On March 16, Alibaba announced internally the establishment of the Alibaba Token Hub (ATH) Business Group, directly overseen by Alibaba CEO Eddie Wu. The new business group integrates Tongyi Labs, MaaS Business Line, Qianwen Business Unit, Wukong Business Unit, and the AI Innovation Business Unit, with a core objective: to rebuild Alibaba's AI business system around the "creation, delivery, and application of Tokens," forming a complete AI production system.

In the mobile internet era, the core of platform competition was traffic access. However, in the era of large models, more and more companies are realizing that what truly determines the scale of AI commercialization is not just model parameters and technical indicators, but rather the production and consumption capacity of Tokens.

From the author's perspective, Alibaba's establishment of ATH may appear to be an integration of AI departments, but at its core, it is reconstructing a new AI production system: from model R&D to platform distribution, and then to application deployment, enabling Tokens to be continuously generated, circulated, and consumed within the system. As AI Agents begin to enter real business processes, this Token-centric system is likely to become the new infrastructure of the AI era.

01: AI Competition is Shifting from 'Model Capabilities' to 'Token Consumption'

In large model systems, Tokens are the smallest units of information processed by the model. Both user-input text and model-generated content are essentially broken down into Tokens for computation.

A common estimation method in the industry is: 1 Chinese character ≈ 1 Token.

Whether it's API call pricing, model inference costs, or computational resource consumption, most AI products use Tokens as their unit of measurement. In other words, while large models may appear to compete on parameter scale, model capabilities, and technical metrics, when it comes to the commercial level, the key determinant of revenue scale is often just one indicator: Token consumption.

The reason is simple: every model call, every content generation, and every inference computation consumes Tokens. The more calls made, the greater the Token consumption, and the higher the corresponding computational resource demand and commercial revenue.

This demand has been noticeably accelerating over the past year.

Internally at Alibaba, the Coding Plan Tokens launched during the Spring Festival became one of the fastest-selling products in Alibaba Cloud's history. Due to overwhelming purchase demand, the originally designed "first-purchase discount" had to be discontinued just two weeks after launch. This phenomenon reflects, to a certain extent, the rapidly rising demand from enterprises and developers for AI computational resources and model calls.

An even greater change is coming from AI Agents.

As large model capabilities continue to improve, more and more AI applications are no longer just answering questions but are beginning to directly execute tasks, such as writing code, organizing documents, handling customer service inquiries, and even automatically completing parts of business processes. The industry generally refers to such systems capable of autonomously executing tasks as AI Agents.

Eddie Wu stated bluntly in his internal memo: A vast amount of digital work will be supported by tens of billions of AI Agents, and these AI Agents will rely on Tokens generated by models to operate, becoming the primary carriers of human-digital world interaction.

If this trend holds, the commercial logic of the AI industry will also change accordingly.

In the past, the focus of competition among large model companies was on model capabilities, such as parameter scale, training data, and evaluation scores. However, as application scale continues to expand, what will truly determine commercial scale is how many times the model is called and how many Tokens are consumed.

Peng Deyu, a well-known tech industry commentator, analyzed in discussions with the author that from this perspective, Tokens are no longer just a technical metric but will gradually become a fundamental production factor in the AI era. Just like computational power in the cloud computing era or energy in the electrical age, whoever can produce, distribute, and consume more Tokens is more likely to gain an advantage in the AI industry.

This is precisely why Alibaba named its new business group Alibaba Token Hub.

The name itself reflects the company's judgment: in the AI era, what truly matters is not just the models but an entire production and distribution system centered around Tokens. Tokens are likely to become the "utilities" of the AI industry.

02: ATH is Essentially an 'AI Productivity Supply Chain'

Centered around the core concept of Tokens, Alibaba has redesigned its AI organizational structure. The newly established Alibaba Token Hub (ATH) Business Group integrates nearly all of Alibaba's key AI capabilities, forming a new business system.

According to the internal memo, ATH currently consists of five core departments: Tongyi Labs: Responsible for the overall R&D of the Qwen large model, creating leading multimodal models, continuously pushing the limits of foundational model capabilities, and providing the most advanced models for the group and the industry; MaaS Business Line: Upgraded from MaaS, it builds an efficient and open model service platform and technical system to support the AI ecosystem across the industry; Qianwen Business Unit: Aims to create the best personal AI assistant; Wukong Business Unit: Set to become DingTalk's most important future business, it builds a B-end AI-native work platform, deeply integrating model capabilities into enterprise workflows; AI Innovation Business Unit: Explores various AI innovation applications and rapidly validates new models and markets.

On the surface, this is merely a departmental integration. However, from a supply chain perspective, Alibaba is actually constructing a complete AI production system.

This system can be broken down into three stages.

The first stage is Token creation.

This is primarily the responsibility of Tongyi Labs. Tongyi Labs handles the R&D of the Qwen series of large models, including continuous upgrades to multimodal model capabilities. The stronger the model capabilities, the larger the scale of Tokens that can be generated and processed, determining the technological ceiling of the entire AI system. In other words, Tongyi Labs is the source of the entire system.

