The generative AI competition has entered its third year, with the leadership of the practical usage market firmly in the hands of Chinese open-source models. While the U.S. maintains an edge in top-performing models, it has struggled to find a counter to the low-cost onslaught, resulting in a significant loss of market share.
According to the IT industry on August 10, the platform OpenRouter, which aggregates various AI models through a single API, reported that as of August 5, eight of the top 10 models by weekly token usage were Chinese. Although models like ChatGPT, Gemini, and Claude are well-known globally, 60% of global token usage is concentrated on Chinese models.
The latest data shows that Chinese models occupy eight of the top 10 spots, with a market share of 88.4% among these models. The top model, DeepSeek's coding-specialized 'V4 Flash,' processed 6.92 trillion tokens weekly, followed by models from Xiaomi and Tencent. Only two U.S. models, OpenAI's 'GPT-5.6 Luna' and NVIDIA's 'NemoTron3 Ultra,' made the list.
In June of last year, OpenRouter's token market share was 74% for the U.S. and 20% for China, but by the last week of June this year, it had flipped to 48% for China and 20% for the U.S. In July, Chinese models claimed all top five positions, pushing China's share of overall routing traffic above 60%, with Chinese models at one point accounting for 63% of the tokens processed by U.S. companies.
The driving force behind this shift is a significant price disparity. The output token cost for DeepSeek's V4 Flash is $0.28 per million tokens, 99% lower than Anthropic's top model, 'Claude Opus 4.8,' which costs $25. ZhipuAI's open-source model 'GLM-5.2' saw its usage increase nearly 50-fold within a month on the development platform Versel, thanks to operational costs at just 20% of U.S. competitors while maintaining stable performance. AI agent startup Lindy reported saving millions by switching from Anthropic's model to DeepSeek's.
The low-cost strategy is supported by China's overwhelming domestic volume. According to the National Data Bureau of China, the daily AI token processing volume in the country has surged to 140 trillion, an increase of about 1,400 times since early 2024. This scale of economy achieved domestically enables the ultra-low-cost supply of open-source models for international markets.
In response to the low-cost offensive, the U.S. has few effective countermeasures. The startup Arci AI, which aims to compete with Chinese models through open-weight models, recently raised only $50 million in funding. Industry insiders explain that major venture capital firms are hesitant to invest, fearing that the success of open-weight models could undermine the valuations of their existing investments in OpenAI and Anthropic. This creates a structural dilemma where capital invested in closed models hinders the counteroffensive of the open-source camp.
Among single-frontier model factions, Anthropic is managing to hold its ground. Although its token processing volume lags behind China, its profitability metrics tell a different story. The average token cost for Claude Opus, according to OpenRouter, is 23 times higher than that of DeepSeek's V4 Flash, and under standard pricing, the gap can widen to 139 times. This indicates that despite losing processing volume, a significant portion of its revenue remains intact.
Leveraging its market dominance, DeepSeek is reportedly planning to go public in China this year. However, on August 6, it announced plans to significantly raise API prices, suggesting that its ultra-low-cost policy has become unsustainable due to the operational resource burden from a surge in users, and profitability has become urgent ahead of its IPO. Companies like Alibaba Cloud, Tencent Cloud, and ZhipuAI have also recently raised prices, indicating that the price-cutting competition from China may have reached its peak.
Domestic companies face a complex situation. SK Telecom announced on July 29 that its own model 'A.X K2' was designed to match the performance of Chinese models like Q1, DeepSeek, GLM, and Kimi, but its competitive pricing in practical use remains a key issue. Other major platforms, including Naver and Kakao, have yet to present objective superiority metrics compared to Chinese open-source models.
Meanwhile, the government is set to announce the results of the second selection for its independent AI foundation model project on August 12. Whether the current strategy of matching performance while not competing on price against China will be effective in the market will likely be assessed after the second selection results are released.
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.