Alibaba introduces Zhenwu V900 as its AI infrastructure ambitions expand

A chip announcement with a system behind it

Alibaba's T-Head unveiled the Zhenwu V900 at the Apsara Conference in Hangzhou, according to TechNode's September 22 report. The accelerator targets training and inference, with mass production and sales scheduled for the first quarter of 2027. That timetable makes this a product announcement rather than evidence of a broadly available V900 cloud service today.

T-Head describes the V900 as three times faster than the preceding M890. It also specifies 216GB of memory, 1,200GB/s of inter-chip bandwidth and native FP8 and FP4 support. The performance comparison is a vendor claim; the report does not provide a reproducible application benchmark showing the same improvement across different models.

Why memory and interconnect matter

The accompanying server design brings together V900 accelerators, ICN Switch interconnects, Panmai networking and Zhenyue storage controllers. T-Head says more than 1,000 accelerators can operate as one system, while its larger cluster architecture is designed to scale much further.

These are different levels of a computer. Memory on an accelerator determines how much model state it can keep close to computation. An interconnect determines how quickly it can exchange information with other chips. A network connecting many servers adds another layer. A large figure at one layer does not establish the performance of the complete application.

The older supernode and the next cloud buildout

In the published transcript of his Apsara keynote, Alibaba CEO Eddie Wu says M890-based supernodes already support inference for models above two trillion parameters and are entering commercial deployment at scale this quarter. That provides context for the V900 announcement: Alibaba is describing a successor to an existing server platform, not its first attempt to combine accelerators into a larger machine.

Wu also set a target of more than 20GW of globally operated data-center capacity by 2032. That is a corporate infrastructure goal, distinct from a V900 order book or the power draw of one chip. The keynote connects the plans through software and system optimization, including servers, networks, models and inference engines.

The different timelines should remain separate. Current M890 deployment, planned V900 production in 2027 and the 2032 capacity goal describe successive parts of Alibaba's strategy. None proves that the later capacity is already installed or that all of it will use V900 accelerators.

What the international expansion actually promises

A separate September 23 Alibaba Cloud announcement puts a nearer timetable on its overseas expansion. The company plans its first cloud regions in Türkiye, Finland and the Netherlands during the following 12 months, alongside additional data-center capacity in Malaysia, Germany, the United Arab Emirates, France and Hong Kong. It says the existing network comprises 107 availability zones across 31 regions.

That release also introduces Smart Studio for businesses packaging and billing model APIs under their own brands, and Smart Fusion for choosing and combining models through a common endpoint. These products describe how customers might consume and sell AI capacity. The announcement does not assign V900 hardware to the new overseas regions, so their planned opening should not be read as a V900 availability schedule.

This gives overseas customers two separate procurement tracks to follow: the chip and server roadmap, and regional service catalogs. A region opening can improve access to Alibaba services without proving that a particular accelerator, memory configuration or software stack is offered there.

The milestones after the launch

For customers, the next concrete questions are when production hardware can be ordered, which software and models are supported, and whether a trial workload can be reproduced at a larger scale. The announcement does not answer those procurement questions with a public price list or independent test.