Chinese artificial intelligence (AI) models have overtaken U.S. models in token processing share on the world's largest AI model brokerage platform, OpenRouter. This shift in market dynamics suggests that the competition between the U.S. and China in AI is moving from model performance to actual usage.
According to an analysis of OpenRouter data by Bloomberg, as of the last week of June, Chinese models accounted for 48% of token processing share, while U.S. models held only 20%. This marks a complete reversal from a year ago, when U.S. models dominated with a 74% share compared to China's 20%.
OpenRouter allows developers worldwide to access over 400 AI models through a single API, with a weekly processing volume reaching 25 trillion tokens, making it a key indicator of model selection in real-world development.
The volume gap is significant, with Chinese models using approximately 18 trillion tokens weekly, far surpassing the 5.5 trillion tokens used by U.S. models.
In June's leaderboard, DeepSpeed's 'V4 Flash' ranked first in token usage, followed by Tencent's 'Hy3' in second place, with six of the top ten models being Chinese open-source models. OpenAI, which once dominated the market, has fallen to fourth place, significantly trailing behind DeepSpeed.
The driving force behind this shift is pricing. The API costs for Chinese models are 10 to 20 times cheaper than those of major U.S. models, with DeepSpeed's V4 Flash being up to 150 times less expensive than GPT-5.5 based on output tokens. As the share of coding and agent tasks in overall platform usage surged from 11% to over 50% since early last year, cost has become a primary criterion for model selection. According to venture capital firm Andreessen Horowitz, about 80% of startups using open-source AI stacks are running Chinese models.
The performance gap has also narrowed. In the Artificial Analysis Intelligence Index, China's top open-weight models, MoonshotAI's 'Kimi K2.6' and Xiaomi's 'MiMo-V2.5-Pro,' scored 54 points each, just six points behind the U.S. leader, GPT-5.5, which scored 60. This is a significant improvement from a year ago when the top Chinese model scored only 22 points.
In response to China's aggressive expansion, the U.S. has begun to take measures. In April, the U.S. House China Task Force and the Homeland Security Committee sent letters to Airbnb and AI coding tool developer Anisphere, expressing national security and data privacy concerns regarding the use of Chinese models. There are concerns that DeepSpeed's explicit use of data inputted into its API for training could be subject to China's National Intelligence Law, which may require companies to provide data related to their operations.
However, in the premium enterprise market, where safety and data protection are paramount, U.S. models still maintain an edge. Anthropic held a 14.8% share of weekly token usage on OpenRouter as of June, ranking second, and the revenue gap with Chinese models is much larger due to higher prices per token.
Korean models have virtually no presence in this competition. Upstage's 'Solar Pro 3' has been listed on OpenRouter since January, but its cumulative processing volume is only in the billions of tokens, with no Korean model ranking among the top users on a platform that processes 25 trillion tokens weekly. Companies like Naver and LG AI Research also have their own models, but their usage in the global developer market remains minimal.
As the government-led independent foundation model (Dopamo) project is still in progress, there is no clear understanding of the scale at which domestic developers and startups are using Chinese models through platforms like OpenRouter. Concerns are growing that domestic models are losing price competitiveness due to factors like electricity costs, and the AI Basic Law lacks regulations on the use of foreign model APIs, leaving little means to curb the penetration of Chinese models in the domestic market.
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.