Source: Tencent Technology
Author: Shi Zhanzhong
In the rapidly evolving landscape of the artificial intelligence industry today, a term once only popular within technical circles—”Token”—is quietly becoming a key entry point for understanding the economic form of AI.
According to the latest data from OpenRouter, the world’s largest AI model API aggregation platform, the total Token usage of global large AI models from March 16 to 22 was 20.4 trillion, of which China alone accounted for 7.359 trillion, representing 36% of the global total.
At the 2026 annual meeting of the ‘China Development Forum,’ Liu Liehong, Director of the National Data Administration, stated that by March this year, China’s daily average Token consumption had exceeded 140 trillion, compared to 100 billion at the beginning of 2024, marking a growth of over a thousand times in two years. In the same month, China’s weekly Token usage of large AI models surpassed that of the United States for three consecutive weeks, making it one of the countries with the highest AI application activity globally.
Behind these numbers lies a usage scenario that is quietly changing—AI usage today cannot be compared to what it was before.
In the past, AI was primarily dialogue-based, providing answers to questions through input and output; but now, what truly operates on the front lines are entities known as Agents, which do not merely answer questions but also perform automated tasks on behalf of humans.
On social media platforms, “Q&M Dental raising shrimp” has become a new trend; some individuals have shared their monthly Token consumption bills on WeChat Moments. Enterprises are also accelerating adoption, with various manufacturers launching “shrimp-like” products and encouraging employees to “delegate tasks to AI,” integrating AI into daily workflows. Some companies have even included “how much work Agents have done for you” in work logs as a quantifiable metric.
Behind this AI wave sweeping across society, Token plays a key role. Whether it is Agent invocation, video generation, or AI usage in enterprise collaboration, what is ultimately consumed are Tokens counted in millions, tens of millions, or even trillions.
Describing Tokens as the “computational currency of the AI era” has become a popular analogy in the industry. However, if we delve deeper into whether this analogy truly holds and what economic order Tokens are reshaping, the answer is far more complex than a simple slogan.
In fact, Tokens are neither traditional currencies nor conventional computing resources but rather a new type of value carrier that combines characteristics of infrastructure, traffic commodities, intangible assets, and intellectual property rights. The pricing logic, market structure, and governance issues surrounding them may signify the beginning of a new economic paradigm.
This article attempts to conduct a systematic review of Tokens from perspectives such as monetary theory, industrial economics, market structure, and public policy. It will be divided into two parts, with this part serving as the first installment, focusing on the essence and definition of Tokens, cost and pricing on the supply side, and consumption logic on the demand side.
1. Is Token the ‘Currency of Computational Power in the AI Era’?
01 It has some similar attributes but is not equivalent.
In the industry, some people refer to Token as the ‘Currency of Computational Power in the AI Era.’ This analogy partially holds true in economic terms but lacks rigor. In fact, Token is not a currency; it is more akin to kilowatt-hours in the electricity world. Kilowatt-hours are not currency, but they measure productivity in the electrification era.
Token indeed possesses two basic attributes of currency. It serves as a unit of measurement for tracking the consumption of AI services and acts as a medium of exchange, converting user payments into usable intelligent outputs. However, Token lacks several key characteristics of currency: universal acceptance, general equivalence, and free circulation.
An OpenAI Token cannot be used on Anthropic’s platform, nor is it equivalent in cost or value to a Token from an open-source small model. It is more like an arcade token or prepaid card—restricted to specific scenarios, non-interchangeable, and without an independent credit system.
A deeper issue lies in the inherent contradiction within Token, which simultaneously fulfills two roles that should ideally be separate: on one hand, it is a ‘measure of cost,’ quantifying the physical consumption of computational power; on the other hand, it serves as a ‘measure of value,’ assessing the utility of intelligent output. Traditional currency only performs the latter function. This duality makes it difficult for Token to achieve stable equivalent exchanges like ‘one dollar.’
