28 SEP 2026 — Alibaba says it will operate more than 20 gigawatts of data centre capacity worldwide by 2032, and it has announced a new AI chip that goes into mass production in the first quarter of 2027. Both came at its Apsara Conference in Hangzhou on 22 September, alongside a plan for language models of 5 to 10 trillion parameters.
These are the company's own figures. Taken together, they describe an attempt to own every layer of AI in China, from silicon to phone apps.
The capacity target
"Our target is that by 2032, the global data center capacity operated by Alibaba Cloud will surpass 20GW," Eddie Wu, Alibaba Group's chief executive, said, according to the company's announcement.
Alibaba did not say how much capacity it runs today, so there is no way to judge the size of the jump from the announcement.
For Southeast Asia, Alibaba Cloud says it will expand its data centre footprint in Malaysia over the next 12 months, alongside Germany, the United Arab Emirates, France and Hong Kong, and open its first regions in Türkiye, Finland and the Netherlands. It currently runs 107 availability zones in 31 regions.
The chip
T-Head, Alibaba's chip design unit, unveiled the Zhenwu V900, an AI processor for both training and running models. Alibaba says it delivers three times the performance of the Zhenwu M890, which it released in May, with 216 GB of memory and 1,200 GB per second of bandwidth between chips. It supports the low-precision FP8 and FP4 number formats used to cut the cost of running models.
The V900 is scheduled for mass production and commercial release in the first quarter of 2027. Alibaba says its Zhenwu chips already serve more than 650 customers, and that a server built around the V900 can link up to 500,000 chips in a single cluster.
The models
Qwen 4, Alibaba's next foundation model, is in training. The company says the later Qwen 4.5 and Qwen 5 series are projected to reach 5 to 10 trillion parameters, the adjustable values that make up a model. It gave no dates.
Parameter count measures size, not quality. A model that large is expensive to run, and the target says more about ambition than about what the models will do.
Alibaba also described a self-improvement experiment. Over a month of automated runs, Qwen3.8-Max completed 33 cycles and raised its Artificial Analysis score from 40 to 45, according to the company. A separate test pointed the model at chip design: across 60 hours and more than 10,000 calls to design software, Alibaba reports, it produced circuit modules 42% smaller in area with no loss of performance. Neither result has been independently checked.
Phones and agents
For consumers, the nearest product is Qwen Intelligence, a platform that lets phone makers build Qwen-powered agents able to carry out tasks across several apps. Honor is its first partner, Constellation Research reports. Alibaba also announced speech, audio and image models, including a live translation model it says cut delay from 2.8 to 2.3 seconds.
For businesses, the company launched AgentCore, a platform to build, run and monitor AI agents, and an Agent Security Center to police them. It says a new context service cuts token use by up to 67% in tasks such as customer service.
Reading the roadmap
The through-line is self-sufficiency. Chinese companies face US limits on buying the most advanced AI chips, and a home-grown accelerator, CPU, server and model family together reduce that exposure. Alibaba's claims for each layer are its own and mostly forward-looking: the chip ships in 2027, and the largest models have no date.
The Malaysia expansion has the nearest effect in Southeast Asia, and the company has given no size or location for it.
View author profile → · Editorial policy
