Alibaba Recentralizes Its AI Org Around a "Token Economy" Bet

Alibaba has reversed a structural philosophy it held for most of its existence. The company that famously split itself into semi-autonomous business units, and that in 2023 announced a six-way breakup into independently governed groups, is now consolidating its AI work under a single executive. Over roughly three months in early 2026, CEO Wu Yongming executed three consecutive consolidation moves that pulled model development, AI research talent, and product ownership into one reporting structure that terminates at his desk.

The framing Alibaba has chosen for this is the "token economy," and the new division names lean into it heavily: the Alibaba Token Hub and the Token Foundry. The vocabulary is doing a lot of work here, so it helps to separate what actually changed from how the company is describing it.

What actually happened

The sequence is concrete. In March 2026, Alibaba created the Alibaba Token Hub (ATH) by merging five units that had operated separately: Tongyi Lab, the Model-as-a-Service (MaaS) business line, the Qianwen division, the Wukong division, and the AI Innovation division. The division of labor inside ATH maps cleanly onto a production pipeline. Tongyi Lab builds the models. MaaS packages and delivers them as services. Qianwen, Wukong, and AI Innovation are tasked with finding application directions, meaning the products that sit on top of the models.

In April, Alibaba stood up a group-level Technology Committee chaired by Wu, with Zhou Jingren installed as Chief AI Architect. The stated purpose was to gather AI scientists who had been distributed across different org units into one strategic framework. Tongyi Lab was promoted to a full Large Model Business Division in the same period.

In June, the Tongyi LLM Division and a unit called Future Life Lab were merged into the new Token Foundry division. Product teams with the internal codenames Happy Horse and Happy Oyster were folded into Token Foundry, which reports directly to Wu.

Strip away the naming and the pattern is a textbook recentralization: research, infrastructure, and products that previously answered to different leaders now answer to one.

What is genuinely new

The interesting part is not that a large company reorganized. It is the specific economic argument Alibaba is using to justify reversing its own long-standing structure.

The financial backdrop is real. Alibaba Cloud's external commercial revenue grew 40% year-over-year in the most recent quarter, and the company reports AI-related product revenue of 8.971 billion RMB with triple-digit growth sustained for 11 consecutive quarters. Eleven quarters of triple-digit growth is the kind of number that reorganizations get built around, because it suggests the bottleneck is no longer whether the models are good enough to sell, but how fast the company can convert model capability into products that generate usage.

Alibaba's internal thesis is that the competitive frontier has moved from raw model quality to the speed of a loop: models become products, products generate usage data, and that data feeds back to improve the models. This is the flywheel argument, and it is not unique to Alibaba. It is essentially the same logic OpenAI, Google, and Anthropic use to justify shipping consumer and API products aggressively rather than sitting on research. What is distinctive is Alibaba choosing to express it through organizational structure, betting that a unified command chain closes the loop faster than coordinating across independent units.

The naming reveals the mechanism Alibaba cares about. Token Hub is positioned around token circulation: producing models, serving them, and delivering tokens. Token Foundry is positioned around token consumption: building products and scenarios that burn large volumes of tokens. The Happy Horse video generation model is described as a deliberate "token sink," and that label is technically honest. Video generation does consume orders of magnitude more compute and output tokens per task than text chat. A single short generated clip can require the kind of sustained inference that a conversational exchange never touches. If your business is selling inference capacity through a cloud platform, products that consume tokens heavily are products that drive cloud revenue.

What to be skeptical about

The phrase "transforming from a model company into a token platform" is the part that deserves a careful read. "Token" here means inference units consumed and billed, not anything resembling a cryptocurrency or a tradable asset, and the press framing blurs that line in a way that sounds grander than the underlying business. The actual claim reduces to something familiar: Alibaba wants to sell more cloud inference, and it is building first-party products specifically to consume that inference at scale. That is a reasonable strategy, but it is a cloud-economics play dressed in new terminology, not a new category of company.

There is also a tension the announcement does not resolve. Alibaba split itself apart in 2023 with the argument that smaller, independent units would move faster and respond to markets better. Recentralizing AI three years later is an implicit admission that decentralization slowed the parts of the company that now matter most. Both can be defended, but a company cannot claim both that autonomy drives speed and that consolidation drives speed without acknowledging that the right structure depends on what you are optimizing for. For AI in 2026, Alibaba has decided the answer is central control. Whether that survives contact with the next strategy cycle is an open question, and the company's own recent history suggests these structures are not permanent.

The deeper bet is that controlling token flow means controlling the value chain of next-generation AI. That holds only if inference volume remains the scarce, monetizable resource. If model efficiency keeps improving and the cost per token keeps falling, which has been the consistent direction for the past two years, then a strategy built on maximizing token consumption could find its margins compressing even as usage grows. A token sink is a revenue source only as long as tokens are priced to matter.

Why it matters

For anyone tracking the Chinese AI market, the Tongyi and Qwen model families have been among the more credible open-weight releases, with Qwen models showing up frequently in third-party benchmarks and fine-tuning workflows. The reorganization signals that Alibaba intends to treat those models less as research outputs and more as the feedstock for a vertically integrated product-and-cloud machine. That likely means faster product releases, tighter coupling between Alibaba Cloud and the models, and more first-party applications designed to keep usage inside Alibaba's own infrastructure.

For the broader market, it is another data point in a clear trend: the major AI players are converging on the same conclusion, that owning the full path from research to deployment to data feedback is the position worth fighting for. Alibaba is making that bet through org-chart surgery rather than acquisitions, which is cheaper and faster but depends entirely on whether one executive can actually run a research lab, an infrastructure business, and a product portfolio as a single coordinated system. The numbers give the strategy room to run. The execution is the part no reorganization announcement can prove in advance.