PODCAST — Stationed in Shanghai for several years, Artefact CEO Édouard de Mézerac explains how China, cut off from US chips, has become one of the world’s leading AI powerhouses. He is the guest on Superpuissance, the podcast by Challenges.

When he arrived in Shanghai in the summer of 2016 with his wife and their two children, Édouard de Mézerac thought he was bringing expertise to the table. His employer at the time wanted to conquer China and relied on his skills in data analytics and machine learning. But things did not go as planned. Among his first projects, he and his team spent 18 months working for the tech department of JD.com, one of China’s e-commerce giants. The error rate in sales forecasting was hovering around 50%; the goal was to bring it down by five percentage points.

He discovered the power of Chinese engineers and the striking gap with the traditional, century-old or half-century-old players he had been supporting until then. In 2018, the founders of Artefact approached him to offer him the role of developing the group’s business in China. He accepted. Having returned to France in 2024 to take the helm of the company, he now coordinates enterprise AI strategy deployments from Paris.

The myth of a widespread lag

And the head of Artefact continues to closely follow developments in AI across China. One might have thought the country was falling behind under the impact of US embargoes on essential components, particularly microchips. China’s first AI strategic plan dates back to 2017, targeting global dominance by 2030. When it comes to large language models, it certainly got off to a late start, but it was already in the race for machine learning, especially in image recognition. One of its leaders, SenseTime, has become a major player in facial recognition.

To bridge the gap, China bet on open source, often in partnership with universities. The Hugging Face leaderboard bears witness to this: eight of the top ten open-source models are Chinese, including five versions of DeepSeek and three of Qwen, Alibaba’s model. Business models vary: Alibaba makes money every time a model runs on its cloud, while Moonshot AI, one of the new AI Tigers, charges for its model once consumption exceeds a certain threshold.

The chip and electricity bottleneck

On closed models, however, he estimates that China is still behind by at least six months. The culprit: computing power. Denied access to the latest Nvidia chips, Chinese industry is being pushed hard by authorities to manufacture its own components. But scarcity has sparked engineering creativity. Their models reportedly consume up to 100 times less electricity and chips than American models. He sees no reason why the country will not manage to catch up: time, talent, and capital are all aligned, even if these chips still depend on components from the Netherlands or Taiwan.

The other major challenge in this tech race is electricity, which accounts for 60% to 70% of the cost of a token, the unit of measurement in AI. Hence the East-to-West data transfer project, Eastern Data, Western Computing: data centers are built near new solar and wind energy sources in the desert West, while demand develops in the major cities of the East. As a result, Chinese models use fewer tokens for the same task, at a lower unit cost—an advantage leading some European and American players to consider using these models.

An optimistic and positive outlook on AI

The contrast is also cultural. Recent studies show that 83% of Chinese people view AI-powered products positively, compared to about 40% in the United States and 41% in France. Illustrating this enthusiasm, he recalls that in Shanghai last March, Tencent engineers were installing OpenClaw—an open-source tool for creating agents—onto passersby’s laptops for under 40 euros. The primary use case cited by new users: managing their stock market portfolio.