Alibaba released the Zhenwu V900 accelerator to power its Qwen AI model family, marking Alibaba's expansion into custom silicon for in-house model training.

Alibaba is building its next AI push from both ends. At the Apsara Conference in Hangzhou on September 22, CEO Eddie Wu unveiled a new accelerator for training and inference while laying out a roadmap for Qwen models that could eventually reach 10 trillion parameters.

The company announced that its next-generation model, Qwen 4, is currently in training, with roadmaps for the upcoming Qwen 4.5 and Qwen 5 model series projected to scale up to 5 to 10 trillion parameters.

The hardware is the Zhenwu V900, developed by Alibaba's T-Head chip division. The Zhenwu V900 delivers three times the performance of its predecessor, the Zhenwu M890, with 216GB of GPU memory and 1,200GB/s of inter-chip bandwidth.

The model roadmap extends much further out. Qwen 4 is currently in training, while Qwen 4.5 and Qwen 5 are projected to scale to between 5 trillion and 10 trillion parameters—up to roughly four times the 2.4 trillion parameters in the company's current Qwen 3.8-Max flagship.

V900 is designed as a unified training and inference processor rather than a chip aimed at only one stage of AI development. Its support for FP8 and FP4 precision improves efficiency for different workloads, and the accelerator can be combined into supernode clusters containing as many as 500,000 cards. The accompanying server design combines the V900 with Alibaba's ICN Switch, Panmai SmartNIC and Zhenyue storage controller.

Training and serving models with trillions of parameters requires thousands of processors to communicate efficiently, along with sufficient memory, networking and storage to supply those processors with data. The V900 aims to reduce inference costs while increasing compute density, seamlessly handling both high-precision model training and ultra-low-precision inference.

Alibaba has not announced a public release date, pricing or performance benchmarks for Qwen 4. The 5 trillion to 10 trillion parameter target applies to Qwen 4.5 and Qwen 5, not Qwen 4 itself. Parameters measure model size, not intelligence directly—a larger model can still underperform on particular tasks depending on its architecture, training data, optimization and inference efficiency.

Eddie Wu, CEO of Alibaba Group, wrote in a letter to shareholders: "Vast numbers of AI agents are poised to take on an ever-greater share of work in the digital economy, each powered by tokens generated from models. Agents will increasingly serve as the primary interface between humans and the digital world."

Alibaba Cloud aims for the global data-center capacity it operates to exceed 20 gigawatts by 2032. The company also unveiled the Yitian 720 and Yitian 730 server CPU roadmap for 2027, extending its in-house hardware effort beyond AI accelerators.

"With our full-stack AI strategy, we have put Alibaba in a superior position to capture the substantial growth of demand for artificial intelligence and AI compute," Wu said.

This strategy reflects constraints facing Chinese AI companies. Huawei accelerated its own next-generation Ascend chip roadmap at its September 17 conference, while U.S. restrictions have limited Chinese companies' access to some of Nvidia's most advanced processors. For Alibaba, building more of the stack internally reduces the number of points at which its AI plans depend on an outside supplier. It does not eliminate manufacturing constraints facing China's semiconductor industry, but it gives the company more control over the hardware and infrastructure it can deploy domestically.

Alibaba's more consequential bet is that its own silicon, Qwen models and cloud infrastructure can be designed to work together efficiently enough to support increasingly large AI systems. If successful, the company gains more control over the economics and availability of its AI stack.

That remains unproven. The V900 does not enter commercial release until Q1 2027, Qwen 4 is still in training, and the 5 trillion to 10 trillion parameter target belongs to future model generations. Manufacturing capacity, real-world chip performance and the eventual quality of those models will determine whether the strategy succeeds. Alibaba has made its direction clear: rather than building one domestic alternative to Nvidia, it is attempting to build enough of the chip, model and infrastructure stack itself that its AI ambitions depend less on what it can obtain from outside China.