If we extend our view to a broader historical scale, we find this is not a new phenomenon. For instance, during the initial commercialization of electricity in the late 19th century, companies had varying voltages, frequencies, and billing methods, all incompatible with each other. After decades of standardization, public utility regulation, and market-oriented trading, electricity evolved into the universally accessible infrastructure we know today.
Token is currently at a similar ‘pre-standardization’ stage. Based on industrial evolution patterns, we can anticipate the emergence of cross-platform interoperability protocols, unified measurement standards, and even financial instruments like Token futures.
However, the most profound significance of Token lies in its role as the first systematic way for humanity to quantify and trade ‘intelligent services.’
Therefore, the focus of our discussion should not be whether it deserves the label of ‘currency.’ Instead of fixating on definitional disputes, we should explore the new economic paradigm it is creating, which may lead to theoretical frameworks surpassing traditional monetary theory.
From the perspective of traditional economics, the analogy between Token and ‘currency’ does not hold.
If strictly measured against the three classic functions of money as outlined in economics textbooks, Token appears inadequate in almost every aspect.
As a unit of account, although it serves as the pricing unit for AI services, the lack of equivalence among vendors and extreme price volatility prevent it from providing a stable value reference.
As a medium of exchange, it facilitates one-way conversion from fiat currency to AI services but lacks universal liquidity and cannot be used in any scenarios outside of AI. As a store of value, its deflationary trend is evident; holding prepaid Tokens will only continue to depreciate, offering no value preservation whatsoever.
Compared with the ‘national credit’ of sovereign currencies, the ‘credit’ of Tokens has four fatal weaknesses.
The first is fragility: companies may go bankrupt, models may become obsolete, and technologies may be disrupted; Tokens lack an external guarantee mechanism.
The second is fragmentation: each company establishes its own credit system, requiring users to evaluate the credibility of different platforms separately.
The third is volatility: the rapid pace of technological iteration in the AI field means today’s leaders may be replaced tomorrow, leaving an extremely short window of credit.
A fourth dimension, often overlooked, is data security trust—by sending sensitive information to platforms through Tokens, users are essentially trusting their data protection capabilities. In the event of a large-scale breach, Token credit would collapse instantly, similar to a bank run.
Ultimately, the credit of currency is built upon a social contract. The recognition of its value by society as a whole often requires decades or even centuries of accumulation. Currently, Tokens have only established a technical contract—you believe the technology is useful, so you purchase Tokens. Transitioning from a technical contract to a social contract requires a lengthy process of credit accumulation, including industry standardization, the establishment of regulatory frameworks, and continuous market validation.
As mentioned earlier, Token is exhibiting new characteristics not covered by traditional monetary theory: it is both a consumer good (end users pay to use it) and a factor of production (developers use it to build AI products). This dual identity is extremely rare in the history of money.
More importantly, it is playing the role of ‘smart access rights.’ In the AI era, without Token, intelligent services cannot be accessed, which endows it with strategic attributes akin to fundamental factors of production.
Measuring Token through traditional monetary theory is like evaluating cars using traffic rules from the horse-drawn carriage era. It’s not that the car is inadequate, but rather that the rules need updating.
From the current perspective, the ultimate form of Token may not necessarily be ‘currency,’ but rather more likely ‘a foundational measurement protocol for the intelligent economy,’ used to define the standard unit and rules for exchanging intelligent services. The field of economics will sooner or later have to construct a new theoretical framework for this novel form of value exchange.
Section 03: The pricing evaluation of Token represents an ‘exchange of intelligent service capabilities.’
Since Token can no longer fit within the framework of traditional currency, could comparing it to a reference closer to its origins, such as ‘traditional computing resources,’ reveal its contours? The answer remains negative.
Whoever controls the pricing anchor will possess the discourse power in the AI economy. There are three fundamental dimensional differences between Token and traditional computing resource pricing, and the distinctions are clearly visible.
