When Token becomes a business, it's time for enterprises to do the math on their AI investments.
Calculating the value of each Token clearly is the watershed for AI implementation
By Luo Zhenhao
Edited by Liu Junhong
In the past two years, the most frequently heard voice in enterprises is "I want to adopt AI", but many people have no clear perception of what AI is specifically used for and what effects it delivers. Too many enterprises choose to "use it first and think later" until one day, an extremely expensive Token bill suddenly comes out, and they find that the investment cannot be accounted for at all.
However, some enterprises have made the accounts completely clear: a leading supermarket chain giant used AI to negotiate supplier contracts, cutting costs by 3%, achieving an ROI of 4 times, and extending the payment period by 35 days...
For enterprise applications, how to make the AI accounts clear is becoming more important than ever.
In the past two years, the most visible change in the AI industry is that models are getting larger and computing power is getting stronger. But after entering the stage of agents and enterprise-level applications, enterprises are facing a more realistic problem: what exactly have the computing power and Tokens spent brought back?
This is also why "Token Economics" has begun to enter the industrial vision.
According to the disclosure of the National Data Administration: at the beginning of 2024, China's daily average Token call volume was 100 billion; by the end of 2025, it jumped to 100 trillion; in March this year, it exceeded 140 trillion, an increase of more than 1000 times in two years. Behind this set of data, the way AI is used is changing. In the past, enterprises were used to calculating how many GPUs they bought and how many data centers they built. Now they are calculating how many tasks an AI investment has completed, how much time has been saved, and how much revenue has been generated.
When AI moves from "generating content" to "working continuously", Token is also transforming from a technical indicator to an operating indicator.
In the past, Token was more of a measurement unit when calling models. Now, as agents start to conduct multi-round reasoning, call tools, pull context, and even allow multiple agents to collaborate for a long time, the demand for Tokens has become more diversified.
In the view of Dr. Lai Junjie, Vice President of Engineering and Solutions for NVIDIA China, an agent task may involve 50 to 500 consecutive model calls, and multi-agent collaboration will further amplify Token consumption. This means that the accounts of enterprise AI are being recalculated.
Therefore, what really deserves attention is no longer just how cheap Tokens are, but how the value of Tokens itself is defined, how demand is predicted, how supply improves efficiency, and how it is finally converted into business returns.
NVIDIA summarizes it as four pillars of Token Economics: Value, Demand, Supply, and Monetization.
From this perspective, what Token Economics actually discusses is a new operating logic after the AI industry moves from technical investment to productivity realization.
If the core issue for enterprises to discuss AI in the previous stage is how much computing power to invest, the real issue at this stage is becoming whether the Tokens produced by these computing powers can form measurable and sustainable business value.
Focusing on this issue, Wang Yisu, founder of Lightcone Intelligence, and Dr. Lai Junjie, Vice President of Engineering and Solutions for NVIDIA China, held a podcast on "Token Economics".
NVIDIA, standing on the side of AI infrastructure, discusses how models, GPUs, networks, reasoning optimization and Serving frameworks jointly determine the unit cost and delivery efficiency of Tokens. As an industry observer, Lightcone Intelligence, from the perspective of enterprise applications, asks whether AI has entered real business and how to put Token consumption, process efficiency and final operating results into the same table.
The following is the transcript of the dialogue, edited by Lightcone Intelligence without changing the original meaning:
Q: Combining your practical experience, let's talk about how to use Tokens more economically and efficiently?
Dr. Lai: Inside NVIDIA, we strongly encourage employees to explore how to integrate AI into daily work processes to make AI an assistant that improves personal efficiency. Our CEO Mr. Jensen Huang also mentioned that every employee will have their own AI assistant in the future. All our internal teams, including engineering, product, legal, marketing, finance and so on, are constantly exploring how to integrate AI into specific work processes. From this perspective, the collaboration between humans and AI is becoming a new way of working.
From my personal experience, the key to using Tokens is not simply to pursue reducing consumption, but to find scenarios where AI can truly amplify personal capabilities, so that each AI interaction can generate higher value. For example, in information sorting, data analysis, problem decomposition or other highly repetitive tasks, allowing AI to participate in more links can help us reduce repetitive work, so that we can devote more time to judgment and creative parts.
Wang Yisu: We can't live without AI now. As content creators, our main expenses are on in-depth content research and video production. I have also tried some small AI coding software. Specifically, data query and in-depth research in daily writing are the most frequent use scenarios, and the video team is also using AI more and more in material generation. In terms of cost, everyone spends about hundreds of yuan per month, and we are using memberships of some foreign AI software and mainstream high-quality domestic AI memberships.
