Alibaba returns to a highly centralized era

Over the past two decades, Alibaba has undergone numerous organizational adjustments.

From the business expansion during the Taobao era, to the launch of the "big middle - platform, small front - end" strategy in 2015, and then to the implementation of the "1+6+N" model in 2023. However, behind these adjustments, there is actually a common thread: decentralizing power.

Decentralization is the most typical organizational logic in the Internet era. It allows more businesses to have independent decision - making power and enables more people to make decisions. Even the "1+6+N" reform is essentially a large - scale power delegation, giving each of the six business groups its own board of directors and independent financing capabilities.

However, an interesting change has emerged today. AI is leading Alibaba back towards centralization.

1. Why has Alibaba suddenly started to centralize power?

In the Internet era, whether it's Taobao, Alipay, Alibaba Cloud, Cainiao, or DingTalk, each business unit of Alibaba has greater independence, and those closest to the customers have the greatest power.

The reason is simple. Internet competition is essentially business competition. E - commerce has its own way of competing, payment has its own approach, and cloud computing has its own strategies.

The greater the autonomy at the front - line, the more vibrant the organization tends to be. Each business needs its own decision - making speed.

However, AI has changed this logic. AI itself is not an independent business; it is becoming the common infrastructure that all businesses rely on. If e - commerce develops its own model, the cloud develops its own model, Quark develops its own model, and DingTalk develops its own model, the result will only be repeated investment of resources, inability to accumulate capabilities, and failure to share data.

The stronger the model capabilities, the more necessary it is to unify. The more expensive the computing power, the more necessary it is to unify. The scarcer the talent, the more necessary it is to unify.

So, in essence, what Alibaba CEO Wu Yongming has been doing in the past three months is solving the same problem: how to reorganize the scattered AI capabilities within Alibaba into a whole.

Step 1: Take back the models.

In March 2026, Alibaba established Alibaba Token Hub (ATH), integrating the Tongyi Laboratory, MaaS business line, Qianwen Division, Wukong Division, and AI Innovation Division.

Tongyi is responsible for model production, MaaS is responsible for model services, and Qianwen, Wukong, and AI Innovation are responsible for application exploration in different directions. The essence of this step is to end the previously relatively scattered AI layout and unify the model capabilities.

Step 2: Gather talents.

In April 2026, Alibaba established a Technology Committee at the group level. Wu Yongming personally serves as the leader, Zhou Jingren serves as the Chief AI Architect, and Wu Zeming and Li Feifei are included in the unified decision - making framework.

Meanwhile, the Tongyi Laboratory was upgraded to the Tongyi Large - Model Division, led by Zhou Jingren. This is a typical gathering of talents. Core AI scientists who were previously scattered in different organizations have been brought into the same strategic system.

Step 3: Take back the products.

In June 2026, Alibaba merged the Tongyi Large - Model Division and the Future Life Laboratory to establish the Token Foundry Division. Zheng Bo led product teams such as Happy Horse and Happy Oyster to be incorporated, and the Token Foundry is directly under the charge of Wu Yongming.

At this point, all products and scenarios that may generate Token consumption have been incorporated into the unified command system directly managed by the CEO.

Taking back models, talents, and products is not just an adjustment of the organizational style; it is a strategic shift.

2. An overlooked fact

There is a very direct background for Alibaba to do this.

One of the biggest misjudgments in the industry in the past year was the assumption that AI competition was still just model competition. However, by 2026, this logic has begun to change.

OpenAI has GPT, Google has Gemini, Meta has Llama, Alibaba has Qwen, Tencent has Hunyuan, and ByteDance has Doubao. The capabilities of basic models are still important, but a key change is that the score gaps of global leading large models on core evaluation sets have significantly narrowed, and the absolute advantage of a single model is disappearing.

What really starts to make a difference is the part outside the model.

Who can turn the model into a product faster, who can make the product generate data faster, who can feed the data back into the model, and who can make the whole system run.

