Aerial cityscape of Changsha, China, showcasing vibrant hills and a sprawling urban skyline under a clear blue sky

China's data regulator plans standards for metering and billing AI tokens

By Himanshu Sakre·

Photo: CHINA YU / Pexels

National Data Administration chief Liu Liehong said on November 20, 2026 the agency will improve token metering, billing and quality standards. No dates or criteria were given.

China's National Data Administration says it will improve standards for measuring, billing and quality-checking AI tokens, the unit in which model use is now sold. The agency's director, Liu Liehong, said on November 20, 2026, at the World Computing Conference in Changsha, that "an industry covering the production, sale and use of tokens is taking shape." It is a short statement with no dates or numbers attached to the standards themselves, but it signals that Beijing wants to regulate AI compute as a metered commodity, and that matters for anyone tracking the physical race.

What Liu actually said

Liu Liehong, Party Secretary and director of the National Data Administration (NDA), spoke at the opening ceremony of the 2026 World Computing Conference. His remarks were posted on the NDA's website on November 21. Reading the original Chinese, he said tokens connect computing power, models and data into a value chain and serve as an anchor for the digital economy. Token activity, he said, is starting to take off, with an ecosystem around token production, circulation and consumption forming.

He described "token factories" emerging, which turn computing leasing, model supply and data capabilities into token production services, and said industry datasets, after model training and inference, become metered, deliverable intelligent services billed by token count and per call. On standards, the agency plans to improve token metering, billing and quality evaluation, guide orderly development and, in the NDA's wording, firmly prevent low-quality duplicate construction. The NDA page gives no token volumes, prices or performance metrics.

The usage numbers behind it

Numbers arrive through a secondary route. The Next Web, which reported the speech on November 21, says Liu said in March that China's daily token use had passed 140 trillion, up more than 1,000-fold from early 2024, and that by June daily use had passed 500 trillion, according to agency data cited by People's Daily. We could not open the People's Daily item, so treat the 500 trillion figure as reported secondhand. The March figure is attributed to Liu by the same outlet.

Daily token counts are an odd statistic. A token is a chunk of text a model reads or writes, and counts depend on each model's tokenizer, on how much of a prompt is cached and on whether agents loop. Without standards, 500 trillion at one provider and 500 trillion at another may not measure the same work. That is precisely the gap the NDA says it will fill, and why we read this as a measurement story first.

“An industry covering the production, sale and use of tokens is taking shape.”

Liu Liehong, director, National Data Administration, as reported by The Next Web, November 21, 2026

Why metering is a policy lever

Standards for billing do three things in practice. They make prices comparable across providers, which helps buyers and pressures margins. They let regulators count and tax compute-based services. And they create a definition of "quality" that can include safety, language and data-source requirements. The NDA mentions quality evaluation but gives no criteria, so we cannot say which of these it will emphasize. It did not say when standards would be published.

The phrase about preventing "low-quality duplicate construction" repeats a familiar Chinese industrial-policy worry: too many near-identical data centers and platforms. It reads as a signal that capacity will be steered, not just encouraged.

The compute context

China's constraint is chips, not appetite. Our explainer on whether China can train frontier AI without Nvidia covers the Huawei Ascend and memory bottlenecks. A separate commentary in Fortune on November 21 by Georgetown's Center for Security and Emerging Technology authors makes a related point from the other side: US-owned data centers with US chips rent capacity to inference providers that serve Chinese language models to global users. The authors note DeepSeek's latest model is offered by at least 16 providers outside China, including Amazon Bedrock, and that Together AI, Baseten and Fireworks serve Chinese models. They write that "Chinese models are running on U.S. chips." They also report that companies serving Moonshot's Kimi K3 with over $20 million in annual sales must negotiate a revenue-sharing license, and that Moonshot has reportedly asked Microsoft, Amazon and Google for 30% of their Kimi K3 revenue. That last claim is reported, not confirmed.

Put together, token standards in China and export-control debates in Washington are two sides of one question: who meters, prices and controls the inference that runs on whose silicon.

What we cannot say

We cannot say what the standards will contain, whether they will bind foreign providers, or whether they will favor domestic models. The NDA gave no timeline. We also cannot independently check the token volumes; they come from agency data as relayed by the press. And we have no comment from companies on how the standards would affect them.

What Western readers should take from it

For investors and builders outside China, the near-term effect is limited: no standard exists yet, and nothing in the statement binds foreign providers. The longer-term interest is comparability. If China defines a token billing unit before anyone else does, its definition may become the default in markets that buy Chinese models, a point worth watching given how many providers outside China already serve them, according to the CSET authors in Fortune.

Our take

Do not read this as a breakthrough. Read it as the state saying it intends to treat tokens like kilowatt-hours: something to be metered, standardized and counted. For the superintelligence race, the interesting number is not tokens per day but who gets to define the unit, because the definition will shape compute accounting, pricing and any future compute-based regulation. We would watch for the first draft standard, whether it names quality tests, and whether any Western regulator or standards body proposes a comparable definition.

Frequently asked questions

What are AI token standards?

Rules for how tokens, the chunks of text models read and write, are measured, billed and quality-checked. China's NDA said on November 20, 2026 it plans to improve such standards but gave no date or criteria.

Who announced China's AI token standards plan?

Liu Liehong, Party Secretary and director of the National Data Administration, at the opening ceremony of the 2026 World Computing Conference in Changsha.

How many tokens does China use a day?

The Next Web reports Liu said daily use passed 140 trillion in March and agency data cited by People's Daily put it above 500 trillion by June. We could not verify the People's Daily figure directly.

What is a token factory?

Liu's term for services that turn computing leasing, model supply and data capabilities into token production, with outputs billed by token count and per call.

Why does token metering matter for the AI race?

Whoever defines the unit shapes compute accounting, pricing and any compute-based regulation. The NDA statement is an intent to standardize, not a rule, and gives no criteria.

Sources

What each one is, and whose it is.

  1. 1

    Liu Liehong remarks at the 2026 World Computing Conference opening, National Data Administration (China) (November 21, 2026)

    OtherIndependent of the vendor
  2. Press reportIndependent of the vendor
  3. 3

    Data centers, U.S. chips and Chinese AI models, Fortune (CSET authors) (November 21, 2026)

    Press reportIndependent of the vendor