On the same day NVIDIA defined token economics from the supply side, Alibaba reorganized its entire corporate structure around the token from the demand side — consolidating all AI operations into a single entity called Alibaba Token Hub, led directly by the CEO. When a two-hundred-billion-dollar company restructures its org chart around a compute primitive, the primitive has become currency.
Alibaba Group announced on March 16 that it is consolidating all of its artificial intelligence operations into a single new business group called Alibaba Token Hub. The entity sits at the same organizational level as Alibaba Cloud and the company's e-commerce divisions. CEO Eddie Wu Yongming leads it directly.
The name is not metaphorical. The token — the fundamental unit of computation generated when a large language model processes or produces text, code, or images — is the organizing principle around which five previously separate divisions have been combined. Research, consumer products, enterprise platforms, smart glasses, and model infrastructure now report through one structure, unified by a single idea: the token is where the value lives.
The Memo
Eddie Wu's internal letter announcing the restructuring reduced the entire AI mission to three verbs applied to one noun: create tokens, deliver tokens, apply tokens.
The compression is the point. When a CEO of a company valued at over two hundred billion dollars distills the AI strategy into a token lifecycle — generation, distribution, application — he is not simplifying for communication. He is revealing the unit of analysis that the organization now optimizes for. Every division must justify its existence by its relationship to the token. Research creates tokens. Infrastructure delivers tokens. Products apply tokens. Divisions that cannot articulate their contribution to this lifecycle have no structural home.
Wu described the industry as approaching an inflection point in artificial general intelligence and committed to leading the new group directly — not delegating it to a subordinate. When a CEO takes personal operational responsibility for a division, the division is either in crisis or at the center of the company's future. In this case, it is both.
The Consolidation
Five units were absorbed into Alibaba Token Hub.
Tongyi Laboratory — the research institution that develops Alibaba's Qwen series of foundation models — continues under CTO Zhou Jingren but now reports through ATH rather than through Alibaba Cloud. The MaaS business line, providing model-as-a-service technical infrastructure, joins as the delivery layer. The Qwen business unit, which operates the consumer-facing Qwen personal assistant app — an application that reached one hundred million monthly active users within roughly two months of its November 2025 launch — represents the application layer, led by Vice President Wu Jia.
A newly created Wukong business unit consolidates DingTalk, Alibaba's workplace communication platform, with Quark-branded devices including smart glasses, under the mandate of building an AI-native enterprise work platform. An AI Innovation business unit rounds out the structure, dedicated to rapid exploration of new AI applications and business models.
The previous organizational structure distributed these units across different reporting chains. Tongyi Lab sat within Alibaba Cloud. Qwen had been spun out as a standalone business group. DingTalk operated independently. The token was already the unit of billing across all of them. What changed is that the token is now the unit of organization.
The financial commitment behind the restructuring is substantial. Alibaba pledged three hundred and eighty billion renminbi — approximately fifty-three billion dollars — to AI and cloud infrastructure over three years, exceeding the company's total spending in these categories over the prior decade. The company has open-sourced over four hundred models, ranging from five hundred million to two hundred thirty-five billion parameters. AI products already account for more than twenty percent of Alibaba Cloud's external customer revenue.
The Fracture
The restructuring did not emerge from a position of strength. Three senior AI executives departed in ten weeks.
In January, Hui Binyuan — the staff research scientist who led Qwen Code, Alibaba's AI coding capabilities — left for Meta. On March 3, Lin Junyang, the technical lead of the entire Qwen project and one of the company's youngest senior executives, stepped down. His post on X was four words: me stepping down. bye my beloved qwen. A colleague commented publicly: I know leaving wasn't your choice. The same week, Yu Bowen, head of post-training for Qwen models — the work of alignment, instruction tuning, and reinforcement learning from human feedback — also departed.
A ByteDance AI executive characterized Lin Junyang as at least a hundred-million-dollar talent. Several younger researchers submitted resignation letters the same day he left. One employee echoed the language from OpenAI's 2024 board crisis: Qwen is nothing without its people.
