Programming no longer requires typing code; it can be handed over to AI through conversation, or even by speaking into a microphone, saying "make me a certain feature." This idea was considered a "fantasy" five years ago, but today, it seems to have become a "basic operation" for programmers.
This is just the "tip of the iceberg" among current AI applications. The AI industry is at a crossroads. On one side is the frenzy on the technology supply side, with Nvidia's stock price doubling and doubling again, new large models emerging continuously, and computing power scaling exponentially. On the other side is the confusion on the demand side, where companies invest in computing power but struggle to calculate whether the investment is worth it. As a value connector, Token is becoming the key to whether AI can deliver value.
Token Becomes the "Measurement" of the AI Era
In early 2024, China's daily Token call volume was about 100 billion. By the end of 2025, this number jumped to 100 trillion. In March 2026, data released by the National Data Administration showed that the daily call volume had exceeded 140 trillion, with a growth of over a thousand times in two years.
What does this mean? Professor Wei Zhexian of Renmin University of China compared Token to electricity: if Token is likened to "electricity in the intelligent era," we are currently in the "stage where electric lights have just been invented." He stated, "Today is still just the prologue. We have not yet truly witnessed its release in all corners."
According to IDC data, the Token call volume in China's Model-as-a-Service (MaaS) market is expected to reach 40,000 trillion in 2026, with revenue of approximately 18.6 billion yuan. Meanwhile, IDC predicts that the global annual Token consumption will rise from 0.0005 Peta Tokens in 2025 to 150,000 Peta Tokens in 2030, with a compound annual growth rate of 3418%. By 2031, the number of global active agents will reach 350 million.
The explosive growth of Tokens has brought a fundamental question to the surface: How much is a Token worth? Who sets the price?
In response, Huang Wei, Deputy Director of the Information and Industrialization Integration Research Institute of the China Academy of Information and Communications Technology, provided his conceptual framework answer. Huang Wei deconstructed the value of a Token into five dimensions: production cost, production efficiency, accuracy, ecological value, and security compliance. Production cost includes chip depreciation, power consumption, model optimization, and system scheduling. "Every time you ask AI a question, it may involve the entire chain from GPU to storage, from the software stack to network interconnection," Huang Wei said.
But the ideal is full, the reality is meager. In practice, Token pricing varies widely across different industries and scenarios. A report from Zhongtai Securities shows that the value difference of Tokens across scenarios can reach up to a hundred thousand times. Tokens in drug R&D can average $1,000 per million Tokens; while Tokens for casual chat might cost only $0.01 per million Tokens. The same one million Tokens could help a pharmaceutical company screen a potential drug molecule, whereas in a social app, it might just be a few lines of casual chat.
Lenovo Group Vice President and China Chief Strategy Officer Abulikemu (hereinafter referred to as "A Mu") stated, "The manuscript fee for a thousand words I write is different from that for a thousand words written by an author." Tokens unify the unit of measurement, but the "intellectual level" behind the Tokens is the key to determining the upper limit of value.
In Huang Wei's view, only "effective Tokens" are truly valuable to industries and enterprises. Wei Kai, Director of the Artificial Intelligence Research Institute at the China Academy of Information and Communications Technology, has also publicly stated, "The economic value of Tokens should not be judged solely by unit price. The industry urgently needs a set of criteria to measure 'high-quality Tokens.'" He even bluntly said, "Low-quality Tokens are just noise in computing power, while high-quality Tokens are the credit of intelligence."
"Three Laws": The First Theoretical Attempt in Token Economics
Just as the academic and industrial communities were debating the "value yardstick" of Tokens, A Mu shared with the author the "Three Laws" derived from Lenovo Group's thinking in token economics. When asked why he had such thoughts, A Mu said that one of his most important daily tasks is helping enterprises and clients calculate the "AI account." Over the past six months, almost every entrepreneur who came to talk to him about AI wore the same expression: anxiety.
Token unit price is indeed falling, and falling fast. But when they check their own backend, total AI expenditure has increased tenfold. How is this accounted for? Should AI be invested in? When can it start generating profit?... These questions plague the vast majority of enterprises that want to apply AI but don't know how to proceed.
