China’s Domestic AI Competition Heats Up

This past summer, at the World Artificial Intelligence Conference in Shanghai, Baidu’s CEO, Robin Li Yanhong, asked a surprising question: Does China have too many AI startups? As he put it: “In 2023, intense competition among over 100 LLMs has emerged in China, resulting in a significant waste of resources, particularly computing power. … How about real-world applications? Who has benefited from them?”

Following the success of ChatGPT and restrictive U.S. sanctions on chips, most expected China to concentrate its resources to compete in the AI space. Instead, Beijing is pursuing a strategy being referred to within China as the “war of [a] hundred models.” The strategy appears to be similar to China’s strategy in EVs, where it offered a wide array of subsidies. The EV strategy resulted in impressive industry leaders such as BYD and Li Auto, but also a glut of over 200 EV manufacturers, many of which are unprofitable. Similarly, many of China’s AI startups are currently facing financial difficulties.

The hundreds of AI startups have driven intense price wars within China, leading some to look overseas. Yuan Jinhui, CEO of the AI startup SiliconFlow, told the South China Morning Post: “No matter if your target audience is business or consumers, in terms of profit, the overseas market is much better than [in China] domestically.” Some startups are even taking steps to “define themselves as global corporations” and go “overseas from the start.” Another reason for the focus overseas are China’s own political requirements. As the Wall Street Journal reported in its July 16 article, “China Puts Power of State Behind AI—and Risks Strangling It,” startups within China are required to submit a data set of “5,000 to 10,000 questions that the model will decline to answer.” With limited funding in a fast-moving field, this can be a distraction and use up valuable resources.

To win internationally, Chinese AI startups will need to be better and cheaper than the competition. China has a number of inherent advantages. The Organization for Economic Cooperation and Development (OECD) reports that China contributed to  more than 20 percent of AI research in 2023; more than the EU and India combined. According to the World Intellectual Property Organization, China also dominates the global race for generative AI patents, having “six times more than second-place U.S.” in the past 10 years. Despite China’s research proficiency, its AI models are behind. In Chatbot Arena, one of the most-watched leaderboards for AI, China does not currently feature in the top 5. The leaderboard is based on user votes in a blind comparison.

Professor Zhu Feida at the Singapore Management University believes it is only a matter of time before China catches up. While China faces limits on access to advanced AI chips, it has an advantage on the equally crucial power supply, where the U.S. “will soon be stretched to its limits.” A leading AI data center requires not only hundreds of thousands of chips, but the energy to power and cool them. China is currently building two-thirds of the world’s wind and solar projects. Additionally, the “hundred models” strategy raises the odds of a single startup coming up with a breakthrough innovation. How did China’s AI ecosystem develop and where are these startups coming from?

Nurturing innovation

China began investing heavily in AI long before the release of ChatGPT. In July 2017, China’s state council put forth the “New Generation Artificial Intelligence Plan,” declaring its desire to build a “first-mover advantage in the development of AI.” The plan also declared that by 2025, “China will achieve major breakthroughs in basic theories for AI” and by 2030, China will become “the world’s primary AI innovation center.” The investments from this plan focused on university research and helped China’s domestic talent base in machine learning and AI.

The “Future of Go” summit in May 2017 is commonly seen as the genesis for China’s “New Generation Plan.” At the summit, Google’s AI program AlphaGo defeated five top Chinese Go players. Google did not plan on spurring massive Chinese investment in AI, with board games being a well-known way to demonstrate computer breakthroughs. In 1996, IBM’s Deep Blue became famous for defeating chess champion Gary Kasparov. As a more complex board game, Go was a natural next challenge for computer science. However, for China, having its top players in its own national pastime defeated by an American company was seen domestically as a “Sputnik Moment.” Beyond investing at the university level, in November 2017 China began tasking Baidu, Alibaba, Tencent, and iFlyTek with building “open innovation platforms” for different sub-areas of AIs, establishing them as national champions for the AI space. All four continue to invest in AI models today and the program has grown to at least 15 companies.

Varied paths to success

A domestic AI startup ecosystem has developed within China, helped by recent government support such as subsidies for data center power and purchasing domestic chips. Five generative AI startups within China—Zhipu AI, Baichuan AI, Moonshot AI, Minimax, and 01-AI—have reached unicorn status. Zhipu AI was founded by Jie Tang, a professor at Tsinghua’s Department of Computer Science. He was tasked by China’s newly created Beijing Academy of Artificial Intelligence to build “China’s first super-scale natural-language AI” model. Seen as a rival  to OpenAI’s GPT-3, the model was completed in 2021 with the startup Zhipu AI launched to develop commercial use cases.

The startup Zero One Everything (01-AI) was launched by Kai-Fu Lee, a Taiwanese businessman and former president of Google China. Lee, who wrote the 2018 book focused on China’s AI advantage, AI Superpowers, had already been investing in AI startups but was inspired to start his own after ChatGPT’s release. His company, 01-AI, is built upon open-source projects like Meta’s Llama series, which his team credits for reducing “the efforts required to build from scratch.” Through an intense focus on quality-control, 01-AI has improved on the public versions of these models. Instruction sets are used in AI to guide models for certain use cases. For the instruction sets in 01-AI’s Yi models, “every single instance has been verified directly by … machine learning engineers” with “multiple iterations based on user feedback.” The startup’s attention to detail seems to be paying off; its “Yi-Lightning” model is currently the top Chinese model on Chatbot Arena. But it’s still behind models from U.S. competitors such as OpenAI, Google, and Elon Musk’s xAI.

One of the best performing Chinese AI models, DeepSeek, is the spinoff of a Chinese quantitative hedge fund, High-Flyer Capital Management, which used high-frequency trading algorithms in China’s domestic stock market. As the Financial Times reported in its June 8 article, “The Chinese Quant Fund-Turned-AI Pioneer,” the fund was originally started by Liang Wenfeng, a computer scientist who began stock trading as a “freelancer until 2013, when he incorporated his first investment firm.” High-Flyer was already using massive amounts of computer power for its trading operations, giving it an advantage when it came to the AI space. DeepSeek built its own “Mixture-of-Experts” architecture, which uses multiple smaller models focused on different subjects instead of a giant, overarching model. As a result, DeepSeek believes its models can perform similar to leading models while using considerably fewer computing resources. According to DeepSeek, their latest model achieved “superior performance compared to closed-source models such as GPT4-Turbo [OpenAI], Claude 3 Opus [Anthropic], and Gemini 1.5 Pro [Google] in coding and math benchmarks.”

Despite the challenges, China’s AI startup ecosystem is highly dynamic and impressive. Top-tier talent, government support, and a strong domestic market position China to potentially become the AI leader. It is yet to be seen whether the “100 models” strategy is the correct one. Perhaps Baidu’s Li is right. If China had limited chip access to only a few companies, it could be more competitive in rankings with the U.S.’s mega-models. It is also possible that if the chips were limited only to China’s tech giants, there would be no startups like DeepSeek willing to take risks on innovation. China’s progress in AI should continue to be closely watched, especially as the new administration’s approach to China comes into view.

About Chase Young ’24

Chase Young is a class of 2024 graduate of the Cornell Jeb E. Brooks School of Public Policy at Cornell University and a research fellow with the Emerging Markets Institute at the Cornell SC Johnson College of Business. Before joining the Emerging Markets Institute, Young interned in the global finance and business management program at JPMorgan Chase and was a research intern for the World Bank’s data development group. Young currently works as a consumer product strategy analyst at Texas Capital Bank.