The AI usage rate in China has exceeded 50%.

Data from US-based OpenRouter shows that as of June, the proportion of China-produced AI used reached 55% calculated by "Token", while the US-produced one accounted for 43%. Back in September 2025, the US share hit 76% and China's was only 20%...

Among services that support flexible selection of different artificial intelligence (AI) models, the usage share of Chinese AI has exceeded half. While improving performance, Chinese AI has adopted an "open source" strategy that facilitates low-cost access for external developers, thus expanding its market share. In the future, price competition in the high-end AI sector will become more intense.

The weekly data as of September 7 from US-based OpenRouter, which can distinguish the usage of different AI models, shows that the most widely used AI model is "Hy4 preview", a new type of AI launched by China's Tencent Holdings.

The agent of Hy4 preview has prominent advantages, allowing modification and commercial use. For example, game development companies can combine it with their own game engines, and use AI to realize character movement and dialogue in games.

Data from OpenRouter shows that as of June, calculated by "Token" that represents the amount of data processed, the proportion of China-produced AI used reached 55%, while the US-produced one accounted for 43%. Back in September 2025, the US share hit 76% and China's was only 20%.

One type of open-source AI called "Open Weight" only discloses the numerical data of learned "parameters (weights)" equivalent to the core content of the model. In addition, there is another type defined as "Open Source", which not only discloses parameters, but also widely releases technical parameters such as basic code and training data.

Strictly speaking, the open-source models of many Chinese and American enterprises should be classified as open weight. In addition to Hy4, the influence of Chinese enterprises such as Alibaba Group, DeepSeek and Moonshot AI is rising rapidly.

Models that do not disclose relevant technologies are called "closed source". Many core models of American technology enterprises such as Anthropic belong to the closed source category. Enterprise users cannot modify the AI itself, while development enterprises can specify the service provision method. This makes it easy to charge high usage fees to recover R&D investment.

Open-source models can be downloaded, used and further optimized by anyone. For China, which has long been restricted by the US government from obtaining cutting-edge semiconductors, it is easier to popularize AI even with limited computing resources it owns.

Driven by the Chinese government's policy of promoting AI applications that can be deployed in various industries and social fields, AI has been gradually applied in many sectors such as manufacturing and medical care.

Some American technology enterprises are also providing open-source models. Typical examples include "Gemma" from Google, "Phi" from Microsoft, and "gpt-oss" from OpenAI.

These are all small models, with lower performance compared to closed-source models. In the past, high-end models adopted closed source, while open-source models were developed and provided as an extension of R&D under the premise of controlling costs. However, with the continuous improvement of the performance of Chinese open-source models, there is a growing trend that American technology enterprises also begin to attach great importance to open source layout.

Among American enterprises, Meta is the one that actively promotes open source. In addition to its flagship open-source model "Llama", the company also launched "Muse Glimmer", an open-weight model for intelligent agents, in August.

US-based NVIDIA announced on September 3 that it will acquire US program sharing website Hugging Face for about 2 trillion yen. Hugging Face is an open-source development data aggregation platform. NVIDIA hopes to master this technology disclosure platform to attract users and convert them into customers of its semiconductor business.

US online media reported in July that against the background of the booming US open-source model market, the US government may restrict domestic enterprises from using open-source models developed by China. As the open-source models of Chinese enterprises have performance close to that of American tech companies and lower prices, the usage of such products by American enterprises is on the rise.

Many American technology enterprises are alert to excessive intervention from the US government. Satya Nadella, Chief Executive Officer (CEO) of Microsoft, stated in an open letter that "Open weights will bring competition to developers, semiconductors, applications and the entire service sector, and drive innovation."

This letter has been signed by more than 270 enterprises and organizations, which also reflects the widespread concern that a few enterprises such as Anthropic and OpenAI are monopolizing advanced AI technologies.

On the other hand, Anthropic, which takes the lead in the cutting-edge AI field, always adheres to the closed-source strategy. Dario Amodei, its CEO, claimed that Chinese enterprises are using Anthropic's AI without permission through "distillation", emphasizing the necessity of taking countermeasures.

Since open-source AI can be provided by a large number of enterprises, it is easily faced with downward price pressure. Recently, DeepSeek lowered the usage fee of some of its self-developed AI models. To compete with Chinese enterprises, OpenAI and Google have also cut the prices of their closed-source models.

Against the background of increasingly fierce price competition with Chinese AI, Anthropic's high-performance models may also be forced to reduce prices.

This article is from the WeChat Official Account "Nikkei Chinese Net" (ID: rijingzhongwenwang), written by Nikkei Chinese Net, and published with authorization by 36Kr.