It was based on these problems in AI application that A Mu proposed a framework he calls an "experimental idea": the "Three Laws" of Token economics. This may be the first time someone in the industry has attempted to summarize the operating laws of token economy in the form of laws.
The first is the Law of Inertia, where the unit Token cost continuously declines. A Mu believes that the unit Token cost will decline continuously and steadily, similar to Moore's Law, but this law has three levels of "inertia."
The first-level inertia is technological innovation in chips, energy, and the model itself. Chip computing power is higher, models achieve higher intelligence levels with the same parameters, and energy consumption is lower. All of these are driving down the unit Token cost.
The second-level inertia is optimization. A Mu stated that by integrating and optimizing "model, compute, and power" holistically, costs can be reduced by up to another 50%. "From hyper-nodes to standard clusters, and then to complete Token factories, optimization at each level is lowering costs," A Mu emphasized.
The third-level inertia is runtime scheduling. In actual usage, intelligent scheduling is used to determine "which intent is distributed to which model, using which GPU, and how to compute," further reducing costs. Lenovo's Token Hub provides exactly this service: multi-model, multi-platform unified access for computing power scheduling and model routing.
A Mu used Lenovo as an example to analyze the layout of the three-level inertia: First level, Lenovo co-develops with domestic and international chip manufacturers to make the design of next-generation GPUs closer to actual application scenarios. Second level, perfecting "back-end engineering" like servers, clusters, and liquid cooling, improving Token output efficiency by over 20% with equivalent computing power. Third level, through a scheduling system called Token Hub, unifying the management of computing power on public clouds, private deployments, and edge devices, ensuring every Token task runs in the most suitable place. The three layers combined can push costs even lower.
The second law is the Law of Acceleration, where the value per unit Token accelerates release. If the first law is about "cost," the second law is about "value." In A Mu's view, the value output per unit Token accelerates due to three factors.
First, the depth of AI integration into workflows. If AI is only used as a Q&A tool by employees, its value is similar to a high-level search engine. However, if AI is embedded into business processes, letting it handle actual work at certain nodes, such as molecular screening, code generation, or automated bid document review... in different scenarios, Token value can differ by 10 times.
Second, the depth of engineering. A Mu gave a detail: Many enterprises now buy AI tools, and employees use them, but the results are average. The reason isn't that AI is poor; it's that the surrounding "engineering" isn't ready, such as unprepared data, unmodified processes, and unrefined agents. He gave an analogy: In the information age, enterprises needed ERP implementation consultants to be stationed for months to get the system running. The AI age is the same, requiring a new role—he calls this the "frontline delivery engineer," who goes deep into the front line to create, embed, and iterate the agent. This process has just begun for most enterprises.
Third, the availability of supporting elements. A Mu broke it down into four dimensions: Are there AI-native talents and organizations? Is there infrastructure like Token factories? Is there a governance system to audit AI ROI, manage knowledge assets, and ensure security? Is there an investment model to calculate the "profit and loss statement" of agents?
The third law is the Law of Singularity. If we plot the cost and value curves of enterprise AI application in the same quadrant. Before a certain "singularity," the cost curve runs above the value curve, and enterprises investing in AI operate at a loss. After passing that "point," the value curve overtakes the cost curve, entering a positive cycle. "Before the singularity, AI helps you save on existing costs. For example, three people with a monthly salary of 30,000 yuan, now using Tokens for 8,000 yuan, saving 22,000 yuan," A Mu further pointed out. "After the singularity, what AI produces is incremental value, things that can help you achieve what was previously impossible."
What is incremental value? It's generating 1 million short video scripts in a day, discovering an effective drug molecule in a year, or enabling someone who hasn't learned programming to create an app. "These are the scaling of innovation," A Mu said. The Industrial Revolution achieved the mass production of industrial goods, the Information Revolution achieved the mass production of data, and the Intelligence Revolution achieves the mass production of innovation.
A Mu even extended his vision further. "If you compare AI to an individual, it is indeed smarter than you. But that's meaningless. How can the intelligence of one person compare to the intellectual network of one million, ten million people working in succession? AIDS research has gone on for over a century, with global scientists passing the baton—this is the real way humans solve major problems. The ultimate value of AI should be benchmarked against such 'human-level challenges.'"
