Token Economy: How Tokens Are Transforming the AI Economy

The Token: The New Unit of the AI Economy

Just as electricity fueled the industrial era and data drove the digital revolution, the token is becoming the new benchmark unit of the artificial intelligence economy.

From computing power to autonomous agents, a complete value chain is taking shape around the production, consumption, and billing of tokens. This emerging economy, which depends on the amount of artificial intelligence consumed, is currently being accelerated by the development of AI agents. While assistants built into smartphones generally perform predefined tasks, more open agents allow users to customize processes, chain together actions, and interact with multiple services—all of which drive up token consumption. An in-depth look with Lucienne Liu, Orange’s China Ecosystem Manager, and Paul Delbouve, an economist specializing in AI.

From Language to Value: What Is a Token?

To understand the token economy, we must first revisit how large language models (LLMs) work. These systems do not process words directly as we read them. Instead, they break down information into basic units called tokens. Depending on the model used, a token can correspond to a word, part of a word, a punctuation mark, or a sequence of characters.

This segmentation, known as tokenization, transforms language into numerical representations that the models can process. The work by Sennrich, Haddow, and Birch, published in 2016, notably demonstrated the value of segmenting text into sub-words for natural language processing.

Thus, every interaction with a generative AI involves tokens: those in the query sent to the model (input) and those in its response (output). In multimodal systems, representation and counting mechanisms also enable the processing of images, sounds, or videos, with modalities that vary depending on the architecture.

The major shift occurs when this technical unit becomes a billing unit. As a result, model providers can market their AI capabilities based on the volume of tokens processed, typically with separate rates for input and output. The token thus takes on a dual role: it represents a unit of information processing, but also a unit for measuring and billing a portion of the models’ activity.

For Lucienne Liu, this development is significant because “unlike data, whose value depends heavily on its context of use, the token provides an accounting unit that links several stages in the AI industry value chain”. However, it is neither a universal currency nor a direct measure of the intelligence produced. Its economic value depends on the model, the task performed, and the conditions of use.

How does the token economy work?

The token economy, or AI token economy, refers to the economic ecosystem in which the production and use of artificial intelligence capabilities are measured, priced, distributed, or optimized using tokens. Its operation is based on a value chain that extends from physical infrastructure to end-user applications.

Upstream are the electricity, data centers, semiconductors, and computing resources needed to train and run models, followed by model providers, cloud platforms, and intermediaries that provide access to AI capabilities, particularly through application programming interfaces (APIs).

Downstream, companies and developers integrate these capabilities into software, assistants, and AI agents designed for users.

Each stage involves costs and opportunities for value creation. The price paid for one million tokens may thus include, directly or indirectly, expenses related to infrastructure, energy, development, and operations, as well as the margins of the various providers.

This approach is gradually transforming AI into an on-demand service, comparable in its economic principle to cloud computing, with pricing that depends on token volumes, user needs, and the degree of service customization, as well as plans combining subscriptions, quotas, and variable billing.

However, contrary to what one might think, not all tokens are created equal. Paul Delbouve explains: “Their cost depends on the sophistication of the model, the type of processing, and the desired level of performance.”

Four Business Models for Monetizing Tokens

By observing the Chinese ecosystem, Lucienne Liu identifies four major commercialization approaches that could be applied in France and Europe.

1. Pay-per-use token sales

Model providers and cloud platforms sell processing capacity accessible via API. Companies pay based on their usage, sometimes with volume discounts, capacity commitments, or throughput guarantees.

This model ties expenses directly to business activity but can make forecasting difficult when usage increases rapidly.

2. Tokens included in a subscription

In this model, token consumption is invisible to the end user. It is factored into the price of software or a service: text generation, image editing, translation, or professional assistance.

The company sells a feature rather than a quantity of computing power. It must nevertheless control its internal costs to maintain profitability.

3. Pay-per-result billing

With AI agents, a new approach is emerging: paying for a completed task rather than a volume of tokens consumed.

An agent capable of searching for a product, comparing offers, or making a reservation can perform several successive operations. The customer then purchases a service or a result, while tokens remain a component of the production cost.

This model, however, raises the question of how to define the expected result and who is liable in the event of failure.

4. Tokens as Service Infrastructure

Finally, some platforms offer orchestration tools that allow users to select different models, distribute requests, and optimize consumption.

Value no longer rests solely on the provision of tokens, but also on the ability to use the model best suited to each need.

What Are the Challenges for Businesses and Telecom Operators?

For businesses, one of the primary challenges is controlling the actual cost of AI. An application used on a large scale can generate millions of requests, with computing costs rising rapidly.

Optimization therefore involves several strategies: choosing models appropriate for the tasks at hand, limiting unnecessary processing, reusing certain results, or routing requests to different infrastructures. Telecom operators have specific advantages in this context. As Lucienne Liu explains, they can, in particular, “leverage their networks, their edge computing infrastructure, their distribution capabilities, and their knowledge of regulatory requirements.”

Edge computing is particularly valuable for certain use cases requiring low latency or local processing of sensitive information.

Operators could also play an orchestrating role by facilitating access to different models and ensuring a quality of service tailored to specific use cases.

These prospects do not mean that tokens will replace data plans, but rather that complementary offerings—at the intersection of connectivity, the cloud, and artificial intelligence—could emerge.

In China, this convergence between telecommunications and artificial intelligence has begun to take shape. In May 2026, several major operators launched commercial offerings that included AI token quotas, sometimes bundled with connectivity, cloud, or cybersecurity services.

Toward an economy of useful intelligence?

While not yet an economic indicator comparable to electricity consumption or data traffic, token consumption provides insight into the intensity of artificial intelligence usage within a company, sector, or region.

As a result, token-based billing could become standardized in the future for transactions between AI providers and integrators, while end users would largely continue to purchase subscriptions or features.

The token economy would therefore not necessarily constitute a self-contained economy, but rather a measurement and billing layer essential to the widespread adoption of artificial intelligence.