Q: For the same 1 million Tokens, calling different models and applying them to different tasks may produce completely different final values. Dr. Lai, can you first talk about what Token is exactly?
Dr. Lai: Token is the basic unit for AI models to process and generate information. It can also be said to be the smallest unit of intelligence produced by AI factories, and it is an important indicator to measure AI workload. But Token itself is not equivalent to value.
We can understand it with a simple formula: the cost per unit of Token is equal to the comprehensive operating cost of the data center per hour divided by the number of Tokens delivered in real time by the data center per hour. The cost here includes many aspects, GPU, CPU, network, storage, power supply and so on. Many enterprises, when evaluating AI infrastructure investment, tend to only focus on one of the costs, such as the price of GPU. But what really determines the cost of Token is often the overall efficiency, that is, how many available Tokens the same investment can produce.
Therefore, from the perspective of Token Economics, we need to pay attention to both the total cost of ownership of the data center and the end-to-end output efficiency, and shift the evaluation criteria from input indicators to output indicators.
Q: That is to say, the value of Token is not single, but related to the infrastructure, model capability, application scenario and final business result behind it. What are the key factors that affect the value of Token specifically?
Dr. Lai: To judge whether a Token is valuable, we can look at two core dimensions.
The first dimension is intelligence content. For the same 1 million Tokens, generated by a cutting-edge reasoning model with long context capability and generated by a lightweight model, the understanding ability, reasoning ability and context information carried behind them are completely different, and the former has higher intelligence density.
The second dimension is interactivity, that is, the speed at which Tokens reach users. For scenarios such as programming assistants and high-frequency dialogue, only Tokens that reach users near real time are valuable; for tasks such as batch processing documents, the delay of Tokens is not so sensitive.
But both dimensions must be based on a premise: value is relative to the use case. A small model post-trained for a specific scenario may achieve or even exceed the effect of a large model with only a fraction of the cost. The additional intelligence premium is only worth paying when it is usable.
Therefore, when making actual evaluations, enterprises can ask themselves several questions: How strong intelligence does this task require? How fast is the response? How high is the cost of result errors? For example, in scenarios such as medical care and finance, accuracy and reliability are hard indicators.
Wang Yisu: Token itself is more like a basic input, and the maximum value it can generate depends on which tasks it is applied to and how enterprises integrate it into business processes. According to my observation, AI scenarios can be roughly divided into two categories.
One category is horizontal scenarios, that is, general needs that exist across industries, such as enterprise knowledge assistants, customer service Agents, office assistants, code assistants and so on. Such applications cover a wide range, and their value is more reflected in helping employees improve efficiency and reduce repetitive work.
The other category is vertical industry scenarios, that is, applications that combine specific industry knowledge and business processes, such as investment research and analysis in the financial field, auxiliary diagnosis in the medical field, equipment operation and maintenance in the manufacturing field, etc. Such scenarios usually require AI to understand more professional data processes and business logic. In both scenarios, Tokens can directly generate economic value.
Q: What is Token Economics? Why has it become an increasingly important framework for enterprises to discuss AI today?
Dr. Lai: From NVIDIA's perspective, Token Economics refers to the economics behind the generation and use of Tokens. As data centers evolve from places for storing and processing data to AI factories dedicated to producing intelligence, Token is the product unit output by AI factories.
Token Economics needs to answer not just "how much is a Token worth", but four things: First, how Tokens are valued, that is, how much the intelligence and interactive experience it carries are worth; Second, how Tokens are consumed, that is, who is using it, how much they use, and in what scenarios; Third, how Tokens are supplied, that is, how to produce Tokens efficiently and at low cost through infrastructure; Finally, how to monetize, that is, how to finally convert into revenue and business value.
This concept is becoming more and more important, largely because AI is moving from the pilot and large-scale training stage to the large-scale application and continuous reasoning stage. With the development of AI agents, a task may require longer context, more rounds of reasoning and tool calls, and the demand and usage of Tokens will change greatly.
After Tokens enter large-scale production and use, enterprises need to pay attention to its output, efficiency and unit cost just like managing a production system. From this perspective, Token Economics connects the technical input of AI and economic output. Enterprises should not only consider whether Tokens can be generated more efficiently and at lower cost, but also consider where these Tokens are used and what kind of business applications and value creation they can support in the end.
Wang Yisu: Let me start with a number that makes me think Token Economics must be discussed seriously. From the beginning of 2024 to March 2026, China's daily average Token call volume has soared from 100 billion to 140 trillion, an increase of 1400 times in two years.
Behind this number, three things are happening at the same time.
