The Rise of Chinese AI: How China is Outpacing the US in Large Language Models

Technology specialist Yin Ruizhi takes a closer look at why China has been able to lower the cost of its large language models compared with the US.

Making Profits at Unbeatable Prices

In early May, an advanced LLM called DeepSeek-V2 emerged in China, achieving excellent performance and nearing the forefront of international standards in core AI technologies. What shocked the industry was the model’s cost-effectiveness, requiring only 1 RMB (US$0.14) per million input tokens and 2 RMB per million output tokens. This is astonishing compared with GPT-4, which costs 100 times more. Even the budget Chinese AI model Moonshot AI costs over 20 times more.

“The cost to process one million tokens is just 0.50 RMB — the rest is profit.” — DeepSeek’s team

With such a huge price difference, a large number of Chinese AI product suppliers are turning to DeepSeek. Many media platforms initially speculated that DeepSeek might be adopting a subsidised strategy to compete for market share. But DeepSeek’s team clarified that the company is still able to make a profit at such a price point, and the cost to process one million tokens is just 0.50 RMB — the rest is profit. Even more shocking is the fact that DeepSeek is an open-source coding model — anyone can build their own model based on it.

The Price War Intensifies

On 21 May, Chinese internet giants entered the price war. The first to be affected was Alibaba Cloud, which announced a significant price cut on its LLMs. This move immediately triggered a positive response from several well-known enterprises in the industry, provoking a subsequent wave of price cuts.

(_download_image) Price cuts sparking a wave of industry-wide price reductions

However, amid this intense price war, Alibaba Cloud’s price reduction strategy is starkly different from other vendors’ focus on adjusting the price of small language models. The former lowered the price of nine of its LLMs in one fell swoop, with its flagship model Qwen-Long — comparable to the internationally renowned GPT-4 — exemplifying the enterprise’s boldness. The application programming interface (API) input price of Qwen-Long has been reduced by a whopping 97%, from 0.02 RMB to 0.0005 RMB per 1,000 tokens.

(_download_image) Alibaba Cloud’s flagship model Qwen-Long, comparable to GPT-4

With this price reduction, 2 million tokens can be bought with just 1 RMB, equivalent to the amount of text in five copies of the Xinhua Dictionary. This price is around the cost of DeepSeek that was announced in early May.

The Advantage of Chinese Engineers

How are Chinese AI firms able to create a product similar to one offered by their American counterparts at a tenth of the cost, and how did this price advantage come about?

The key lies in the large number of engineers in China, and the unique circumstances for the industry. China’s top software engineers are all focused on price optimisation of its AI LLMs, and China pumps in nearly ten times the number of talents for this purpose as compared with the US.

(_download_image) China’s top software engineers focused on price optimisation of AI LLMs

In sum, for the few major AI platforms in China, they have assembled around ten times the talent compared with the US in the same field to work on cost-effectiveness — the efforts of Chinese software engineers have greatly reduced the cost for AI LLMs.

The significant cost difference between Chinese and US LLMs