Alibaba's quarterly report gives three growth rates for Alibaba Cloud. Revenue from AI cloud and computing power services rose 53% year on year, external customer revenue grew 64%, and recurring revenue from external customers increased 67%. Three metrics for the same business, each growing faster than the last.

The 11-percentage-point gap between 53% and 64% reflects the difference between internal and external customers, with the external business growing faster. External customer revenue grew 45% in the previous quarter and 64% this quarter, which the company said was its fastest pace in nearly six years.

The three-percentage-point gap between 64% and 67% comes from excluding computing equipment solutions. According to the report, recurring revenue primarily comes from pay-as-you-go or subscription-based public cloud services, as well as license and software renewals. Subscription and renewal revenue will return next year; equipment sales must be won all over again. Once one-off equipment deliveries are stripped out, the remaining business is growing even faster, indicating that recurring revenue drove this quarter's growth.

In absolute terms, revenue from AI cloud and computing power services reached 60.973 billion yuan in the quarter ended September 30, up 53% year on year. Revenue from AI-related products posted triple-digit growth for the 13th consecutive quarter.

14.5 Yuan Left From Every 100 Yuan in Revenue

On revenue of 60.973 billion yuan, adjusted EBITA was 8.838 billion yuan. Divide 8.838 by 60.973 and the result is 14.49%, rounded to 14.5% in the report.

EBITA is earnings before interest, taxes and amortization. Put more plainly, for every 100 yuan in revenue, 14.5 yuan remains at this level.

That ratio is higher than it was a year ago. Revenue rose 53%, while EBITA jumped 121%, more than twice as fast. The resulting margin expansion is arithmetic, not rhetoric. A 14.5% margin is not especially rich for a cloud business, but it is moving in the right direction.

M890 Is in Service; V900 Is Due Next Quarter

T-Head's M890 underpinned Alibaba Cloud's AI computing power supply this quarter. The report said its in-house AI supernodes powered by the M890 can efficiently run inference for large models with more than 2 trillion parameters.

The weights of a 2-trillion-parameter model are far beyond the capacity of a single card. Support at this scale therefore depends on the supernode, which pools memory across multiple cards and uses high-bandwidth interconnects to partition and distribute the model. In a conventional server cluster, traffic between machines during model partitioning travels over general-purpose networks. A supernode replaces that layer with dedicated high-bandwidth interconnects, so cross-card communication does not leave computing power sitting idle.

The report did not disclose the M890's computing power, memory capacity or interconnect bandwidth, nor how many cards each supernode contains. It provided only the upper limit of 2 trillion parameters. Service is already available in the Ulanqab and Zhongwei regions; this year's push has been a large-scale rollout, not an initial launch.

The next generation will not arrive until next quarter. The Zhenwu V900 is scheduled to enter mass production and go on sale then. The report described it as “the highest-performing domestically developed Chinese AI chip currently available,” offering no figures to support the claim.

The same section that discusses the V900 also mentions clusters containing 500,000 cards. The report emphasizes a “newly designed AI data center network architecture.” A 500,000-card cluster is not simply a chip challenge; it is an interconnect and scheduling challenge. As clusters grow, so do their failure domains. A network disruption can force an entire batch of jobs back into the queue. Delivery is likewise scheduled for next quarter.

4 Billion Downloads and Four Claims Without Numbers

On the model side, the latest development is Qwen4-Max. The report said it “delivers higher performance with lower token consumption” and continues to improve in two key areas: agents and coding. It did not quantify either the token savings or the performance gains.

The figures it does provide are cumulative. The Qwen family has surpassed 4 billion downloads worldwide and spawned more than 300,000 derivative models, making it the world's No. 1 open-source model family, according to the report. That ranking is based on downloads and derivative models, not usage or revenue. And 4 billion refers to downloads, not models or users: repeatedly pulling the same weights adds to the total.

The Wan3.0 upgrade is described in four phrases: longer generation, versatile creation, comprehensive reference capabilities and real-world fidelity. All four are qualitative, with no supporting figures.

Qwen's iPhone Debut Will Have to Wait Until December

Qwen has been integrated into Apple Intelligence. Users in China will be able to access its text and image understanding and content-generation capabilities directly through Siri, Writing Tools and Visual Intelligence.

The rollout has two stages. Testing opened with the iOS developer beta on November 4, but the service will not launch for users in China until the official release of iOS 27.3 in December. It is the only consumer-facing gateway in the quarterly report with a firm date, and that date falls next month.

Five Industries on the Customer List

The report attributed cloud growth to five industries: internet, automotive, embodied intelligence, financial services and biotechnology.

In automotive, it named BYD and NIO, which are upgrading intelligent cockpits to support multi-turn conversations, complex intent recognition and task execution. In finance, Ping An of China is using Alibaba Cloud for full-stack optimization across models, inference and computing power scheduling. Its partnership with CPIC Technology has expanded from cloud infrastructure to AI platform development and use-case incubation.

On the enterprise side, Alibaba offers Qwen Workplace. Its “Enterprise Context” feature, launched in September, organizes information from corporate documents, meeting minutes and internal communications, enabling agents to execute tasks based on organizational knowledge and workflows. Customers already using it include Yihai Kerry Arawana, FAW Toyota and Guming.

In this quarterly report, the good news on computing power is slated for delivery next quarter, while the model's new consumer gateway is due in December. What actually reached the books this quarter was 60.973 billion yuan in revenue and 8.838 billion yuan in adjusted EBITA.