Can ‘token exports’ give China an edge in the AI era?

As Chinese artificial intelligence companies continue to push for the adoption of their flagship models overseas, they have taken on a new role as a key source of so-called “token exports” to the global market. Chinese AI models accounted for four of the top 10 models in terms of token consumption on popular AI model marketplace OpenRouter from March 18 to April 18, highlighting their growing visibility in global developer usage of AI systems. Domestic token demand has been surging at an exponential pace. Data from China’s National Data Administration showed that daily token consumption had exceeded 140 trillion by March 2026, a more than 1,000-fold increase from 100 billion in early 2024. The figure also rose over 40 per cent in the first three months of this year compared with levels at the end of 2025. The South China Morning Post examines the evolving compute ecosystem and how China’s token-driven AI services are expanding internationally, albeit with structural limitations. What does OpenRouter data reveal? OpenRouter, a US-based model aggregation platform, integrates more than 300 leading AI models and processes over 30 trillion tokens per month. It accounts for roughly 1 per cent of all trackable global token consumption, according to research by Frank Nagle of the MIT Initiative on the Digital Economy. The platform provides a unified application programming interface (API), allowing developers to switch between models and underlying compute providers in real time depending on cost, performance and use case. Through OpenRouter, global developers can access and deploy models across borders, with inference often routed to distributed cloud data centres. OpenRouter data also shows that the most token-intensive applications are increasingly driven by autonomous agent systems such as OpenClaw, alongside code-generation tools such as Claude Code used by individual developers. However, it should be noted that high rankings on OpenRouter do not necessarily translate into proportional revenue or global dominance, as the platform captures only a small fraction of global token usage and excludes large volumes generated through consumer applications and enterprise systems operated by firms such as OpenAI, Anthropic and Google. What are token exports and how do Chinese companies benefit? The term “token export” is increasingly used to describe cross-border AI inference, where tokenised requests are processed in geographically distributed compute environments rather than locally. Usage data from OpenRouter suggests that Chinese large language models have gained rapid traction among global developers. Before this year, the platform’s top three models by token consumption were US firms OpenAI, Anthropic and Google. Starting this year, however, models from Chinese companies including MiniMax, Moonshot AI, Zhipu AI, StepFun, DeepSeek and Xiaomi have intermittently ranked among the global top tier in token usage. Many Chinese models are also frequently released for free or with highly permissive pricing on OpenRouter, further boosting their token consumption share without necessarily translating into proportional revenue. China’s API pricing is significantly lower, typically one-fifth to one-twentieth of Western rivals, according to OpenRouter data. MiniMax M2.5 costs about US$0.3 per million tokens versus US$5 for Anthropic’s Claude Opus, while Alibaba Cloud’s Qwen 3.5 is priced at roughly US$0.11 per million tokens, around one-eighteenth of Google DeepMind’s Gemini. “Currently, domestic companies still primarily focus on cost-effectiveness, trying to make their prices more attractive,” said Duan Bing, an analyst from Nomura. “Their gap with the most advanced overseas models is also narrowing, so overseas users have real usage scenarios [for Chinese models],” he added. What are the structural limits of token exports? What is often described as “token exports” remains, in physical terms, a form of model distribution rather than true cross-border commoditisation of compute. Xu Shan, director of international development at the Institute of AI under the China Academy of Information and Communications Technology, noted in an article published on the website of the Ministry of Industry and Information Technology that OpenRouter did not fully reflect global token consumption patterns. Xu added that many Chinese models appearing on OpenRouter were deployed via overseas cloud infrastructure, meaning usage was often hosted outside China rather than processed in domestic data centres. China’s cost-based advantage remains structurally fragile. Most leading domestic AI firms reported 2025 revenues of between 100 million yuan (US$14.6 million) and 700 million yuan, while companies such as OpenAI and Anthropic have reached reported annualised revenues of over US$25 billion and US$30 billion respectively, highlighting a significant monetisation gap. While Chinese models are gaining visibility through open platforms and aggressive pricing, analysts said the gap between usage share and commercial scale suggested that China’s role in the emerging token economy remained primarily one of distribution rather than full-scale economic dominance. In other words, rising token consumption does not yet equate to control over the global AI value chain. Nomura’s Duan, meanwhile, expected token prices to pick up when domestic models penetrated into enterprise scenarios, leading to an increase in users’ willingness to pay. “In the future, it’s not necessarily true that large-scale token models in China will primarily be used by overseas users, domestic demand may also grow,” he said.

Can ‘token exports’ give China an edge in the AI era?

.css-1c6uqr6{color:inherit;font-weight:inherit;font-size:inherit;font-family:inherit;line-height:inherit;overflow-wrap:break-word;}As AI tokens emerge as a new commodity, China’s cost advantage is reshaping global model competition, though monetisation gaps remain

What does OpenRouter data reveal?

As AI tokens emerge as a new commodity, China’s cost advantage is reshaping global model competition, though monetisation gaps remain