China is writing its own rules on AI

The best Chinese artificial intelligence models are less than a year behind their US counterparts, according to experts. Kai-Fu Lee, former head of Google China, said last week that the technological gap between the two nations was narrowing and predicted that AI apps would soon proliferate much faster in China, which is the only country that can get close to challenging US dominance on AI. The 2024 edition of Tortoise’s Global AI Index, launching on Thursday, shows the gap between China and the US on AI capacity is still significant – in some areas like private investment it has actually widened. But the data may be understating China’s position. In the past few years its AI capabilities have become harder to measure, as China builds a software ecosystem separate from the West.

Where China lags behind in model capability, it’s probably due more to the shape of the country’s AI ecosystem than to any gulf in technological sophistication. Developing large-scale AI models is capital- and compute-intensive. While the US market is dominated by a handful of tech giants, China’s is currently fragmented and involves a wider number of players, including established tech firms (Baidu, Alibaba) and startups like Moonshot and Kai-Fu Lee’s own 01.ai. These companies are fighting for market share in what has been called China’s “war of 100 models”. When the market consolidates around a few players, they are likely to concentrate resources building on fewer, but more powerful, models.

US policymakers’ efforts to slow China’s AI development by banning the export of powerful Nvidia GPUs have meanwhile had limited success. In our Global AI Index, Tortoise research on cited use of Nvidia chips for AI training in AI research papers shows that while researcher use of powerful H100 and A100 chips in China still lags behind the US, it is growing rapidly (a 95 per cent increase from 2022 to 2023). Chinese institutions are circumventing the export bans through third countries – the FT recently revealed that the cost of renting A100 chips is now lower in China than the US.

The obstacles to China’s AI development are political rather than technological. Since 2021, a regulatory crackdown on tech companies in China has spooked investors, with overseas funders largely pulling out. The Chinese tech scene has never fully recovered. Private investment in Chinese AI companies was down from a peak of $4.5 billion in 2021 Q2 to $1.5 billion in 2024 Q2, even as US investment hit a record of $26 billion in the same quarter.

Chinese regulators have also introduced some of the world’s most comprehensive rules specifically on generative AI, including requirements that LLMs “embody core socialist values”. The nature of generative AI, particularly in image generation, causes extra headaches. US AI companies have sometimes struggled to prevent image models from producing racist or sexual imagery. The list of problematic imagery in China’s censorship regime is a lot longer.