Alibaba has laid out one of its clearest AI infrastructure roadmaps yet. CEO Eddie Wu says the company plans to train a next-generation model with between five and ten trillion parameters, while Alibaba Cloud is targeting more than twenty gigawatts of data-centre capacity by 2032. At the same event, Alibaba’s T-Head chip unit unveiled the Zhenwu V900 accelerator. Those are significant targets, but they are still targets, and the distinction between an announced roadmap and delivered capacity matters.

The model plan is for a system in the five-to-ten-trillion-parameter range, succeeding Alibaba’s current generation. That does not mean a ten-trillion-parameter model exists today. It is a development goal. On infrastructure, Eddie Wu said Alibaba Cloud aims for global data-centre capacity to exceed twenty gigawatts by 2032. The company is therefore describing expansion across both the software and physical sides of AI.

Alibaba also introduced the Zhenwu V900 AI accelerator through its T-Head semiconductor unit. Alibaba says the chip delivers three times the performance of its predecessor and expects mass production in early 2027. That performance figure is a company claim rather than an independent benchmark result, so the useful milestones will be production volume, deployment and measured performance once the hardware is operating at scale.

Taken together, the announcements show Alibaba trying to coordinate models, accelerators and cloud infrastructure rather than treating them as separate projects. Greater control of those layers could give the company more flexibility over how it deploys AI workloads. But it would be premature to call that compute independence. Alibaba still has to manufacture hardware, expand data-centre capacity and work through supply constraints.

Alibaba has acknowledged that supply-chain bottlenecks remain a constraint. That is important because a roadmap measured in gigawatts depends on far more than a model specification. The announced chip has to reach mass production, infrastructure has to be built and equipped, and the capacity target has to translate into usable computing resources. The scale of the plan is clear; the pace and economics of delivery are not. The company’s stated ambition therefore needs to be read alongside those dependencies. A large parameter count can describe model scale, while gigawatts describe infrastructure capacity; neither measure on its own tells us how capable, efficient or commercially useful the resulting services will be.

The next useful evidence will be concrete rather than promotional: whether Zhenwu V900 reaches mass production on schedule, how widely it is deployed, what independent performance data emerges, and whether Alibaba reports measurable progress toward its 2032 capacity target. Those milestones will show how much of today’s vertically integrated roadmap becomes operating infrastructure.

For now, the confirmed story is straightforward. Alibaba is expanding its AI ambitions across models, custom silicon and cloud capacity at the same time. The strategy is notable, but the model size, chip performance and twenty-gigawatt target should be judged as separate claims and milestones as evidence arrives.