---
title: "Tokens Get a Standard Unit. AI Services Get a Yardstick."
date: 2026-11-22
category: Policy & Governance
site: NeuroAI
canonical: https://neuroai.site/a/na-policy-token-metering-standards-1122
language: en
---

# Tokens Get a Standard Unit. AI Services Get a Yardstick.

> China's Cyberspace Administration counted 133 new generative AI services that completed filing in September and October. Three of them provide generative AI on the phone itself.

China's Cyberspace Administration counted 133 new generative AI services that completed filing in September and October. Three of them provide generative AI on the phone itself. One provides anthropomorphic interactive service. One existing filing changed its information.

That is 133 services in 61 days, or about 2.18 a day. The daily figure is our division, not an official statistic.

## Three words, three jobs

Separately, China's National Data Administration said it would improve token metering, billing and quality-assessment standards, and expand the application scenarios for token services.

Metering answers how much. Billing answers how much is charged. Quality assessment answers whether the output was any good.

The first two sit on one chain: a common quantity has to exist before a common price can. The third runs on a different axis. It measures the other side of the same output, whether a given number of tokens produced something usable.

## Metering is not billing

The two words travel together so often that they read as one thing. They are not.

Metering is counting. Billing is pricing. A bill is a unit price multiplied by a quantity, and the quantity has to be counted before the multiplication means anything. Three quotes, one priced per token, one per character, one per minute, cannot be placed in one column until the units match.

The order in the phrase puts metering first. That order describes a dependency, not a schedule.

## What a token has to settle

Text enters a model as tokens. The model counts tokens, the service bills by count, and the customer pays for the count.

But the count depends on the splitter. The same paragraph, cut two ways, yields two token counts. At the same unit price, it yields two bills.

A metering standard faces three questions before the unit means anything: who splits the text, on what rule, and which end of the request is counted. Without those answers, a token is each vendor's own token.

## Two actions in one sentence

The data administration's sentence has a front half and a back half. The front half is about where token services get used. The back half is about how they are counted, charged and assessed.

These point at different work. One is adoption, the other is specification. They sit in the same sentence but do not move on the same timetable.

The sentence also names three standards and then says "and other standards," which leaves the list open. What else belongs in it is not written, and we will not fill the space.

## Three of 133

The three on-device services are about 2.3% of the 133. That ratio belongs to this two-month window. It is not a share of the year and not a share of the market.

Read the arithmetic carefully. The three are inside the 133, not added to it. The one anthropomorphic service is also inside. 133 plus 3 plus 1 is not 137, because 3 and 1 are labels on members of the set.

The anthropomorphic service is about 0.75% of the total. Whether one service can carry both labels is not stated, and we do not merge them.

## A change is not an arrival

One filing changed its information. That is not a new service. It is an existing registration whose details moved.

So 133 and 1 describe different events: 133 services entered, and 1 had its paperwork amended. Adding them would produce a service that does not exist.

Filing is a registration procedure, and the published word is that it was completed. Materials went in, information was recorded, the process closed. The notice contains counts and categories, nothing more.

## The scale is not in the text

There is no standard text yet. No definition of a token, no rule for counting input against output, no decision on where tokens generated on a phone are counted.

So the measurable part today is narrow: one sentence from the data administration, three named standards, and one two-month filing count with two sub-labels.

Two months can also be read as 66.5 services a month. The total does not change when the denominator does, but the rhythm does. A monthly rate and a daily rate describe the same 133, and neither one extends beyond the window.

## The unit is named, not defined

For anyone who buys AI by the token, the practical fact is this. The unit has entered a standards process, and the process has named the three things it covers. It has not published a scale.

Until it does, two vendors' token counts remain two vendors' counts. A buyer comparing quotes is still converting between private units, and every conversion is an assumption.

That is the state of the yardstick. The word is on the ruler. The marks are not.

*Disclaimer: This article is based on publicly available information and is for reference only. It does not constitute investment advice.*

---

Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-policy-token-metering-standards-1122
Free to quote with attribution and a link to the original.
