As artificial intelligence large models accelerate their integration into everyday applications, the era of the "token" has arrived, triggering a wave of development across the nation—from token factories to token supermarkets, from data annotation bases to AI model workshops, a diverse range of new business forms is emerging. In June of this year, Shaanxi Province launched the first token trading platform in western China—the Shaanxi Silk Road Token Trading Platform. After two months of operation, what results have been achieved? How many steps are involved in taking a token from production to transaction? And as the token consumption era dawns, what will be the next major focal point for competition?
When you hear the phrase "trading platform," your mind might conjure up images of a noisy hall, bustling crowds, and the incessant ringing of telephones. But at the Shaanxi Silk Road Token Trading Platform, none of these familiar scenes exist. In reality, the token trading platform has no physical hall at all. Shi Kui, Executive General Manager of the Shaanxi Silk Road Data Trading Center, offered an analogy—it functions much like the e-commerce platforms we commonly use, such as Taobao or JD.com.
Shi Kui explained that the token trading platform is not a physical office space, but rather an online, digitized trading system. Just like e-commerce platforms, users browse, test, place orders, call upon, and purchase token products entirely online, as if shopping in a virtual supermarket. The key difference lies in what's being traded: e-commerce deals in clothes, appliances, and daily goods, whereas the token trading platform facilitates the exchange of data products and AI capabilities measured in "tokens."
A token is a unit of data measurement—simply put, the computational unit consumed by a piece of text, a report, or a Q&A session can all be quantified in tokens. Shi Kui describes this model as a "front shop, back factory" structure. The "front shop" is the token trading platform, operational since June, serving as the online supermarket. The "back factory" is the token factory still under construction, responsible for the continuous production of "goods."
Shi Kui elaborated that the "front shop" is the platform itself, akin to a store or supermarket for token transactions in the open market. The "back factory" supports the "front shop" through a self-operated model, acting as the true token processing plant. It focuses on data processing, data annotation, model fine-tuning, model integration, and the development of intelligent agents, alongside providing a suite of data API services to ensure a steady supply of products on the shelves.
So, what exactly is stocked on this "supermarket's" shelves? In the two months since its launch, the platform has already begun signing performance contracts. Shi Kui introduced that the platform offers two main categories of token capabilities: one is general-purpose token products, and the other is application capability products. For example, by feeding decades of desensitized diagnostic data from a top-tier hospital into a model, it can be fine-tuned into an AI diagnostic assistant. When a user needs it, they can simply call it with one click, and the system will quickly return professional results.
Shi Kui explained that one category is general-purpose token trading products, which include the widely used models like DeepSeek, Qwen, and Doubao. The second category is application capability token products, which are more focused on specific industries, particularly for "one-person companies" (OPC). These businesses require digital employees to handle services such as finance, accounting, legal, and administrative tasks, helping these small and micro enterprises reduce operational costs.
What is the specific process for a business or developer looking to purchase tokens on the platform? Shi Kui detailed that users register on the platform, complete real-name verification, and can then make purchases, much like buying goods on any other e-commerce site. The platform offers APIs for general-purpose tokens. Once integrated, every successful call to a model or intelligent agent is metered and settled based on the unit price. Users can clearly see from their bills how many tokens have been consumed by each model or agent, ensuring complete transparency in the consumption process.
How is pricing determined? This is also a key concern in the token consumption era. Shi Kui noted that tokens are priced per million units. The cost for calling a general large model is only a few cents per million tokens, making it extremely affordable at present. The platform's billing methods are also flexible and diverse: one is a prepaid model where users top up their account and deduct from the balance; another is a subscription model with annual or monthly payment options.
Shaanxi's exploration is just one step in the national strategy for the token consumption era. In recent months, there has been a flurry of policy signals in the token field. In March of this year, the National Data Administration officially designated the Chinese translation of "Token" as "词元" (Ci Yuan), positioning it as the "value anchor of the intelligent era." In July, Yu Ying, Deputy Director of the National Data Administration, explicitly encouraged business model innovation based on token applications and explored new trading models like token exchanges. From the central government to local authorities, the policy framework for tokens is being rapidly established.