Alibaba's own data also verifies this judgment. In the fourth fiscal quarter of fiscal year 2026, Alibaba Cloud's external commercial revenue increased by 40% year - on - year, and the revenue from AI - related products reached 8.971 billion yuan, maintaining three - digit growth for 11 consecutive fiscal quarters. The AI revenue structure is shifting from self - developed model calls to the MaaS platform and upper - layer AI applications, and the growth engine has undergone a qualitative change.

Today's AI competition is becoming more and more like cloud computing competition. The model is becoming the new electricity, water, and coal. It is of course important, but what really determines competitiveness is not just the power plant, but the entire power transmission system.

Wu Yongming said in the Alibaba earnings conference call in March: The long - term improvement of model capabilities requires the joint efforts of the model end, the application end, and the MaaS end.

This statement reveals Alibaba's judgment. The model cannot grow independently. It must be used, enter scenarios, receive feedback, and then return to the model to form a cycle.

Alibaba's organizational adjustments in the past three months are essentially aimed at making this flywheel spin faster.

3. Alibaba's real intention

Many people may not have seriously thought about why it's called Token Hub and why it's called Token Foundry. In fact, the names say it all.

What is the revenue unit of future AI companies? It's not the number of users, not advertising clicks, not software subscriptions, but Tokens.

OpenAI sells Tokens, Anthropic sells Tokens, and Gemini also sells Tokens. Behind this is a rapidly forming Token economy. Whoever controls the flow of Tokens controls the distribution rights of the next - generation AI value chain.

ATH is responsible for Token circulation, model production, MaaS services, and Token delivery. And Token Foundry is responsible for Token production, to be precise, creating Token consumption. One is responsible for sending out Tokens, and the other is responsible for burning more Tokens. This is the real relationship between the Hub and the Foundry.

So why is Happy Horse important? Video generation is naturally a Token black hole. The amount of Tokens consumed by a video task may be hundreds of times that of a conversation. The Happy Horse - 1.0 model topped the text - to - video and image - to - video tracks on the Artificial Analysis platform, which means it has the ability to output high - quality video content on a large scale, and each video consumes millions of Tokens.

Why is the Future Life Laboratory important? Agent and workflow products are essentially Token - consuming machines. When enterprise - level Agents start to undertake tasks such as automated report generation, multi - round business reasoning, and cross - system operation execution, the amount of Tokens consumed by a single task is dozens of times that of an ordinary conversation.

From this perspective, when looking at this merger again, what Alibaba really cares about is no longer just making a good product. It is clearly creating more Token demand. Alibaba is transforming from a model company into a Token platform.

4. From Model First to Agent First

At the Alibaba Cloud Summit on May 20th, Zhou Jingren, the head of Alibaba's Tongyi Large - Model Division, clearly stated that large models are undergoing a core paradigm shift, from aligning with human preferences to aligning with task goals. In the past, we pursued models that could speak well; now, we require models to be able to achieve results.

In the past two years, the most important capabilities of large models were chatting, answering questions, explaining problems, and writing content. But in the next stage, it's not like that. The most important thing in the next stage is to complete tasks, such as writing code, calling tools, operating software, managing workflows, and achieving a complex goal. This is the Agent.

So, Alibaba's most important breakthrough recently is not in chat capabilities, but in coding capabilities. The code - writing ability of Qwen - 3.7 leads the world, which is essentially paving the way for the Agent era. Code generation is naturally an Agent task. Given a goal, the model plans, executes, and corrects itself until the task is completed.

The same goes for Happy Horse and the Future Life Laboratory. Although they seem to be in different directions, they are actually doing the same thing: transforming the model from a tool for answering questions into a productive force for completing tasks.

So, the real change in this organizational adjustment is not about who reports to whom, but that Alibaba's strategic focus is shifting from Model First to Agent First.

The model is the engine, and the Agent is the whole vehicle. Alibaba doesn't want to just sell engines anymore.

