Beijing’s deadline for proposals on its AI token standards passed on 15 November. The call asked what those standards should cover. It did not set any rules.
That leaves a simple-sounding question open: when companies and governments in China talk about tokens, are they counting the same thing? For anyone buying AI services, the answer has commercial weight. A token count can help explain a bill. On its own, it cannot show how much useful work was done.
Beijing’s notice sought standardisation needs across four areas: token production, circulation, invocation and value conversion. They include cross-model conversion and consistency testing. It is a call for work to be proposed, not a finished rulebook. Beijing Municipal Administration of Government Services and Data, 20 October 2026
Nor is Beijing starting from scratch. Four national efforts are already visible, each at a different stage. The open question is less whether rules will be written than what they will make comparable, and what they will leave out.
A count is not a measure of value
Our explainer on AI token pricing showed how usage becomes a bill. The next question is what evidence should stand behind the quantities that appear in contracts, invoices and policy claims.
Three things often travel together under the word “token”:
- how a system counts its technical units;
- which usage becomes chargeable;
- what a reported total says about the result.
They are related, but one does not answer the others. A standard usage record could make an invoice easier to reconcile without making different models equally capable. Agreement on counting would not, by itself, turn a national usage total into a measure of productivity. A common label helps only if the boundaries behind it are clear.
Five efforts, five stages
Work on token rules is happening in several places at once. None of the five is an issued token-counting standard.
A definition in draft. On 7 April 2026 TC609, the national data standardisation committee, published a consultation draft of Data — Basic terminology that includes an entry for token. This is proposed wording at the consultation stage, not a final definition. A definition alone also does not settle how usage should be billed. TC609 notice and linked draft, 7 April 2026; draft, PDF page 8, clause 3.1.32
Counting rules, proposed. China’s national standards platform lists a project called Intelligent computing — Token service — Part 2: Counting rules, with the national intelligent-computing standardisation working group, SWG32, responsible. Its public-notice period ran from 30 September to 30 October 2026. The proposed scope covers input, output, cached, reasoning and multimodal tokens, as well as counting methods and output interfaces. It also envisages checks by users, third-party auditors and regulators. In other words, the aim goes beyond a single usage total: it is a record that another party can inspect. The page examined does not show that an effective standard has been issued, and the project’s approval stage beyond the published information still needs checking. National standards platform, public-notice dates 30 September–30 October 2026; accessed 15 November 2026
A metering framework, in drafting. TC28, the national information technology standardisation committee, oversees Artificial intelligence — Token metering framework and test methods, registration number 20264237-Z-469. Registered on 7 August 2026, it is a guiding technical-document project marked as being drafted. It is not an issued standard. Official project register, accessed 15 November 2026
A list of needs. In October TC609 published a needs list whose entries include multimodal token conversion, token measurement and service quality. A listed need is not evidence that a matching standard has been approved or brought into force. TC609 notice, published 15 October 2026; official attachment
Beijing’s local call, which closed on 15 November.
These five should not be read as one process. The evidence examined does not show a hierarchy among them, or that any one has completed work proposed by another. How they relate remains a question for further reporting.
Jiangsu’s vouchers follow spending
Jiangsu shows why the accounting matters, and why a policy’s details count for more than its label.
Its token-voucher scheme supports eligible purchases of intelligent-computing services, including rented computing capacity. The combined provincial and municipal contribution is in principle capped at 50 per cent of actual expenditure, up to 5 million yuan per applicant. Cities set the specific support rates, and budgets limit what is paid.
Applicants must provide contracts, invoices, proof of payment and proof of service use. Discounts, rebates and refunds do not count as eligible spending, and the same spending cannot be supported twice. The government is not paying a fixed amount for each token generated.
The notice sets out a support framework. It does not show how much has been paid out. Cities must report 2026 payments by the end of January 2027, which does not mean none could have been made earlier. No verified payment total was available for this analysis, and that is not proof that no payments have been made. Jiangsu Development and Reform Commission, document dated 30 September 2026, published 9 October 2026, attachment 1
Subsidising documented spending and measuring productive output are different jobs. Better usage records might help with the first without solving the second.

Large numbers need small print
National totals raise a similar problem. At a press conference on 24 March 2026, Liu Liehong, director of the National Data Administration, said daily token calls had passed 140 trillion by March, up from 100 trillion at the end of 2025 and 100 billion in early 2024. These are Liu’s figures; SouthwardTech did not conduct an independent national count. NDA press-conference record, 24 March 2026
A report published in October 2026 by the China Academy of Information and Communications Technology (CAICT) and its co-publishers puts daily domestic consumption in June at about 500 trillion. SouthwardTech read the preface and selected methodology pages of the 42-page report. The methods examined describe monitoring the performance of public-cloud APIs. They do not establish a comparable national series of token volumes. CAICT report, October 2026, PDF pages 3, 10–12 and 38–39; accessed 15 November 2026
Both figures have identifiable sources. That does not make them one verified series, because the material examined does not show that they share coverage or counting definitions.
Different reporting dates are not evidence of a contradiction, let alone misconduct. But without compatible definitions, a growth curve drawn between them would carry more authority than the evidence supports.
What buyers should ask
Beijing’s call does not set rules for international purchases. Overseas customers still have a stake, because these efforts concern the usage records behind the AI services they buy.
A useful procurement test is whether a supplier’s records explain four things:
- the service supplied;
- the usage counted;
- the amount charged;
- how a disputed total can be checked.
Whether the service did the task well needs a separate assessment. This is an editorial checklist, not a statement of existing legal obligations.
Researchers Mao Kejun and Liang Yongjian made a similar case in China Economic Times in June 2026, calling for disclosed counting rules, retained records and third-party checks. Their analysis was later hosted by the National Development and Reform Commission. It is an argument for accountability, not a binding billing regulation. Mao Kejun and Liang Yongjian, originally 3 June 2026; reposted 29 September 2026
Whether a usage record can be reconciled with a bill is a more useful test of standardisation than whether every token becomes interchangeable. A shared record could make an invoice, a subsidy claim or a market statistic easier to question. It would not make every model’s output equally useful.
