Large language model development in China thrives, but geopolitics may spell trouble

China’s large language models (LLMs) for generative artificial intelligence were off to a late start catching up to US-based Open AI’s ChatGPT, due to heavy state-imposed regulatory requirements. The first batch of eight approved LLMs were only released to the public in August 2023, led by Baidu’s Ernie Bot.

Since then, however, China’s generative AI landscape has become crowded with new entrants. The list of approved LLMs has grown to 117. IT giant Alibaba alone has invested billions of dollars in at least five of the leading generative AI startups, in addition to developing its own open-source LLM, Qwen. Tencent, ByteDance and Huawei are among other large tech firms working on their own generative AI offerings.

Despite strict regulations on training data and output, many Chinese LLMs are competitive in performance on global leaderboards. They perform especially well in Chinese-language tasks.

But great power competition with the US may yet throw a wrench in China’s progress. Many Chinese models rely on open-source development frameworks as the basis for their design, especially on Meta’s Llama series. The US Commerce Department is mulling restrictions on open-source models, although this may be difficult to implement. Chinese companies would find it much harder to develop LLMs using only Chinese technology at this point. Export restrictions from the US on advanced semiconductors will also continue to negatively affect innovation at the highest level.

However, geopolitics can be a boon to Chinese platforms: OpenAI has blocked programmatic access to their services by Chinese users, and other US companies might follow with similar steps in response to security concerns in Washington. Some Chinese companies have since rushed to offer discounts to developers to switch to their services. Restrictions on tech exports will contribute to a growing split in the tech ecosystem inside and outside China.

MERICS Analyst Wendy Chang: “China was off to a slow start in building large language models but has made much progress since. What happens next is as much about tech as about geopolitics: the US could hobble Chinese developments if it manages to cut off hardware and software access. China’s own progress in building AI chips and indigenizing software architecture will also be key to its staying competitive in the AI race.”