Alibaba used the Apsara conference on September 22 to present an AI strategy much broader than a new model launch.
The company unveiled the Zhenwu V900 chip, confirmed that Qwen 4 is in training, presented plans for future models with 5 to 10 trillion parameters and set a target of more than 20GW of data center capacity by 2032.
The numbers are large, but it is important to distinguish what has launched from what is still planned.
Qwen 4 is not currently available, Zhenwu V900 is scheduled for commercial release only in the first quarter of 2027, and many infrastructure targets concern the coming years.
The real story is Alibaba's attempt to build every AI layer itself: chips, computing systems, cloud and models.
What is Zhenwu V900?
Zhenwu V900 is a new AI chip from T-Head, Alibaba's chip division.
According to the company, it is intended both for training models and running them after training.
Alibaba describes it as the most powerful AI chip developed in China and claims three times the performance of its previous M890.
These are company claims.
No independent benchmark has yet been published that would allow the comparison to be verified.
Important details are also missing, including full power consumption, price, manufacturing yield and performance in different training and inference scenarios.
The chip is not yet a commercial product
This may be the most important detail in the announcement.
Zhenwu V900 is not currently available for commercial purchase.
According to the published schedule, commercial production and release are due in the first quarter of 2027.
Alibaba also plans to bring supernode systems to commercial scale in the fourth quarter of 2026.
A headline presenting the chip as an existing alternative to NVIDIA would therefore be premature.
This is a significant stage in Alibaba's strategy, but performance and the ability to manufacture at scale still need to be tested.
What is happening with Qwen 4?
Alibaba announced that Qwen 4 is already in training.
But it has not launched and is not currently available to developers.
The company also presented a plan to scale Qwen 4.5 and Qwen 5 to 5 to 10 trillion parameters.
That is an unusually large number, but parameter count alone does not mean a model will be better.
Modern models can use architectures in which only some parameters are active for each task.
Data quality, the training process, reinforcement learning, inference efficiency and the tools surrounding the model can also matter as much as the total count.
There is therefore no basis yet for saying that a future Qwen model at this scale will be "several times more powerful" than smaller models.
Why Alibaba wants to control every layer
The move comes against the backdrop of US restrictions on exports of advanced AI chips to China.
Chinese companies are looking for ways to reduce their dependence on NVIDIA and a supply chain outside their control.
Alibaba is attempting this through a combination of domestic hardware, cloud infrastructure and its own models.
From a business perspective, this strategy resembles the idea of full-stack AI: trying to control as many layers as possible between the data center and the service the customer receives.
If it succeeds, Alibaba Cloud could offer customers a system in which the model, hardware and infrastructure are designed to work together.
The big target: more than 20GW
Alibaba set a target of more than 20GW of data center capacity worldwide by 2032.
This is a very ambitious target.
Building such capacity requires not just chips but electricity, cooling systems, network equipment, advanced memory, real estate and local infrastructure.
The number therefore matters as an indication of the direction the company plans to take, not as proof that this capacity already exists.
Up to 500 thousand chips in a cluster
Alibaba claims its architecture can connect up to 500 thousand chips in a cluster.
Here too, distinguish a design capability from a commercial system proven at this scale.
It is not yet clear whether the company has actually operated a system of 500 thousand chips, under what load or for how long.
This is precisely one question the market will need to examine as the hardware moves toward commercial production.
What it means for developers
In the short term, not much changes.
There is no Qwen 4 to open and test yet, and V900 is not an available product either.
But the longer-term implications are broader.
If Alibaba succeeds in connecting models, chips and cloud, developers and companies may gain another alternative to the American ecosystem.
This matters especially in markets where price, data sovereignty or export restrictions make choosing an infrastructure provider a strategic decision.
What to take from the announcement
Alibaba presented a clear direction: it does not want to be merely a company running models on other companies' hardware.
It wants to build the hardware, computing systems, cloud and models.
But most of the plan's most ambitious parts still lie in the future.
Zhenwu V900 is due commercially in 2027. Qwen 4 is still in training. The models with 5 to 10 trillion parameters are on a roadmap. The 20GW target concerns 2032.
This is a major story about competition in AI infrastructure, but it is not yet proof that Alibaba has closed the gap with NVIDIA or other Western players.