First, there is a leap from cost-based pricing to value-based pricing: the price of traditional computing resources closely aligns with hardware costs, following a ‘cost-plus’ logic; whereas Token pricing has detached from hardware cost anchors and shifted towards ‘value-based pricing.’ For instance, different versions of large models of GPT running on the same GPU cluster may vary in Token prices by tens of times, with the difference stemming from model capability rather than the hardware itself.
Second, there is a paradigm shift from time leasing to capability purchasing. The essence of the traditional model is ‘renting hardware time,’ with payment made per hour regardless of whether users efficiently utilize the computing power. In contrast, the Token model ties payment to ‘actual intelligent output,’ making it closer to ‘pay-per-performance.’ This continues the evolution of cloud computing from IaaS to SaaS—progressively moving away from physical resources and toward value generation.
Third, there is a divergence from linear value to non-linear value. Renting one more hour of CPU usage results in roughly linear growth in output; however, Tokens exhibit significant ‘user skill premium’: spending the same one thousand Tokens, a cleverly crafted prompt may generate a hundred times more commercial value than a verbose question. The input-output ratio highly depends on the user’s cognitive level and usage skills—a phenomenon absent in traditional computing resources.
Ultimately, the pricing of traditional computing resources measures the ‘time occupancy of material resources,’ while Token pricing measures the ‘exchange of intelligent service capabilities.’ This marks the first time in human economic history that ‘intellectual output’ has been broken down into quantifiable and tradable atomic units. It is not only a new form of performance-based billing but also the beginning of a pricing paradigm for the intelligent economy era.
II. What is the cost structure of Tokens?
01 The pricing power of Tokens: Scarcity is key
The production cost of Tokens consists of four key elements: chips, electricity, data, and talent. However, there is no single ‘anchor’ determining the price; instead, it is a multi-anchor system that dynamically shifts over time.
The term ‘anchor’ here refers to the critical element in the cost structure that truly determines the price floor, much like how a ship relies on an ‘anchor’ to fix its position. Similarly, prices need a hard constraint to ‘anchor’ their lower limit.
At different stages, the key factors determining the price floor of Tokens vary, and pricing power will gradually shift among the four elements in sync with technological evolution.
If we expand this migration path along the timeline, three distinct phases become clearly visible:
In the short term (one to two years), chips are the primary anchor. For instance, the current shortage of NVIDIA GPUs means chip costs account for 60-70% of total inference costs, and the availability of chips directly determines the supply and price of Tokens.
How does the scarcity of GPUs progressively affect Token prices? Upstream, NVIDIA’s limited GPU production capacity requires orders to be queued for over a year. Midstream, AI companies like OpenAI and Anthropic cannot secure enough cards, limiting the scale of inference clusters they can build. Downstream, the volume of Tokens that APIs can output is constrained, keeping prices naturally high, with the ultimate burden falling on every AI user.
In the medium term (three to five years), electricity will become a rigid constraint: As chip supply increases and architectural innovations reduce chip prices, electricity—constrained by physical laws and compounded by growing societal attention to the surging energy consumption of AI data centers—will emerge as the non-compressible baseline cost.
In the long term (over five years), talent and knowledge density will dominate pricing. Chips and electricity represent physical costs, which will continually be reduced by technological advancements; however, the research talent and data accumulation required to train top-tier models are scarce resources, forming the core anchor for high-end Token pricing.
Thus, the pricing anchor is shifting from GPUs → energy → talent.
02 The Two-Tier Structure of Tokens: Cost as a Floor, Capability Driving Differentiation
Even with pricing power clarified, why do Token prices still vary? The key lies in distinguishing between two types of costs—inference cost and training cost.
Inference cost refers to the marginal cost of producing each Token, primarily driven by chips and electricity; training cost represents the one-time investment in creating model capabilities, dominated by computing power, talent, and data.
