Small lexemes drive the large market of the intelligent economy (Big Data Observation · New Forms of the Intelligent Economy)

（Original title: Our country’s average daily token usage surpasses 140 trillion; companies’ willingness to pay to improve efficiency is strong—small tokens unlock a huge intelligent economy market (Big Data Observation · New Forms of the Intelligent Economy)）

Headline Highlights

A token is the smallest information unit used by large models to process data. Our country’s average daily token usage has grown significantly, reflecting not only the surging vitality of the intelligent economy, but also that a new set of business logic is accelerating its evolution. In the intelligent age, with tokens that can be measured, priced, and traded, artificial intelligence is expected to become a basic resource that supports society’s operations like water and electricity.

On the user side: looking up information, writing code, rewriting copy; on the enterprise side: financial risk control, intelligent customer service, code correction… Today, every scenario where AI applications are rolled out depends on massive calls to tokens.

A token is the smallest information unit processed by a large model. In March this year, China’s average daily token usage exceeded 140 trillion, increasing more than a thousandfold over two years. How should we understand tokens? What is the relationship between tokens and the AI industry, and what are the future development trends? Reporters conducted interviews.

Explosive growth in token usage is driven by improvements in model capabilities

Some people are puzzled: what is different between tokens and commonly used data?

“Under the hood, an AI large model is a complex mathematical computation system. It cannot directly read text, understand sounds, or see images the way humans do.” Tang Huabin, deputy director of the Network and IT Technology Research Institute at China Mobile Research Institute, said that AI large models must first transform various raw data into a “language” they can recognize, and then use computations to produce results. Tokens are the most fundamental unit in that “language.”

More specifically: text tokens are like “LEGO bricks”—a single word or a single Chinese character is broken down into blocks of tokens; audio tokens are like “notes on a musical score”—a sound segment is cut into extremely short time slices, with each slice containing pitch and volume; video tokens are like “puzzle pieces”—each frame of an image is cut into small squares, and continuity over time must also be considered…

“From the perspective of a large model, tokens across these three modalities are completely different. In the model’s eyes, it doesn’t feel like it’s reading text or watching videos—it only feels like it’s processing a sequence of extremely complex numbers.” Wu Di, head of intelligent algorithms at Volcano Engine, said.

Countless basic tokens make up the “cells” that drive the intelligent economy to run. In March this year, the weekly token usage of AI large models in China consecutively ranked first for three weeks, making China one of the countries with the highest global activity for large model applications.

“Explosive growth in token usage is most directly caused by improvements in model capabilities. Each time model capabilities advance, more application scenarios get unlocked, which further drives an increase in token usage for large models.” Wu Di explained. Taking Seedance (Doubao video generation model) 2.0 as an example, generating a one-minute video consumes roughly more than 1 million tokens.

At the same time, new application forms and new business models will also drive a large jump in token usage. Tang Huabin said that recently emerging “agents” are different from traditional single-round Q&A. Their operation often entails longer context windows, more frequent model calls, more rounds of task decomposition, and continuous feedback during tool execution.

“For simple tasks, if you can finish it with only a single round or a few rounds of tool calls, token consumption may only be a few thousand. If you need dozens of rounds of tool calls, you may need at least tens of thousands of tokens or even more.” Wu Di said. The emergence of new industries such as agents has once again put token usage on a fast-growth track.

Token usage concentrates in areas with high information density and fast iteration cycles

6 billion is the token usage amount for Li Jiayi, founder of Aiwen Da Technology, over the past year.

Entering the AI Origin Community located in Haidian District, Beijing, in a studio of about 16 square meters, Li Jiayi’s team is tuning an AI toy they just finished designing.

“This toy needs to have interactive capabilities, which is inseparable from the corresponding software system. In the past, developing software systems of the same scale would take at least half a year and would also require collaboration among multiple people.” Li Jiayi said. With deep enabling from AI large models and assisted programming, consuming hundreds of millions of tokens, the development cycle has been compressed to two months.

At the beginning of 2025, after Li Jiayi—who is not from a computer science major—came into contact with AI-assisted programming tools, he used new technology to break through the professional barriers of software development. “In one year, we efficiently completed the development and design of two APPs and one AI toy. Not long ago, we also brought our self-developed AI hardware product to the Consumer Electronics Show (CES).” Li Jiayi said. AI assistance not only greatly reduces development costs, but also creates more possibilities for innovation exploration by small and micro enterprises.

Software development is one of the typical scenarios for AI applications to be deployed. Overall, the distribution of token usage shows clear industry- and scenario-specific characteristics. It mainly concentrates in areas with higher information density, faster product iteration cycles, and closer links between models and production systems.

Taking Doubao large model 2.0 as an example, Wu Di explained that in terms of industry classification, the internet industry has the highest share of token usage. Consumer electronics, finance, new retail, and business services follow closely behind.

In terms of scenarios, the largest share is for the processing and analysis of unstructured information; education, content creation, and “search and recommendation” come next.

Looking ahead at development trends, agent application scenarios such as software development, deep research, and personal assistants may become important areas for explosive growth in token usage. “Especially software development: currently, AI is shifting from simply writing code to understanding an entire project. It can not only detect errors and omissions and automatically optimize, but can even complete the entire development task autonomously through agents. With long code context and many rounds of interaction, enterprises have a strong willingness to pay to improve efficiency. This will become a huge growth point for demand,” said Tang Huabin.

Tokens are measurable; artificial intelligence is expected to become a basic resource for the functioning of society

“Since the end of January this year, some model companies have set performance records where their revenue in 20 days surpassed their total revenue for all of 2025. Behind these figures is a new business logic based on token-based billing that is accelerating its evolution.” Liu Liehong, director of the National Data Administration, said.

Looking back at the development course of the AI industry, in a past period of time, the sector focused on competing model performance. There was a lack of a quantifiable bridge between technological innovation and business deployment, making it difficult to form a virtuous cycle of “technological iteration—value output—continued investment.”

“Tokens themselves have characteristics of being measurable, anchoring compute power and energy consumption, and supporting cross-modal general settlement, enabling them to become the unit for settling technology supply and business demand. To put it simply: the intelligence produced by a large model is like electricity, the intelligence computing center is like a power plant. Electricity is measured in kilowatt-hours, and intelligence calls are billed using tokens.” Huang Shan, strategic management director for Lenovo China’s Infrastructure Business Group, said. Looking to the future intelligent society, artificial intelligence is expected to become a basic resource that runs like water and electricity—available on demand, paying for exactly what you use.

From the perspective of tokens, to build a new form of the intelligent economy, what advantages does our country have? What are the key areas to focus on for development?

Algorithmic innovation continues to deliver breakthroughs. Domestic large models gradually narrow the gap with global top technologies through optimization of underlying architectures. Wu Di said that Chinese large-model vendors have made a great deal of algorithmic innovation, continuously optimizing inference costs and response speeds, and can complete complex tasks with fewer tokens.

Infrastructure has leading advantages. For every token generated, it calls compute power from data centers, accompanied by power consumption. Our country has already built a global energy system with the most complete categories and the largest scale. With ample power supply, strong power grids, and a viable market, Tang Huabin said that with continuously improving electricity generation supply and steadily decreasing electricity costs, our country can effectively reduce the cost of token usage.

“Energy consumption and compute power are two key cost components of token calls. Producing each token with less compute power and less energy consumption reflects the production capability and efficiency of basic infrastructure such as intelligence computing centers.” Huang Shan said. To build secure, efficient, and inclusive AI infrastructure, it is necessary to continuously improve large-model inference efficiency and reduce the cost per token, so that AI can move toward large-scale deployment and application to the greatest extent possible.