China Hits 500 Trillion Daily Tokens as Agent Workloads Rewrite the AI Race
As of June 2026, China's average daily token call volume has surpassed 500 trillion, placing Chinese large models firmly in the global first tier. Current flagship models are now updated almost monthly, with the competitive focus shifting toward agent deployment and ecosystem building, driving an explosion in inference compute demand. In the first week after the official release of Tencent Hunyuan 3, token call volume grew 68 times compared to the previous generation Hunyuan 2.
Signals of the Token Explosion
Tokens are the smallest units of information processed by large models. From an average daily volume of 100 billion tokens in early 2024 to 500 trillion by June 2026, a more than thousandfold increase over two years reflects the industry's transition from the parameter arms race of the 'hundred-model war' into the deep waters of large-scale application. Global comparisons are even more telling: OpenRouter data shows that during the week of June 15–21, 2026, the weekly call volume of Chinese AI large models reached 18.81 trillion tokens, holding the global top spot for eight consecutive weeks, while the US recorded only 5.76 trillion in the same period. Domestic large models occupy the top four global positions, breaking a long-standing pattern of overseas dominance.
China's AI call volume leads globally for eight consecutive weeks
Accelerating Iteration and Shifting Competition
Feng Wen, chief architect at Xiyu Technology, noted that in the second half of 2025, vendors planned to iterate every three months, but now new models are released every four to six weeks. This drastically compressed iteration speed means the competitive focus has shifted from 'model IQ' to agent deployment and ecosystem building. The large-scale deployment of agents is the core driver of this explosion—a single task requires repeated retrieval, context reading, tool invocation, and multi-round feedback, causing inference compute to grow exponentially. Data disclosed by Liu Feng, general manager of Tencent's Smart Industry division, confirms this: Hunyuan 3's first-week token call volume grew 68 times over Hunyuan 2, a direct manifestation of surging application-side demand.
Agent deployment drives inference demand
Cost Advantage and Ecosystem Building
Liu Liehong, director of the National Data Administration, officially translated 'Token' as '词元' (word unit) for the first time at the government level, pointing out that it provides a quantifiable basis for business model implementation. The price of domestic tokens is only a fraction of comparable overseas products, a cost advantage that makes Chinese AI applications more economical on the enterprise side. Policy support is equally critical. The 'AI+' initiative is advancing in depth, data factor market-oriented allocation reforms are landing, the 'East Data West Computing' project continues to operate, and the government work report for the first time listed 'compute-electricity coordination' as a new infrastructure project—together forming the infrastructure guarantee for massive token calls. By the end of 2025, China had built over 100,000 high-quality datasets, totaling 890 PB. When models move from 'able to chat' to 'able to work,' the true value of the token economy begins to materialize. Chinese large models have established a global leading position with 500 trillion tokens, but the next stage of competition will revolve around whether agents can truly embed into production workflows and whether ecosystem moats can continue to deepen.
References
[1] 我国日均词元调用量已突破 500 万亿，中国大模型稳居全球第一梯队 -36 氪。https://www.36kr.com/newsflashes/3957440108871042 [2] 日均词元调用量已突破 500 万亿，中国大模型稳居全球第一梯队。https://finance.sina.com.cn/jjxw/2026-08-27/doc-iniptxkv5790290.shtml [3] 我国日均词元调用量已突破 500 万亿 - 界面新闻。https://www.jiemian.com/article/15008302.html [4] 我国日均词元调用量已突破 500 万亿 - IT 之家。https://www.ithome.com/0/995/136.htm [5] 我国日均词元调用量突破 500 万亿，中国大模型稳居全球第一梯队 - 同花顺。https://m.10jqka.com.cn/20260827/c679345743.shtml