Current Token pricing mainly reflects inference costs, while the premium for high-end models essentially represents the amortization of training costs. Tokens are evolving from homogeneous commodities into quality-tiered products: the anchor for low-end Tokens is electricity and maintenance, whereas the anchor for high-end Tokens is talent and R&D. The overall direction of the pricing anchor is shifting from ‘material scarcity’ to ‘intelligence scarcity.’
One point to add is that if price were merely a result of cost transmission, it would fail to explain a phenomenon—why can Token prices for different models under the same company vary by tens of times?
There exists a gap here, which can only be explained by introducing value-based pricing logic: for instance, GPT-5 can perform complex reasoning, code generation, and specialized analysis tasks that GPT-4 struggles with, and its leap in capability supports a premium far exceeding costs.
Users demonstrate a clear willingness to pay for ‘smarter AI,’ and corporate clients are willing to pay a premium for reliable business decision-making, much like they would pay higher fees for a senior consultant rather than a junior assistant.
From a market strategy perspective, this significant price disparity also reflects ‘second-degree price discrimination’: manufacturers offer product lines of varying quality, allowing users with different payment capacities to self-select into tiers. High-priced versions target corporate clients, while low-priced versions focus on customer acquisition and maintaining user bases, with a single product portfolio covering the entire demand curve.
The combination of these two forces constitutes the complete mechanism for the current Token price formation: the bottom is supported by the cost transmission caused by GPU bottlenecks, while the top is driven higher by the value premium of model capabilities.
At a deeper level, this structure of ‘capability stratification determining price stratification’ marks the first time humanity has directly priced ‘intelligence levels’ through market mechanisms. For instance, lawyers charge by the hour, doctors charge per consultation, whereas AI charges based on the ‘quality tier of cognitive units.’ A new economic order of ‘intelligence-tiered payment’ is taking shape.
Third, how will Token prices and industry cycles evolve?
01 Beyond Moore’s Law: Token Prices Exhibit a ‘Downward Curve’
When GPU bottlenecks are no longer the dominant factor (this is the premise), Token prices will continue to decline along a long-term curve extended by Moore’s Law, and the rate of decline will be more astonishing than anyone anticipated. The price war for Tokens will make AI accessible worldwide, just as the price war for electricity made electricity available globally.
Consider a set of real numbers: When GPT-4 was first released in March 2023, the API price was $30 per million input Tokens and $60 per million output Tokens; today, models with equivalent intelligence have seen API prices drop significantly, with some models experiencing reductions exceeding 95%.
Behind this lies the synchronous superposition of three downward curves:
• At the hardware level, NVIDIA’s H100 typically achieves about three times the LLM inference efficiency compared to the previous generation A100, and the new B200 takes it a step further from the H100;
• At the framework level, inference optimization tools such as vLLM and TensorRT-LLM increase the throughput of the same card several times over;
• At the architecture level, MoE models (e.g., DeepSeek-V3) reduce inference computational costs by several multiples under similar capability levels through sparse activation mechanisms.
The token of foundational models is highly likely to approach economic zero within the next five years, potentially becoming as inexpensive as tap water, where most users will no longer need to consider costs. The true turning point lies not in the price itself but in the business model.
For instance, ChatGPT’s free version offers GPT-4o access to all users, while Google’s Gemini 2.0 Flash emphasizes low-cost, high-throughput pricing strategies. These changes indicate that “token-based billing” is no longer the sole model, and the industry is moving toward a coexistence of various business models, such as “free basic services + value-added payments.” Vendors’ revenue streams have expanded from single token charges to subscription services, enterprise solutions, and closed-loop ecosystems.
Of course, the trend toward zero token costs does not equate to zero AI usage costs. When tokens become nearly free, new scarcities will emerge: high-quality data, human attention in processing AI outputs, and verification of the credibility of AI-generated content—economics will always chase and focus on the next scarce resource.
Phase 02: The Industry Cycle Tokens Will Experience—Gradual Stratification
If we shift our focus from the price of individual tokens to the entire supply-side market, this multi-year arms race for computing power will ultimately follow a typical industry cycle: “short-term scarcity → mid-term surplus → price wars → industry consolidation.”
