China is pioneering a new way to export its AI capabilities, turning raw computing power directly into a digital service for the global market.
This emerging model is called 'token exports.' It involves selling AI model processing power, known as inference, to overseas users through an API from data centers located in mainland China. Instead of exporting physical chips, China is exporting the results of their computation. A pilot program in Shantou, Guangdong, has already validated this approach, with daily call volumes surging from 1 billion to over 10 billion tokens in just one month.
This strategic shift is driven by a confluence of three key factors. First, escalating U.S. tech controls are a primary catalyst. Washington's tightening restrictions on advanced AI chips, including a recent rule targeting the overseas subsidiaries of Chinese firms, have made acquiring high-end hardware increasingly difficult. In response, China is pivoting from a model that moves restricted chips to China to one that exports finished AI services from China.
Second, China has been methodically building the necessary domestic infrastructure. The country is weaving a nationwide compute network, developing its own powerful accelerators like Huawei's Ascend series, and leveraging low-cost green energy to power its data centers. Provinces like Guangdong also benefit from superior undersea cable connectivity to Southeast Asia, making it a natural and low-latency target market for these new API exports.
Finally, a supportive legal framework for data governance is solidifying. Beijing has been running pilots for compliant cross-border data processing, such as the 'lai-shu jiagong' (offshore data processing) program in Shantou. These initiatives, along with maturing data-export assessment procedures, provide the legal scaffolding for companies to confidently handle foreign data, process it within China, and return the results abroad.
In essence, this isn't just a clever workaround to sanctions. It's a fundamental pivot in China's digital trade strategy, combining its strengths in infrastructure and policy to shift from exporting physical hardware to selling high-value, scalable AI intelligence as a service.
- AI Inference: The process of using a trained AI model to make predictions or generate outputs based on new, unseen data. It's the 'live' operational use of an AI model.
- Token: In the context of large language models, a token is a unit of text, which can be a word, part of a word, or a punctuation mark. AI services are often priced based on the number of tokens processed.
- lai-shu jiagong: A Chinese term for 'offshore data processing.' It refers to a pilot program where data from overseas is legally brought into China for processing and the results are then exported.