China Wants to Standardize the Token, Like a Kilowatt-Hour
China just told the world how it thinks about AI: as a utility you meter. On Saturday the National Data Administration posted remarks by its director, Liu Liehong, from the opening of the 2026 World Computing Conference in Changsha on November 20. The line that matters is the last one. The agency will "improve standards for token metering and billing and quality evaluation," widen the use cases for token services, and "resolutely avoid low-quality duplicate construction." Original (Chinese): https://www.nda.gov.cn/sjj/jgsz/jld/llh/llhldhd/1121/20261121163827783032901_pc.html
Liu's framing is worth reading slowly. Tokens, in the official Chinese term "ciyuan," are called the "value anchor" of the digital economy, the link that carries value between compute, models and data. "Token factories" are named as a new form of industrial organization: rent the GPUs, supply the model, add the data, sell tokens as a service. Industry datasets get turned into intelligent services that are "metered in tokens and priced per call." That is a cloud price sheet rewritten as industrial policy.
The numbers behind the push are big. In March, Liu said China's daily token use had passed 140 trillion, more than 1,000x early 2024 (Yicai). By June it passed 500 trillion, according to agency data cited by People's Daily. Bloomberg broke the English version, TNW has the roundup: https://thenextweb.com/news/chinas-data-regulator-plans-standards-for-measuring-and-billing-ai-tokens . What is missing: any draft, any timeline, any definition. Nobody has said what a "standard token" is when every tokenizer splits text differently.
That gap is the interesting part. A token from a 5B model and a token from a frontier model are not the same product, the way a kilowatt-hour from any plant is. A billing standard has to pick: count raw tokens and reward bloated outputs, or price by quality and admit tokens are not a commodity. Add the "quality evaluation" clause and the state is effectively signing up to grade model output. Meanwhile Rhodium Group, per the New York Times, estimates Chinese big tech capex doubles to $139B this year with AI-lab revenue under 8% of that. The "avoid duplicate construction" line reads like a regulator who has seen that ratio too.
For anyone building agents, this is the real tell. Agents are the heaviest token consumers there are, and the country that standardizes the meter first gets to define what an agent's work costs. Watch for the draft: if it counts tokens instead of tasks, it will reward exactly the agent loops everyone is trying to make cheaper.
← Back to all articles
Liu's framing is worth reading slowly. Tokens, in the official Chinese term "ciyuan," are called the "value anchor" of the digital economy, the link that carries value between compute, models and data. "Token factories" are named as a new form of industrial organization: rent the GPUs, supply the model, add the data, sell tokens as a service. Industry datasets get turned into intelligent services that are "metered in tokens and priced per call." That is a cloud price sheet rewritten as industrial policy.
The numbers behind the push are big. In March, Liu said China's daily token use had passed 140 trillion, more than 1,000x early 2024 (Yicai). By June it passed 500 trillion, according to agency data cited by People's Daily. Bloomberg broke the English version, TNW has the roundup: https://thenextweb.com/news/chinas-data-regulator-plans-standards-for-measuring-and-billing-ai-tokens . What is missing: any draft, any timeline, any definition. Nobody has said what a "standard token" is when every tokenizer splits text differently.
That gap is the interesting part. A token from a 5B model and a token from a frontier model are not the same product, the way a kilowatt-hour from any plant is. A billing standard has to pick: count raw tokens and reward bloated outputs, or price by quality and admit tokens are not a commodity. Add the "quality evaluation" clause and the state is effectively signing up to grade model output. Meanwhile Rhodium Group, per the New York Times, estimates Chinese big tech capex doubles to $139B this year with AI-lab revenue under 8% of that. The "avoid duplicate construction" line reads like a regulator who has seen that ratio too.
For anyone building agents, this is the real tell. Agents are the heaviest token consumers there are, and the country that standardizes the meter first gets to define what an agent's work costs. Watch for the draft: if it counts tokens instead of tasks, it will reward exactly the agent loops everyone is trying to make cheaper.
Comments