First, let’s examine the capacity side. Currently, major technology companies are investing tens of billions of dollars annually in building AI data centers. In 2024, global tech giants collectively exceeded $250 billion in capital expenditures on AI infrastructure: Microsoft alone announced an $80 billion investment in AI data centers for the fiscal year 2025. Meta disclosed capital expenditures of nearly $400 billion in its 2024 annual report, while Amazon and Google’s annual capital spending has reached the hundreds of billions. Even sovereign funds from the Middle East have begun direct involvement—MGX Fund of the UAE, in collaboration with Microsoft and BlackRock, announced the creation of an AI infrastructure fund.
By 2026, instead of contracting, these investments will continue to increase significantly. Converting these investments into GPU clusters and data centers will require two to three years to fully materialize and release capacity.
Meanwhile, although demand for agents and automated workflows is indeed growing explosively, enterprises’ procurement pace lags far behind the deployment of capacity. The demand curve represents a smooth integral, while the capacity curve exhibits step-like pulses. The mismatch between the two curves almost inevitably leads to medium-term oversupply or even excess capacity.
The result of overcapacity will be price wars, which will gradually lead to a stratified landscape:
One layer comprises the low-end token market, used for basic conversations and simple tasks. Competition here will be extremely fierce, with companies fighting on price, squeezing profit margins to very low levels—similar to the price war phase during the early days of cloud computing.
Another layer is the high-end Token market, used in complex reasoning and professional scenarios, where effectiveness and stability are more valued. Whoever is smarter and more reliable can command a higher price.
Among these, open-source models (such as Llama and Mistral) will play an important role as price “challengers.” Open source will not eliminate closed-source giants, but it will establish a price ceiling for low-end Tokens and expand Token supply from a few major players to thousands of independent deployers, profoundly altering the market structure.
Fourth, who is consuming Tokens, and where will this consumption trend lead?
01 When Agents begin working 24 hours a day, Token demand accelerates.
As Agents and automated workflows become widespread, the growth in Token demand is transitioning from a “normal curve” into a more complex, and even somewhat counterintuitive, pattern. This can be summarized as: “Overlapping S-curves → J-shaped explosion → Sublinear convergence.”
In the initial phase (2023–2027), Token growth will primarily be driven by human users. This follows a typical S-curve: an increasing number of people are beginning to use AI tools for writing emails, revising resumes, coding, learning assistance, and other expanding use cases. However, this curve has a natural cap—the time and attention available to humans. Since there is a limit to how much time one person can spend using AI daily, overall growth, though steady, will not exhibit an extremely steep slope.
During this phase, it is fundamentally still “humans using AI.” Token consumption is directly tied to human behavior. Agents act as a super amplifier of Token demand.
The real turning point will come from the second S-curve—growth driven by Agents (2025–2030). Once Agents begin integrating into enterprise systems, the logic of Token consumption changes entirely: it no longer follows “human activity” but rather aligns with “machine processes.”
For example, in customer service scenarios, one Agent can handle hundreds or even thousands of conversations simultaneously, automatically retrieving knowledge bases, generating responses, and summarizing records. In software development, a task may be broken down into multiple Agents working together—writing code, running tests, identifying bugs, and conducting reviews, with each step consuming Tokens. In e-commerce and operations, Agents can autonomously generate content, place advertisements, analyze data, and optimize strategies in feedback loops, forming a continuously operating closed system.
These changes, when combined, create a noticeable “multiplier effect.”
First, Agents can operate 24 hours a day, no longer constrained by human work-rest cycles; second, multi-Agent systems will generate recursive calls—Agents invoking other Agents and having results verified by additional Agents, with each layer adding to the Token cost. As automation costs continue to decline, an increasing number of previously ‘not worthwhile’ long-tail tasks are being activated, further expanding the application boundaries of Tokens.
