United States still maintains its leadership in AI development infrastructure, both in hardware and software. As a result, China currently still needs to import chips and relies on U.S. technologies.
- Nevertheless, despite infrastructure limitations, China has been able to develop AI models with intelligence comparable to those of the United States, reflecting the country's strong domestic technological capabilities and the rapid progress over the past few years.
- In the future, AI competition may extend beyond advanced technology development to include the practical application of AI across the economy and various industries, which could become a decisive factor for long-term competitive advantage.
Infrastructure Layer
Developing AI models requires advanced chips for data processing and storage (hardware), as well as software systems that operate alongside these chips to execute instructions. These components work together to ingest data, train models, and further develop them into functional AI systems.
Hardware: China still faces limitations in its ability to manufacture advanced chips.
The United States remains the global leader in the semiconductor industry. In 2024, U.S. chip companies accounted for roughly half of worldwide semiconductor sales (Figure 1).
The United States' key advantage lies in its advanced chip design and manufacturing technologies, which are highly complex and require specialized expertise. This has enabled the U.S. to consistently maintain its leadership in the global market, particularly in the processing chip segment, which is critical for AI development.
Although China is now capable of producing its own advanced processing chips, their performance still lags behind that of the United States (Figure 2).
Major Chinese companies such as Huawei have developed their own processing chips, including the Ascend series, to support domestic demand. However, the performance of Chinese chips still lags behind that of the United States. Tests show that Huawei's newly developed Ascend 910C, which is now being deployed in China, delivers only about 60% of the performance of NVIDIA's H100, even though the H100 is no longer the latest U.S. chip.
This highlights the clear technological gap between China and the United States—especially when compared with NVIDIA's Blackwell series (B200 and B300), which are the current flagship chips and offer significantly higher performance than the H100.
In addition, China continues to face insufficient manufacturing capacity to meet domestic demand, despite ongoing investment in expanding production.
The main source of pressure comes from restrictions on China's access to advanced chip manufacturing equipment, particularly EUV lithography machines from the Dutch company ASML. These machines incorporate critical components and technologies originating in the United States. As a result, the United States can enforce export controls on such equipment under the Foreign Direct Product Rule (FDPR), which states that any product or equipment manufactured anywhere in the world using U.S. technology is subject to U.S. export regulations.
As a result, these restrictions prevent China from mass‑producing advanced chips, thereby slowing the progress of its domestic semiconductor development relative to that of the United States.
For these reasons, China still needs to increase its imports of chips.
China's chip imports have continued to rise over the past several years, reflecting growing demand that is driven by the expansion of the AI industry (Figure 3). In 2024, U.S. chips accounted for as much as 70% of total chip usage in China, while the use of domestically produced chips remained limited (Figure 4).
Software: The United States continues to lead in the systems that operate alongside chips for AI training.
Beyond the processing capabilities of chips, the software used for issuing instructions and training models is equally important. Even when two chips offer the same level of computational power, differences in the efficiency of the training software can lead to significant variation in the quality of the resulting AI models.
This is one of the key reasons NVIDIA continues to maintain its leadership in the global AI chip market. The company's software ecosystem—particularly CUDA, which is designed to work seamlessly with NVIDIA's own chips—remains highly effective at unlocking the full performance of the hardware in real‑world applications.
When comparing NVIDIA's software performance with that of its direct competitor, AMD's ROCm (Figure 5), CUDA consistently delivers faster and more efficient performance across nearly all categories of computational workloads used in AI, including training, inference, parallel computing, and optimization.
Meanwhile, there are issues with the software developed by Chinese chip developers:
Although Huawei has been advancing its CANN software ecosystem alongside its Ascend‑series AI chips (such as the Ascend 910C), with the goal of supporting domestic use of Chinese chips both now and in the future as demand rises, AI developers in China still report that Ascend chips running on CANN often encounter stability issues, especially in complex operations. In addition, they requires a high level of coding expertise to unlock the chip's full efficiency. This stands in contrast to NVIDIA's CUDA, which is more stable and easier to use.
Therefore, in terms of AI infrastructure—both advanced processing chips (hardware) and software for model development (software)—the United States still maintains a clear technological advantage in the AI industry.
Model Layer
Despite these constraints, China has managed to develop AI models whose intelligence is on par with those of the United States.
Although China lags behind in both chips (hardware) and training software, Chinese AI models have nonetheless reached performance levels that are nearly comparable to those of the United States.
Data from Chatbot Arena shows that the performance gap between Chinese and U.S. models has been steadily narrowing (Figure 5). At the same time, China's AI development has become more concrete, with the launch of models such as DeepSeek, Qwen, Kimi, and ERNIE onto the global market. These models demonstrate intelligence that can compete with leading U.S. models like OpenAI, Copilot, and Gemini (Figure 6).