First, enterprises used to focus on computing power input, but now they need to focus on intelligent output. The previous logic was how many GPUs you bought, how much electricity you used, and how many data centers you built, which was accounted for from the input dimension. After the emergence of Token Economics, enterprises have to calculate another account: I spend 1 dollar to buy Tokens, how many tasks can it help me complete? How much time has been saved? How much revenue has been generated? This is the accounting from the output dimension. From input to output, this is the watershed for the entire AI industry to switch from technical competition to commercial implementation.
Second, new measurement units have emerged within enterprises to re-manage AI investment. The most typical example is that some companies proposed DAA, Daily Active Agents this year. Why did this happen? Because an enterprise only looking at how many Tokens it has burned is the same as a steel mill only looking at how much coal it has burned. The amount of coal burned does not represent the quality of steel output. IDC also released a DAA report. We also observed that some operators began to redesign their business models according to "number of users + AI terminal penetration rate + Token consumption + economic value per unit Token".
Third, AI management within enterprises is becoming a joint matter for CFOs and CIOs. In the past, AI investment was managed by CIO alone, but now the magnitude of Token consumption has reached the level that CFOs must intervene. More and more enterprises have begun to set up positions dedicated to managing AI resource consumption. This position needs to understand both models and finance, and the question to be answered is: What exactly has every AI investment brought back?
To sum up, Token Economics is an inevitable product of the transition of enterprise AI investment from the computing power era to the intelligence era. This concept has evolved from discussions among a few people to industrial consensus in the past year, which is why it has become an increasingly core framework for enterprises to discuss AI today.
Q: When agents enter enterprises, what changes will happen to the use of Tokens? What are the common types of agent applications at present, and what are the different demands for Tokens?
Dr. Lai: From the perspective of workload and actual use cases, I will roughly divide common agents into three categories.
The first category is dialogue and assistant type, such as customer service assistants and knowledge Q&A. This kind of application requires timely response, high latency requirements, needs to support a large number of concurrent users, each task is relatively short, and Token consumption is relatively easy to predict.
The second category is workflow-type agents. For example, let the agent help you complete a research report. It needs to understand the task first, then retrieve materials, call tools, analyze information, and generate results. Generally, it will go through multiple rounds of model calls, and the context is relatively long.
The third category is more autonomous, long-running agents, including multi-agent collaboration, which may run continuously for tens of minutes, several hours or even longer, constantly call various tools, update the context, and even start sub-agents during the process.
NVIDIA has done some tests ourselves. An agent task may involve 50 to 500 or more consecutive model calls. In the multi-agent scenario, Token consumption will be further amplified. Therefore, after the development of agents, Token demand is becoming more and more unpredictable. How long the context is, how many rounds of reasoning are needed, how many tools are called, how many agents are running at the same time, and what the response speed requirements are, all these will change the demand structure of Tokens.
Wang Yisu: To add a trend judgment, in the next one to one and a half years, AI office agents are likely to become the largest scenario of enterprise Token consumption, because office is one of the few scenarios that have both high-frequency use and clear willingness to pay. In Q2 2026, the capital expenditure of some leading technology companies skyrocketed by 70% to 300%, and AI office accounted for a large part of it. In Dr. Lai's classification, office agents are moving from workflow type to autonomous collaboration type, and Token consumption will be further amplified.
Q: To apply Token Economics to business in practice, does enterprises have a set of frameworks that can be used for decision-making?
Dr. Lai: When enterprises consider AI investment, the core needs to answer several questions: Does this investment generate value? How many resources do I need to invest? How to ensure the stable operation of AI capabilities? Can it finally be converted into business returns? Focusing on these issues, we can use the four pillars of Token Economics proposed by NVIDIA to understand.
The first is the value of Token: whether the output of AI really solves business problems. The second is the demand for Tokens: enterprises need to understand the demand for Tokens in different users and different scenarios, including usage scale, task type, peak demand and so on. The third is the supply of Tokens: enterprises need to focus on how to improve Token supply efficiency and reduce unit cost through model, computing power and infrastructure planning. Finally, what many people care about is how to monetize Tokens: how to convert the use of Tokens into actual business value and commercial returns. Combining these four pillars, enterprises can comprehensively evaluate Token investment from multiple dimensions of Token demand, supply and value.
Wang Yisu: These four pillars are not only about technology, but also the core issues of enterprise operation, so they need to be designed from the top-level architecture. I map it to four roles inside the enterprise, and I think this mapping relationship is very useful.
CEOs and CFOs care about monetization, how much revenue and efficiency every AI investment brings back, and the accounts must be clear; business leaders care about value, whether AI output can be really used by the business team, the situation that the demo is amazing but no one uses it after implementation is too common; product and operation teams care about demand, whether the peak load can be withstood, each additional work order processed by a service-oriented agent will consume 5000 to 15000 more Tokens. An enterprise with 5000 employees has 10000 work orders per day, which is 5