In this process, a classic economic phenomenon—the ‘Jevons Paradox’—will become highly evident: increased efficiency does not reduce overall resource consumption but rather activates more use cases due to lowered barriers, ultimately driving up total demand. The same applies to Tokens: the cheaper and more efficient they become, the more they are used.
However, this growth will not continue indefinitely. Enterprises must always face a fundamental constraint: the cost of Tokens must be less than the value they create. When certain Agents consume large amounts of Tokens without generating sufficient returns, they will be optimized or phased out.
Therefore, after experiencing rapid expansion in the early stages, the growth in Token demand will gradually return to rationality, transitioning from a steep J-curve to a more moderate ‘sublinear growth.’
From a deeper perspective, the real shift is that the primary consumers of Tokens are transitioning from humans to machines.
This means that the upper limit of the Token economy is no longer determined by population size or human time but depends on the total value that the entire economic system can create, as well as the underlying computing power and energy supply capacity.
In terms of the timeline, the period from 2026 to 2028 is likely to be a critical inflection point. With the large-scale deployment of Agents within enterprises, the demand for Tokens could experience a one- to two-order-of-magnitude leap in a short time, leading to a true ‘J-curve explosion.’ At that stage, what we will see is no longer ‘people using AI,’ but the entire economic system autonomously operating with AI.
This also explains why the growth of Tokens is fundamentally not driven by an increase in user numbers but is instead a story of rising automation levels and economic restructuring.
02 The tipping point of Token demand: Watch for ‘two critical thresholds’
For the price of Tokens, different groups exhibit significant variations in their price sensitivity.
Corporate users are highly sensitive to price. When a company is calculating the cost of ‘replacing some human labor with AI,’ every single cost element will be meticulously analyzed. If the price of tokens drops by half, tasks that were previously deemed unprofitable immediately become feasible. For instance, what was once limited to providing AI customer service only to VIP clients can now extend to all customers; tasks that were handled in English alone can now cover over a dozen languages. As a result, usage typically does not increase by just 50%, but instead doubles or even triples.
However, most end consumers do not perceive changes in price because the majority use subscription services that are either free monthly or fixed-fee based, temporarily shielding them from price signals.
Nevertheless, with the recent rise of Agent-type products like ‘Lobster,’ many C-end consumers are once again being drawn into the logic of consumption-based pricing, and their sensitivity to price is quickly aligning with that of corporate users.
The true determinant of the upper limit of Token market growth is not solely dependent on behavioral changes among these two types of existing users, but rather the collective influx of a large group of people who have never used AI before. This represents the largest potential source of Token demand.
Once the price crosses a certain critical threshold, it does not merely stimulate an increase in existing user demand, but unleashes the latent demand potential of hundreds of millions of new users and entirely new application types.
When will this explosive growth occur? It requires the simultaneous fulfillment of two conditions.
The first is the price threshold. For example, when the cost per million Tokens drops to $0.1 or lower, meaning the cost of completing a task with AI becomes less than one-tenth of human labor costs, then the question of ‘whether to use AI for this task’ no longer needs to be debated—it becomes the default choice.
The second is the ‘threshold’ of awareness. Many people do not avoid using AI because it is unaffordable, but because they simply do not know what AI can do for them. For instance, drafting contracts, organizing invoices, translating documents, planning trips, or providing personal health consultations—these scenarios remain ‘heard of but rarely used’ for most people. The dissemination of this awareness itself takes time and a societal process.
These two thresholds interact to form a self-accelerating flywheel: lower prices encourage more people to try AI—trying leads to word-of-mouth and broader awareness—broader awareness generates greater demand—and increased demand drives further reductions in costs and prices. Once this flywheel starts spinning, the growth in Token demand will no longer be linear but will explode suddenly.
A surge in Token consumption does not necessarily equate to real value creation.
If Token is becoming a new fundamental production factor in the AI era, can its consumption be used to measure the operational level of an economy, similar to electricity usage or steel production?