5. Zhou Jingren has been "extracted"

Many people interpret Zhou Jingren's transfer to the position of Chief Scientist of Alibaba and leading the establishment of the AI Future Research Institute as a reward, while others think it's a promotion in name but a demotion in reality. But this is not the most important thing.

More importantly, Alibaba is extracting top - notch scientists from the business system. This is what Google and OpenAI would do.

Zhou Jingren led the establishment of the Tongyi large - model team from scratch, promoted the Qwen series to become a globally leading open - source model, and joined the Alibaba Partnership at the end of 2025. Letting such a model founder break away from daily business management and turn to frontier exploration precisely shows that Alibaba's emphasis on basic research is increasing, not weakening.

Alibaba's strategy is clear. There must be someone responsible for today, and there must also be someone responsible for tomorrow.

This dual - track system is a mature practice in the AI industry. OpenAI has its Chief Scientist Jakub Pachocki leading the frontier research team, focusing on the next - generation model architecture, while an independent product team is responsible for commercialization; Google has merged DeepMind and Google Brain into Google DeepMind to promote research and products in a unified way.

Alibaba's latest architecture also adopts the dual - wheel drive of a frontier research institute and a commercial division, essentially replicating the same model.

Token Foundry is responsible for today, turning AI into a growth engine and creating Token revenue. The AI Future Research Institute is responsible for tomorrow, exploring next - generation model technologies and reserving capabilities for the next - stage competition. This is actually a very classic division of labor, where short - term commercialization and long - term scientific research are starting to separate. Zhou Jingren is not simply getting a promotion; he is being placed in another position.

6. Wu Yongming is becoming Alibaba's AI chief designer

If there is a main thread behind the three reorganizations in three months, it is that AI is changing from a business to the entire operating system of Alibaba.

ATH is directly under the charge of Wu Yongming, Token Foundry is directly under the charge of Wu Yongming, the Technology Committee is led by Wu Yongming, and the AI Future Research Institute reports to Wu Yongming. What does this mean?

It means that AI is no longer just a direction under Alibaba Cloud, nor a division within the group. It has directly become the top - priority project of the CEO. This level of centralization is comparable only to when Jack Ma personally took charge of the "big middle - platform" in 2015 in Alibaba's history. However, the middle - platform at that time supported the business, while today's AI is going to redefine the business itself.

Over the past two decades, Alibaba's growth has come from e - commerce and later from the cloud. In the future, Alibaba hopes that growth will come from AI. Since Wu Yongming returned as CEO in 2023, he has repeatedly emphasized that AI - driven development is the most core strategy for Alibaba in the next decade. From giving up some non - core businesses to scaling back the "1+6+N" spin - off, and then to these three AI organizational reorganizations, he is shifting the company's strategic focus towards AI.

As a result, power is starting to concentrate on AI, the organization is starting to move closer to AI, and resources are starting to tilt towards AI.

This is also why there have been so many adjustments in three months. It's not because Alibaba likes to adjust the organization. In essence, AI has become an irreversible strategy for Alibaba.

[Beyond the Page] Words:

From Taobao to Alipay, from cloud computing to Cainiao, from the "big middle - platform" to the "1+6+N", Alibaba's growth history is almost a history of continuous power delegation. The biggest proposition in the Internet era is efficiency, and efficiency comes from more people making more decisions.

However, new changes have emerged in the AI era: When intelligent capabilities become the underlying infrastructure of a company, power naturally tends to concentrate on those who control the infrastructure.

The most notable thing this time is that Alibaba is starting to answer a question from twenty years ago again: When there is only one most important direction for a company's future, should power be decentralized or centralized?

For the past twenty years, Alibaba's answer has always been the former. And today, Wu Yongming seems to be giving a new answer.

This article is from the WeChat official account "Beyond the Page", written by Huahua and Banjun, and published by 36Kr with authorization.