China’s AI Advantage

Faster, Cheaper Commercialization of Large Models, Says Kai-Fu Lee

As AI technology moves from breakthrough innovations to widespread commercial applications, China’s AI landscape is gaining confidence in its ability to compete globally, particularly with Silicon Valley.

According to Kai-Fu Lee, former Google China president and founder of AI startup 01.AI, Chinese AI firms have a unique edge over their U.S. counterparts: faster and more cost-effective deployment of large language models (LLMs).

At a recent event, 01.AI announced the launch of its flagship pre-trained model, Yi-Lightning, which has achieved a global ranking of sixth place in the LMSYS leaderboard, surpassing OpenAI’s GPT-4 May version and Anthropic’s Claude 3.5 Sonnet.

This marks a significant milestone for Chinese AI, as Yi-Lightning has become the top-ranking model from China.

Lee, speaking to Chinese media outlet TMTPost on October 21, highlighted this achievement:

Yi-Lightning is the first Chinese large model to beat most U.S. competitors, including OpenAI’s GPT-4 May version. It’s a world-class model that not only excels in performance and inference speed but also comes at a very low cost, making it ideal for both app and enterprise use cases.

China’s Strength: Commercialization Efficiency and Cost-Effectiveness

While Lee acknowledges that Chinese AI remains behind the U.S. in certain respects, he counters the notion that China lags by a decade or more. Based on the rapid advancements demonstrated by companies like 01.AI, Lee believes the gap has narrowed to just a few months in key areas.

In an op-ed for the Financial Times, Lee elaborated on China’s strengths, particularly in producing affordable AI inference engines.

“China excels at delivering AI applications that are both efficient and cost-effective,” Lee said. “The country’s vast pool of highly skilled and hardworking engineers gives China a distinct advantage over the U.S.”

Lee emphasized that while China may not be able to conduct unprecedented, open-ended research on par with the U.S. when budget constraints are lifted, it can reliably bring AI solutions to market more efficiently and at a lower cost.

A New AI Frontier: Cost-Effective Models

Founded in May 2023, 01.AI has quickly risen to prominence, becoming a unicorn in November 2023 after raising billions of dollars in investment.

The company is led by Lee and aims to build a global AI-first productivity platform. Its latest model, Yi-Lightning, is a testament to the company’s cost-efficiency and technical prowess.

Unlike U.S. firms that may spend exorbitant amounts on AI research, 01.AI’s model pre-training costs a mere $3 million, only 3% of what OpenAI reportedly spent training GPT-4.

Yi-Lightning has been optimized for various enterprise applications, and the model can be fine-tuned at a fraction of the cost of U.S. counterparts.

Lee highlighted the model’s cost-effectiveness: “Yi-Lightning can generate answers at a speed 40% faster than before, with an inference cost of just 0.99 yuan ($0.14) per million tokens—significantly lower than OpenAI’s smallest model, which costs $0.26 per million tokens, and GPT-4, which costs $4.40.”

The ability to offer such cost-effective AI solutions is a core part of 01.AI’s strategy, positioning the company to compete in both international consumer markets (To C) and domestic enterprise markets (To B).

In the latter, Yi-Lightning is already being deployed by retail and e-commerce companies, with one hospitality client reportedly seeing a 170% increase in GMV (gross merchandise volume).