The most critical obstacle lies in the “heterogeneity” of Token consumption. A single Token may generate economic value that varies by ten thousand times depending on the context. Some Tokens are utilized to support pivotal business decisions, creating significant value, while others are consumed in aimless conversations with almost no economic output. Adding these two types of Token consumption indiscriminately would yield a figure that has little relevance for measuring economic output.
There is no doubt that at least two prerequisite issues need to be addressed.
First, it is necessary to distinguish between “productive consumption” and “consumptive consumption.” Only the Token consumption genuinely used to complete work tasks and generate commercial value has a stable positive correlation with GDP growth.
From this perspective, the volume of enterprise API calls might serve as a better economic proxy variable than “total Token consumption,” as it effectively filters out a substantial amount of consumptive chatter and entertainment-related usage.
Second, a measurement standard for “Token economic efficiency” needs to be established, i.e., the ratio of how much economic value a unit of Token can create. If this indicator continues to rise, it indicates that AI’s contribution to the economy is becoming more efficient; if, conversely, it keeps declining, it could signal bubbles or inefficiencies. This metric itself serves as an important economic diagnostic tool.
Even after resolving these two prerequisites, the relationship between Token consumption and economic output is still unlikely to follow a straight line but rather resembles an S-shaped curve.
In the early stages, businesses are still learning and experimenting with how to use AI, leading to rapid growth in Token consumption but lagging output. In the middle stage, applications gradually mature, and economic output expands rapidly alongside Token consumption. In the later stage, diminishing marginal returns set in, and output growth slows down. Different industries and countries currently occupy different positions on this curve, which explains why opinions about “AI’s impact on GDP” remain varied, as each economy moves at its own pace.
An easily overlooked reverse trend is also at play: as model capabilities improve, reasoning optimization progresses, and user proficiency increases, the number of Tokens required to accomplish the same task actually decreases.
Code that once required thousands of Tokens to write well a year ago now may only require hundreds. This means that the total growth in Token consumption will likely lag behind the growth in economic value — “Token efficiency” will become a continuously improving metric, further diluting the significance of Token consumption as a direct indicator of output.
The more fundamental challenge lies elsewhere: a significant amount of Token value does not pass through transactions and therefore is not captured by GDP statistics. For example, when a student uses AI for academic tutoring, an individual uses AI for health consultations, or a creator uses AI to enhance personal productivity, all these Token consumptions generate real value. However, the majority of it does not enter the transaction process and thus is not included in GDP calculations. According to the existing accounting system, this portion of value exists objectively but remains unmeasurable, much like ‘dark matter.’
As an increasing amount of value creation occurs outside of traditional transactions, as AI makes non-market services readily accessible, and as individual efficiency improvements far exceed the capacity of conventional statistical methods to capture them, the familiar ‘discourse system’ we rely on to forecast (or ‘evaluate’) the economy may no longer keep pace with the rapidly advancing artificial intelligence era.
Therefore, the narrative surrounding the Token economy is not primarily about whether it will eventually become a currency; instead, the focus is on the redistribution of efficiency, structure, and control that underlies it.
Whoever can create more value with fewer Tokens will gain mastery over new productive forces.
Whoever can build a more efficient Agent system will possess a new organizational advantage.
And whoever can define standards, control entry points, and establish ecosystems may attain a position akin to ‘new digital infrastructure’ in this restructuring.
Transitioning from selling Tokens to delivering outcomes is analogous to shifting from ‘selling electricity’ to ‘selling illumination.’
In this sense, Token is not the answer; rather, it serves more as a signal.
It reminds us that:
We are moving from an era of ‘using tools’ to an era of ‘building automated systems.’
From ‘labor-driven growth’ to ‘machine-driven economy’;
From ‘incorporating AI into the production function’ to ‘the production function itself being rewritten by AI.’
And this, perhaps, marks the true beginning of this transformation.
Editor /rice
