HANGZHOU, China – Alibaba has unveiled a new artificial intelligence processor and plans for a future model with as many as 10 trillion parameters, expanding its effort to build a domestic, full-stack AI infrastructure spanning chips, models and cloud computing.
The Chinese technology company introduced the Zhenwu V900 at its annual Apsara Conference on Tuesday. Chief Executive Eddie Wu described it as the most powerful AI chip in China and said it delivers three times the performance of Alibaba’s previous-generation Zhenwu M890 processor.
Alibaba also said its Qwen team is training the next-generation Qwen 4 model and outlined a roadmap for future Qwen 4.5 and Qwen 5 models that could scale to between 5 trillion and 10 trillion parameters.
The announcements represent a broader strategy than simply developing another AI accelerator. Alibaba is attempting to control more of the computing stack needed to train and operate increasingly large models, at a time when Chinese technology companies face restrictions on access to some advanced foreign AI chips.
New chip targets large-scale AI computing
The Zhenwu V900 was developed by Alibaba’s T-Head semiconductor unit.
Alibaba said the processor is designed for both high-precision model training and lower-precision inference, the computing process through which trained AI models generate responses. The company said V900 processors can be connected into clusters of up to 500,000 chips for large AI workloads.
Mass production and commercial release are scheduled for the first quarter of 2027, according to Alibaba.
The three-times-performance figure is an Alibaba comparison with its previous-generation M890 chip. It is not an independent benchmark against competing processors such as Nvidia’s latest AI accelerators.
■ Related News
Alibaba launched the M890 only in May, making the rapid progression from that processor to the V900 part of a broader effort to accelerate its domestic AI-chip development.
Alibaba plans a much larger Qwen model
The company’s model ambitions are moving even faster in scale.
Alibaba said Qwen 4 is currently being trained, while future versions in the Qwen 4.5 and Qwen 5 series are projected to reach 5 trillion to 10 trillion parameters.
Parameters are numerical values learned during model training and are commonly used as a rough measure of model size. A larger parameter count does not by itself establish that one AI model is more capable than another because architecture, training data, computing methods and post-training techniques also affect performance.
Alibaba’s current flagship Qwen3.8-Max has 2.4 trillion parameters, according to the company and current reporting.
A model reaching 10 trillion parameters would therefore represent a substantial increase in scale compared with Alibaba’s current flagship.
The company has not announced a completed 10 trillion-parameter model or demonstrated that such a model will outperform competing systems. The figure is part of Alibaba’s future development roadmap.
The chip and model plans are linked
Alibaba’s announcements connect the model roadmap directly to its semiconductor and cloud strategies.
The company says it is building what it describes as a full-stack AI system encompassing foundation models, proprietary chips, cloud infrastructure and AI agents.
That approach addresses one of the central constraints facing companies developing frontier AI: access to sufficient computing capacity.
Training increasingly large models requires enormous quantities of processing power, memory, networking and electricity. Operating those models for millions of users requires still more infrastructure.
Alibaba is therefore also expanding the physical computing capacity behind its AI strategy.
The company plans to increase Alibaba Cloud’s global data-center capacity to more than 20 gigawatts by 2032. Wu said demand for AI computing is growing rapidly and that supply-chain constraints are limiting the speed at which the company can expand.
China is building alternatives to foreign AI processors
Alibaba’s announcement comes as Chinese technology companies accelerate efforts to develop domestic alternatives to Nvidia’s leading AI processors.
U.S. export controls have restricted Chinese access to some advanced AI chips and semiconductor-manufacturing technologies. Those restrictions have increased the strategic importance of domestic processor development for Chinese companies seeking to expand AI computing capacity.
Huawei has also accelerated development of its Ascend AI-chip family and recently announced additional processors and large-scale computing systems.
The result is a growing Chinese effort to develop an AI ecosystem that relies less heavily on foreign hardware.
Alibaba’s approach differs in that the company is simultaneously developing processors, foundation models and cloud infrastructure for commercial customers.
Scale will remain a major challenge
Designing an AI processor is only one part of building a competitive AI computing ecosystem.
Alibaba will need to manufacture the V900 in sufficient quantities, integrate it with high-bandwidth memory and networking infrastructure, and provide software that allows developers to use it efficiently.
The company’s ability to scale its data-center infrastructure will also depend on electricity, advanced semiconductor supply chains and other physical infrastructure.
Alibaba acknowledged that supply constraints are already limiting the pace of expansion. Wu said demand for AI computing is outpacing the company’s ability to supply it over the medium and long term.
The planned 500,000-chip clustering capability therefore represents an infrastructure ambition as much as a processor specification.
Alibaba’s AI strategy extends beyond the chip
Wu also presented the company’s AI development as part of a larger transition toward what Alibaba calls the “Machine Intelligence” era.
The company highlighted progress in what it describes as recursive self-improvement, including automated experimentation and optimization of its Qwen models. Alibaba said Qwen3.8-Max completed 33 automated improvement cycles over more than a month and improved its Artificial Analysis score from 40 to 45.
Alibaba also said one internal chip-design experiment used an AI model to make more than 10,000 electronic-design-automation tool calls over more than 60 hours, reducing chip area by 42% without reducing performance.
Those are company-reported results and have not been independently established by JournosNews.
The broader strategy is clear: Alibaba wants AI to contribute not only to cloud services but also to the development of the hardware and software used to build future AI systems.
A larger model does not automatically mean a better model
The planned 10 trillion-parameter scale is significant, but parameter count alone cannot determine how competitive the resulting system will be.
Modern AI development increasingly emphasizes efficiency as well as raw scale.
Techniques such as mixture-of-experts architectures, improved data selection, post-training, reasoning systems and specialized inference methods can allow models to achieve stronger results without simply increasing the number of parameters.
Alibaba’s own emphasis on model architecture and data optimization reflects that broader shift. The company says its next generation of Qwen models is intended to handle more complex and longer-horizon tasks.
The eventual performance of a 5 trillion- or 10 trillion-parameter model will therefore depend on how Alibaba trains and deploys it, not simply on the number attached to the model.
The next test is execution
Alibaba’s latest announcements show that China’s AI competition is moving beyond individual model releases.
The company is attempting to build a vertically integrated system in which its own chips provide computing for its own models and cloud customers, while its expanding data centers provide the infrastructure needed to operate them at scale.
The V900 is not yet in mass production, and the 10 trillion-parameter model remains a future target.
That makes execution the next major test.
Alibaba will have to turn its chip-performance claims into commercially available processors, expand manufacturing and data-center capacity, and translate larger models into measurable improvements for users.
For China’s AI industry, the significance of the announcement lies not in a completed 10 trillion-parameter system but in the scale of the infrastructure Alibaba says it is preparing to build.
Reporting Credit: Alibaba Group / Alibaba Cloud — Zhenwu V900 announcement, Qwen model roadmap, AI infrastructure plans and company-reported AI development results; U.S. Department of Commerce, Bureau of Industry and Security — semiconductor export-control framework affecting advanced computing technology exports to China.