How Enterprises Move from Anxiety to a Positive Cycle
Theoretical discussion aside, real-world Token bills don't wait. In 2026, a problem that caught all enterprises off guard surfaced. According to a report by the Financial Times, the American ride-hailing giant Uber had exhausted its annual AI budget by April 2026, forcing management to cap each employee's monthly usage fee for AI programming tools at $1,500. Internal calculations at Meta showed that, maintaining the current growth rate of employee calls, internal AI usage alone would cost tens of billions of dollars in 2026. Executives at Amazon even publicly warned employees "not to use AI just for the sake of using AI."
This is not an isolated phenomenon. Data from the FinOps Foundation shows that in 2026, AI inference costs accounted for over 80% of enterprise AI total budgets. Tan Dai, President of ByteDance's Volcano Engine, publicly calculated: If an enterprise has 1,000 employees, each calling models 100 times a day, at the then-market average price, the annual Token cost could be as high as tens of millions of RMB. "Many enterprises haven't even calculated this expense, thinking AI is just buying a membership."
Where is the problem? First, the "quality discount" of Tokens. The phenomenon of "secretly reducing precision" mentioned by Huang Wei in the salon is not an isolated case. Token market quotes vary widely: some offer fixed monthly fees but with hidden caps, charging by Token beyond the limit; some charge based on input + output volume but have a minimum billing unit—even if you only input 10 Tokens and output 20, you might be charged for 100 Tokens. It is difficult for users to compare prices horizontally, and budget planning becomes virtually meaningless.
Second, the "Token inflation" caused by agents. Gartner's analysis in March this year showed that Token consumption in agent scenarios is 5 to 30 times that of ordinary conversations. An agent completing a task may trigger 10 to 20 model calls, each call requiring "thinking" through a large chain of thought. You ask AI "help me plan a trip to Yunnan," it might first plan the itinerary, check flights, hotels, attractions... each step is a string of Tokens.
"Many clients come to me saying AI is too expensive," A Mu said. "I ask them what scenarios they are using it for, and they say their employees use it every day to write weekly reports." Writing weekly reports is a relatively low Token value activity. This example is somewhat extreme, but it points to a core issue: There is no inherent good or bad in Tokens; the key lies in where they are used. Used for writing weekly reports, a million Tokens might be worth less than $1; used for drug target discovery, a million Tokens could be worth $1,000, with a cross-scenario Token value difference up to a hundred thousand times.
In response, the China Academy of Information and Communications Technology is promoting a set of "High-Quality Token Service Standards System," constructing a standard framework from four dimensions: service quality, operational capability, production capability, and security capability. Huang Wei proposed a key proposition: Users should have the "right to know," and service providers must disclose what kind of computing power, what version of the model, and what level of precision is used in the backend. "Good computing power can be more expensive; products from two generations ago can be cheaper. But users must know what they are buying."
Explorations on the industry side are more pragmatic. Lenovo launched Token Factory, transforming Token production from a "handicraft workshop" to a "standardized workshop," allowing enterprises to call on demand and pay per Token. The three major telecom operators have successively launched Token packages, selling AI computing power like data plans. Payment giant Stripe spent about $1 billion to acquire Metronome, a startup specializing in Token usage measurement for large model companies like OpenAI and Anthropic. The capital market is already betting on the "Token metering" track.
"When people say Tokens are expensive now, they are comparing them to the electricity bill from a power plant. But only after the power plant is built and electricity arrives do refrigerators, TVs, and air conditioners follow. The real value lies in the refrigerators and TVs, not in the electricity bill," A Mu further pointed out.
Tokens now have prices, bills, laws, and even philosophical questions. But who can surpass that "singularity" depends on who can make every "breath" of AI generate real value.
"The ultimate goal of AI is not layoffs; AI is to scale innovation itself." This goal may sound like a slogan now. But think about how people felt a hundred years ago hearing "every household will have electricity"—it might be similar to how we feel hearing this statement today.
(By Leo Zhang, ToB Casual Talk; Author: Zhang Shenyu; Editor: Yang Lin)
