AI starts billing per hour: 48 hours for tokens to be placed on Double 11 shelves

In 2026, Double 11 added a new product that has never been seen before on the shelves: Token.

The Zhitong Finance App learned that on the evening of November 2, Li Jiaqi's live broadcast room sold 100,000 copies, totaling about 2 trillion tokens; almost the same night, the JD Double 11 live broadcast room launched token packs for more than 10 big models, including JD Cloud, DeepSeek, Zhi Spectrum, and Volcano Seed. Sales increased 20 times month-on-month in 2 hours, and “Token” searches increased 7 times month-on-month. At the Taobao recharge center, tokens are on the same page as phone bills, utilities, and coal.

A unit of measurement that originally only belonged to the developer's back office broke into mass consumption shelves within 48 hours. It's not a simple marketing event; it's a sign that AI is moving from a “model race” to an “infrastructure race.”

A token is the unit of measurement for AI — every time the model generates a piece of text or image, it is measured as a token. In the past two years, it was just a cold figure for the API back-office of the big model manufacturer: Kimi entered 13 yuan and exported 68 yuan per million tokens; Meta charged 4 to 20 US dollars per million tokens; of Anthropic's revenue of US$16.286 billion in the first half of the year, consumer revenue billed by token accounted for 80%. Today, Token's “sales location” has changed—from the developer back office to the live streaming shopping cart. When a unit of measurement can be bought directly by ordinary consumers, it is no longer a technical indicator, but the smallest unit of a business.

Demand-side outbursts have supported this transformation. According to the caliber of the National Data Administration and CCID Research Institute, the average number of daily word calls across the country increased from 100 billion in early 2024 to over 175 trillion in June 2026. The “Ten-Year AI Outlook II - Personal Agent” released by CICC on November 4 estimates that under high-frequency usage, personal intelligence consumes about 360 million tokens per month, and the hardware cost is about 23 US dollars/month; by 2030, China's personal agent market is expected to exceed 510 billion yuan. According to the QuestMobile report on the same day, Doubao has been used 72 times per person per month. User habits have been developed, and AI's “electricity consumption” is taking off.

The commercialization of tokens benefits not only those who sell tokens. Along the “production-mass-production-transport-trade” chain, clear card holders have emerged.

The production side is the model manufacturer. Alibaba's (09988) Qianwen Office is the protagonist of this round of token commercialization. The 0.01 yuan monthly card essentially uses tokens as a customer acquisition tool. The goal is to keep users in the Qianwen model ecosystem. For large manufacturers, tokens are traffic entrances.

A special “etymology factory” has appeared on the mass production side. The Beijing No. 1 Jiyuan factory built by Softcom Power (301236.SZ) has been put into operation, with Nissan Token reaching 1.4 trillion dollars; Changshan Beiming (000158.SZ)'s Yuanyuan factory was online for more than ten days and connected to more than 100 customers. The word element factory packages models, computing power, and inference systems, and charges according to token output — tokens have changed from “model accessories” to products that can be mass-produced independently.

The delivery end is a cloud vendor. Established cloud service providers such as Jinshan Cloud (03896) are responsible for stable delivery of tokens to customers: the server stock exceeds 100,000 units, the EBITDA profit margin is about 32%, and the customer structure and delivery capacity have been verified. The higher the number of token calls, the more stable the flow of this “pipeline” will be.

The scheduling and transaction side is a layer where tokens are operated as “standard products”. Guangdong-Gangwan Intelligent Computing (01396) is stuck in these two areas: the self-developed “quantum power” computing power cloud platform supports unified scheduling of computing power across regions, chips, and cloud environments to solve the “production and transportation” of tokens; at the same time, as the operator of the Guangdong Word Exchange Center, it participates in token transactions and services to solve the “measurement and transaction” of tokens; and the 10-billion-level word element production base built in Hohhot uses green voltage to lower the production cost of tokens. On the operational side, the listing rate of its computing power resources has stabilized at over 95%. Of the new orders signed this year, more than 95% are about three to five years long, and over 95% have advance payments.

The four links put together are a complete token product chain: model factory pricing, word element factory mass production, cloud vendor transportation, and operator organization transactions. Li Jiaqi and JD's live broadcast room sold the front end; what caught up in the last few rings was the number of calls for each day after that.

Looking at these few things together, the logical chain is clear: tokens are placed on the shelf → call volume is activated → consumption continues to increase → value of all aspects of production, scheduling, and transaction is re-priced.

Within 48 hours, Li Jiaqi sold 2 trillion tokens, and JD gave out 1 billion tokens. These numbers may have the color of Double 11 marketing, but the changes they point to are real: when Tokens appear in the same recharge center as phone bills, water, electricity, and coal, the AI's “electricity” will already begin to be billed on a per cent basis. The token on the shelf is a coming-of-age ceremony for AI to move from the world of technology to infrastructure for people's livelihood. And all pay-per-use businesses ultimately have to answer the same questions — who produces, who transports, and who measures.