The second stage is Token delivery.

This is mainly undertaken by the MaaS Business Line. MaaS is essentially Alibaba's large model service platform for enterprises and developers. Through API interfaces, development tools, and platform services, it opens up Qwen's model capabilities to enterprise clients. Enterprises can rapidly develop AI applications on this platform, and model capabilities are distributed to various industry scenarios through the platform.

The third stage is Token application.

The application layer mainly includes Qianwen, Wukong, and the AI Innovation Business Unit. These products are AI applications directly oriented towards (facing) users and are where Tokens are heavily consumed. Among them, the Qianwen Business Unit primarily targets C-end users, aiming to create personal AI assistants and accumulate user scale through features like chatting, searching, and content creation.

Meanwhile, the Wukong Business Unit is more oriented toward the enterprise market, positioned as a B-end AI-native work platform. Its core mission is to embed model capabilities into enterprise office systems and business processes, enabling AI to directly participate in daily enterprise operations.

To understand this system more intuitively: Tongyi Labs is like a "power plant" responsible for producing energy; the MaaS platform is like a "power grid" responsible for delivering energy to various locations; and Qianwen and Wukong are like "electrical devices" that convert this energy into specific application scenarios.

The "energy" flowing through this system is Tokens.

Through this structure, Alibaba reconnects its previously dispersed large model R&D, platform capabilities, and application products, forming a complete chain from technological origins to application deployment.

To some extent, this is also Alibaba's first attempt to organize its AI business in a manner closer to an industrial system, no longer just focusing on individual models or products but building an entire production and distribution system centered around Tokens that can operate sustainably.

03: The Real Issue This Adjustment Aims to Solve is 'AI Fragmentation'

In terms of technical strength alone, Alibaba has consistently been in the first tier of China's large model industry.

Over the past two years, the Qwen series of models has maintained high popularity in the open-source community. On platforms like HuggingFace, Qwen models have consistently ranked among the top in download counts and have performed on par with or even surpassed some overseas models of similar scale in multiple evaluation benchmarks. This means that Alibaba is not lacking in technical accumulation when it comes to foundational model capabilities.

However, the problem lies not in technology but in organizational structure.

Over the past few years, Alibaba's AI capabilities have actually been dispersed across multiple systems. For example: Tongyi Labs was responsible for large model R&D; Alibaba Cloud handled computational resources and model service platforms; DingTalk explored enterprise-grade AI applications; and business lines like Taobao and Tmall were also developing their own AI tools.

These teams advanced products and technologies independently. In the early stages of AI's explosion, this model was not necessarily a bad thing. Exploring multiple technical routes simultaneously could help the company identify potential opportunities more quickly.

However, as AI competition continues to escalate, the issues with this dispersed model have gradually emerged.

Internally, dispersion means resources are prone to redundant investment. Different teams may repeatedly develop similar functionalities for the same issue, increasing collaboration costs.

Externally, product perception has become increasingly blurred.

If users are asked, "What AI do you use?" many will directly answer: Doubao, Kimi, or DeepSeek. However, when it comes to Alibaba's AI, the situation often becomes complicated.

Is it Tongyi Qianwen? Or the Qianwen App? Or the AI assistant in Quark?

The same company, the same set of technologies, yet multiple different product forms emerge, making it difficult for ordinary users to form a clear perception.

As the industry enters a new phase, this issue has become more pronounced.

In the past, the large model industry was primarily in the technological exploration phase, with companies focusing more on model capabilities, such as parameter scale, training data, and evaluation scores. However, as computational resource investment continues to increase and large model R&D costs rise, AI competition has gradually shifted from pure technological competition to a comprehensive competition involving computational power, data, and application scale.

In such an environment, excessive resource dispersion means decreased efficiency and difficulty in concentrating strategic resources.

Google encountered a similar issue.

In 2023, Google merged Google Brain and DeepMind into the new Google DeepMind, with the core objective of reducing internal resource dispersion and aligning research, products, and technical routes more uniformly.

Alibaba's establishment of ATH appears to have similar considerations from a logical standpoint.

Through the new business group structure, large model R&D, cloud platform capabilities, and application products previously dispersed across different departments have been reintegrated into a single system, directly overseen by the group's CEO. This approach reduces cross-departmental collaboration costs and allows for more concentrated resource allocation.

Peng Deyu pointed out: In other words, Alibaba is no longer just advancing multiple AI products simultaneously but is attempting to reorganize these capabilities to form a complete system from model R&D to platform distribution and application deployment.

Therefore, the author believes: From this perspective, ATH is not just an organizational adjustment but also represents Alibaba's strategic shift in the AI era: ATH connects model R&D, platform capabilities, and application products into a closed loop, enabling Tokens to flow and be consumed within the system.

The bullets are now in the gun. What truly matters next is API call volume, user scale, and enterprise adoption speed. In the AI era, Tokens are the "electricity" driving the entire system, and whoever controls Token flow will control the initiative in future competition.

The cover image and illustrations in this article are sourced from the internet, with copyright belonging to their respective owners. This article does not constitute any investment advice.