Alibaba is expanding its AI infrastructure ambitions far beyond model development. At its Apsara Conference, the company said Alibaba Cloud is targeting more than 20 gigawatts of global data-centre capacity by 2032, putting physical computing infrastructure alongside chips and models at the centre of its AI strategy.
The scale of the target matters because AI infrastructure is increasingly constrained by power, cooling, networking and data-centre availability rather than processors alone. Training and serving large models requires dense computing clusters, and companies are competing for electricity and suitable sites at the same time they compete for GPUs and AI accelerators.
Alibaba’s plan is part of a full-stack approach. The company is developing its own processors through T-Head, building large cloud systems, expanding the Qwen model family and promoting agent-oriented services. The objective is to control more of the infrastructure between the silicon and the application.
Twenty gigawatts is an infrastructure statement
A data-centre target measured in gigawatts is fundamentally a power and capacity commitment. One gigawatt represents a large amount of continuous electrical load, and the actual computing capability delivered by that power depends on cooling systems, server efficiency, networking equipment and the mix of workloads.
Alibaba’s target therefore says as much about its expectations for future AI demand as it does about its cloud business. If model inference becomes a routine component of search, productivity software, commerce, coding and enterprise applications, cloud providers will need much more capacity close to customers.
The company is also positioning its cloud network as an international platform. Alibaba has data centres and availability zones across multiple regions and has continued adding locations as customers seek local data processing and lower latency.
Chips, models and infrastructure are converging
Alibaba’s decision to develop Zhenwu accelerators is directly connected to this expansion. The company has described the Zhenwu V900 as a major performance step over its predecessor and is planning clusters capable of supporting very large AI workloads.
The chip strategy can reduce Alibaba’s dependence on imported accelerators, although it does not eliminate the need for a broad supply chain. AI data centres still require high-bandwidth memory, networking, storage, power equipment and advanced manufacturing capacity.
On the model side, Alibaba is preparing much larger Qwen systems. Reuters reported that the company is developing a model with between 5 trillion and 10 trillion parameters, although parameter count alone does not determine a model’s usefulness or operating cost.
The economics will be harder than the headline
Building 20GW of capacity is a long-term capital commitment. Data centres require land, grid connections, cooling infrastructure, network connectivity and equipment, and AI clusters have relatively short technology cycles compared with buildings that may operate for decades.
That creates a utilisation problem. A facility built for one generation of accelerators must remain economically useful when newer processors arrive. Cloud providers therefore need flexible infrastructure that can accept different generations of compute and support both training and inference.
Alibaba’s advantage is that it can coordinate multiple layers of the stack. Its cloud business already has customers, its semiconductor unit can design specialised processors and its model teams can optimise software for the hardware. The challenge is making those pieces competitive on cost and reliability while expanding internationally.
The 20GW target is consequently less about one data-centre project than a statement about the scale Alibaba expects AI computing to reach. Whether the company reaches that figure will depend on capital spending, power availability, customer demand and the economics of operating increasingly dense AI clusters.