The departures were not coincidental. Tongyi Lab had been restructuring the Qwen team from a vertically integrated unit — where a single team handled pre-training, post-training, language, multimodal, and code together — into a horizontally organized system with separate functional teams. Lin had advocated for the opposite direction: merging language, image, video, and code teams into a unified multimodal effort. His management scope was being reduced. The tension between vertical integration and horizontal specialization — the same structural question that every technology company faces when a product team reaches a certain scale — resolved in favor of horizontal. The vertical leaders left.
Eddie Wu convened an emergency all-hands meeting at Tongyi Lab the day after Lin's departure. Chief People Officer Jiang Fang acknowledged insufficient internal communication about the restructuring. Hao Zhou, recruited from Google DeepMind — where he had contributed to Gemini 3.0 and multi-step reinforcement learning — was brought in to lead post-training. The organizational wound was treated. Whether it has healed is a question that Alibaba's fiscal third-quarter earnings, scheduled for March 19, may begin to answer.
The Precedent
When a company reorganizes its structure around a single unit of value, the unit has graduated from a metric to a currency.
Google's transformation in the mid-2000s is the clearest parallel. The company had search, advertising, YouTube, Android, Gmail, and Maps — each a separate product with its own logic. The structural unification was the click. Every product existed to generate clicks. Every click existed to be monetized through advertising. The entire organizational architecture — engineering, sales, business development — was oriented around generating, routing, and pricing clicks. The click was not the product. The click was the denomination. Google did not sell search. It minted clicks.
Netflix followed a similar arc a decade later with the subscriber. Content, technology, marketing, international expansion — every decision was evaluated against its impact on subscriber count. A hundred-million-dollar original series was not a content investment. It was a subscriber acquisition cost. The subscriber was not a customer. The subscriber was the unit that the entire organization optimized for, reported on, and structured itself around.
Alibaba has now done the same with the token. Create tokens — research and pre-training. Deliver tokens — model-as-a-service infrastructure. Apply tokens — consumer products, enterprise platforms, smart glasses. The token is not the product. The Qwen app, DingTalk, Quark glasses — those are products. The token is the denomination that makes them commensurable. It is the unit that allows a CEO to evaluate whether a research lab, a messaging platform, and a pair of smart glasses belong in the same division. They do, because each one creates, delivers, or applies the same primitive.
The Timing
The same day Alibaba announced Token Hub, Jensen Huang stood on a stage forty-five hundred miles away at GTC 2026 in San Jose and spent two hours defining the token economy from the supply side.
The Definition documented the keynote: NVFP4, a four-bit numerical format that delivers fifty petaflops of inference compute. Vera Rubin, three hundred thirty-six billion transistors optimized for token generation. Groq's inference architecture, integrated to capture the operating expenditure half of the compute economy. Every announcement was about making tokens cheaper, faster, and more abundant.
NVIDIA defines how tokens are manufactured — the silicon, the format, the interconnects, the energy. Alibaba Token Hub defines how tokens are organized — the research that creates them, the infrastructure that delivers them, the products that apply them. One is the supply side of token economics. The other is the demand side. Both announced on the same day.
When both sides of a market independently reorganize around the same primitive on the same day, the primitive has stopped being a technical detail. The click was a technical detail in 1998. By 2005, it was the unit that structured a three-hundred-billion-dollar company. The subscriber was a technical detail in 2007. By 2013, it was the unit that structured a hundred-billion-dollar company. The token is a technical detail in the way that the byte was a technical detail in 1975 — a low-level abstraction that will become the organizing principle of the next generation of corporate value.
Alibaba's stock rose approximately three percent in premarket trading. The market's response to a two-hundred-billion-dollar corporate restructuring was roughly the same as its response to a two-percent chip announcement. Both were confirmation, not surprise. The token was already the unit. The org chart is just catching up.
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)