However, to advance its AI development, Financial Times reports that China has been importing advanced U.S. processing chips through third countries, as well as conducting AI model training in overseas data centers to circumvent U.S. restrictions on its access to high‑performance chips.
One of the key factors driving the development of China's AI models is the steadily increasing number and quality of AI talent.
Driven by investments in education and policies aimed at attracting researchers back to the country, China has rapidly expanded its pool of AI specialists. Recent data shows that researchers from the United States and China together account for 57.7% of all AI researchers worldwide, with approximately 63,000 based in the U.S. and around 53,000 in China.
China's AI talent base has grown dramatically over the past few years—rising from about 10,000 people in 2015 to more than 52,000 in 2024. This sharp expansion reflects the rapid development of a high‑quality talent pool, which is essential for advancing AI innovation.
In addition to talent-related factors, government policy support in China has also been a key driving force accelerating the country's AI model development.
The Chinese government has played a significant role in creating an environment that systematically supports AI research and development. This has been achieved through long‑term investments in infrastructure, funding for research, and initiatives that promote the practical adoption of AI across various industries. These efforts help reduce structural constraints faced by the private sector, accelerate experimentation and the development of new models, and enable China to offset its disadvantages in hardware and training software.
This demonstrates China's growing potential in the race to develop AI models, suggesting that the gap between China and the United States will continue to narrow in the future.
Application Layer
Beyond infrastructure and model development, the actual adoption of AI across economic sectors and industries is another crucial factor that reflects a country's ability to generate added value from AI technologies.
Although the United States remains the leader in chip technology and AI models, China places greater emphasis on the widespread application of AI.
According to the latest data from the China Internet Network Information Center, China recorded 515 million AI users in the first half of 2025, nearly double the number at the end of 2024. This reflects how AI has already integrated into daily lives for a large share of the population.
AI adoption in China is spreading across regions and industries, for example:
- Chongqing and Chengdu are focusing on applying AI in the automotive and electronics industries, including the development of pilot projects for smart vehicles and smart factories.
- Wuhan and Xi'an, which serve as major hubs for universities and key research institutions, are driving the use of AI in industrial automation, robotics, and high‑precision manufacturing. As a result, cities in China's inland regions are emerging as strategic centers for advanced manufacturing.
- Shenzhen, China's innovation hub, is advancing the application of AI in industry, robotics, and hardware development, supported by a startup ecosystem backed by global technology companies such as Huawei, DJI, and Foxconn.
Moreover, the Chinese public appears far more open to adopting new technologies. A survey by Stanford University shows that 83% of Chinese respondents view AI as beneficial, compared with only 39% in the United States. This openness has allowed China's AI user base and experimentation to grow rapidly and significantly (Figure 8).
In addition, when looking at the actual usage of AI models, China has shown remarkable progress during 2024–2025. The share of AI model deployments using Chinese models has continued to rise, eventually surpassing those of the United States and Europe (Figure 9). A key factor behind this trend is the Chinese government's strategy of promoting open‑source AI models, enabling companies, developers, and institutions to freely use or customize them. This has driven the widespread adoption of Chinese models both domestically and internationally. In contrast, U.S. developers tend to focus on closed‑source models accessible only through APIs, which limits their flexibility and broader usage.
Even open‑source models from the U.S. that were once highly popular—such as Meta's Llama—are now facing stronger competition from Chinese models like Qwen (from Alibaba) and DeepSeek. According to recent data from the Hugging Face platform, Qwen has already surpassed Llama in total downloads, indicating a clear rise in global adoption of Chinese open‑source AI models.
Overall, these factors show that China is emerging as a leader in the application layer of AI. Its promotion of open‑source model development, combined with a large and ready user base, has accelerated the widespread adoption of AI across the economy.
Outlook for the Future of the Chinese and U.S. AI Industries
Therefore, the landscape of AI competition between China and the United States in the coming years may diverge from the traditional view that focuses solely on the race for cutting‑edge technologies. The essence of competition may increasingly shift toward the ability to effectively apply AI across economic sectors and industries, rather than simply possessing the most advanced technical capabilities.
From this perspective, technological advancement alone may not be the decisive factor if a country is unable to translate AI's potential into broad‑based economic benefits—such as improving productivity, reducing costs, and enhancing the competitiveness of the manufacturing and services sectors in tangible ways. Within this context, China's current level of AI technology, while perhaps not at the global frontier, may still be sufficient to meet domestic needs and aligns with the structural priorities of the Chinese economy.