Where will the next competitive battleground be? In Wuxi, Jiangsu Province, the country's first city-level "Token Supermarket" was launched in May this year. Over 30 mainstream large models are now available for businesses to choose from, just like products on a supermarket shelf. Enterprises only need to register one account to access all models in a one-stop manner, eliminating the need for separate negotiations. Wuxi's "Token Supermarket" also adopts a pay-as-you-go billing model, significantly lowering the barrier to entry for small and micro enterprises. Since its operation began in May, the "Token Supermarket" has served over 50 companies in Wuxi.
If the "Token Supermarket" is a "convenience store" for the application layer, then the integrated computing network connects the dispersed computing resources nationwide. In Shaoguan, Guangdong Province, the "East Data, West Computing" hub node for the Guangdong-Hong Kong-Macao Greater Bay Area is transforming into a massive "computing power factory." Rows of server racks serve as production lines, and the chips within them are the ever-operating "digital workers." It is this invisible computing power that supports over 140 trillion token calls daily across the nation.
From Shaanxi's "front shop, back factory" to Wuxi's "Token Supermarket," from Shaoguan's computing infrastructure to the data and token factories springing up nationwide, various entities are actively positioning themselves. But what exactly is the relationship between these concepts—token factories, data factories, and token trading platforms? Zhang Xianghong, a member of the National Data Expert Advisory Committee and a professor at Beijing Jiaotong University, explains that a token factory is a new type of production operation that uses data as a key production factor, intelligent agents as core production tools, and computing power as new infrastructure. It encapsulates data, algorithms, and computing power into tokens as a unified intelligent capability. The token factory is composed of data factories, intelligent agent factories, and computing power factories.
High-quality datasets are the raw material foundation for tokens. A data factory is the basic operation for large-scale production of high-quality datasets, consisting of stages such as a reserve workshop, a production workshop, and a pilot workshop. An intelligent agent factory is the basic operation for large-scale production of intelligent agents, involving steps like scenario selection, task creation, model selection, agent production, agent tuning, and agent pilot testing. A computing power factory is the operation for large-scale production, monitoring, and scheduling of computing power, composed of computing centers, computing clusters, and a national integrated computing network.
As the token consumption era arrives and various regions make their moves, what pressing issues need to be addressed? Zhang Xianghong analyzes that the "measuring stick" for token usage is not yet standardized. He points out that during the training phase and the inference input phase, tokens are the smallest language unit and the basis for input billing in large models. However, different manufacturers currently use different tokenization techniques and rules, resulting in different token counts derived from datasets of the same size. Tokens are also the basis for billing user output. Yet, different models from different manufacturers, and even different models from the same manufacturer, employ different inference paths and chain-of-thought lengths. This leads to significant discrepancies in token consumption when solving the same problem across different vendors and models. Additionally, issues of inconsistent metering and opaque pricing exist. Unless these shortcomings are resolved, the large-scale development of token consumption cannot be realized.
Experts point out that the next competitive focus is not on building more platforms or factories, but on achieving breakthroughs in foundational systems such as standardizing tokens, standardizing measurement, and ensuring pricing transparency. Zhang Xianghong notes that in the current pricing system, token prices are essentially unilaterally set by leading large model companies. Algorithms are a core barrier; companies do not open their data or accept third-party monitoring, making it impossible for external parties to measure, verify pricing, or conduct audits. Regarding measurement units, data of the same scale can yield different token counts depending on the tokenization method used. The amount of computing power that one kilowatt-hour of electricity can convert is not fixed, the number of tokens that equal computing power can generate is not fixed, and the value of an equal number of tokens across different large models is not fixed. To date, there is no standardized token measurement unit comparable to one kilowatt-hour of electricity, one ton of water, one liter of oil, or one gigabyte of data traffic.
The curtain has risen on the token era, but the journey from "usable" to "user-friendly," and from "having a market" to "having a price," is still a long one.