Alibaba Unveils Zhenwu V900 AI Chip, Massive Qwen Models, and 20 GW Compute Ambitions
Alibaba made a major statement at its Apsara Conference in Hangzhou, China, outlining a bold roadmap for artificial intelligence infrastructure, custom silicon, and next-generation foundation models. The company revealed new details about its upcoming Zhenwu V900 accelerator, discussed its long-term compute expansion plans, and confirmed that future Qwen AI models could reach an enormous 4 trillion to 10 trillion parameters.
The announcements signal Alibaba’s growing determination to compete at the highest level of AI hardware and cloud computing, especially as demand for large-scale model training continues to surge worldwide.
Alibaba Zhenwu V900 chip targets 216GB of on-package memory
One of the biggest highlights was Alibaba’s upcoming Zhenwu V900 AI chip, which is expected to launch in the first quarter of 2027.
According to the company’s latest details, the Zhenwu V900 will feature 216GB of on-package memory. That is a major jump for AI accelerators and would give the chip a strong advantage in memory-heavy workloads such as large language model training, inference, multimodal AI, and advanced generative AI applications.
The V900 is also expected to support native FP8 and FP4 formats, which are increasingly important for improving AI performance and efficiency. These lower-precision formats allow models to run faster and consume less power while maintaining strong accuracy in many AI workloads.
Alibaba says the new accelerator will deliver roughly three times the performance of its previous Zhenwu M890 chip. Based on the M890’s reported performance level, this suggests the Zhenwu V900 could offer around 1.8 petaflops of peak FP16 compute power per chip.
ICN Switch fabric could allow massive AI clusters
Alibaba also highlighted its ICN Switch interconnect technology, which is designed to link AI chips together at extremely high speeds. The Zhenwu V900 is expected to support chip-to-chip interconnect bandwidth of around 1.2 TB/s through this fabric.
The company claims the interconnect can allow roughly 1,000 Zhenwu V900 chips to operate like a single accelerator. Even more striking, Alibaba says a full V900 cluster could scale up to 500,000 chips.
If achieved, that would create a massive pool of compute and memory. With 216GB of memory per chip, a 500,000-chip cluster would offer around 108 petabytes of total memory across the system. That level of scale would be aimed at training and serving some of the largest AI models in the world.
Alibaba is building a full rack-scale AI system
Rather than relying only on individual chips, Alibaba is also developing a complete rack-scale solution built around its own hardware ecosystem.
The system is set to include Yitian CPUs, Zhenwu V900 AI accelerators, ICN interconnect technology, Pangu network interface cards, and Zhenyue storage controllers.
This approach shows that Alibaba wants tighter control over the entire AI infrastructure stack, from processing and networking to storage and system-level optimization. For cloud providers, owning more of the hardware platform can improve efficiency, reduce bottlenecks, and create better performance for large-scale AI workloads.
Alibaba plans 20 GW of compute capacity by 2032
Beyond chip development, Alibaba outlined an aggressive expansion plan for global data-center and compute capacity. The company is targeting more than 20 gigawatts of compute capacity by 2032.
That is a massive figure and reflects the growing energy demands of artificial intelligence. Training frontier AI models, running inference at scale, and supporting enterprise cloud workloads all require enormous amounts of power, cooling, and data-center space.
Estimates have suggested Alibaba could reach around 5 GW of compute capacity by the end of 2026. If that proves accurate, the company would need to add roughly 2 GW to 3 GW of additional capacity every year through 2032 to hit its long-term goal.
This expansion would place Alibaba among the world’s most ambitious AI infrastructure builders.
Qwen 4.5 and Qwen 5.0 could reach 10 trillion parameters
Alibaba also shared major news about its future Qwen AI models. The company said its upcoming Qwen 4.5 and Qwen 5.0 models are expected to span between 4 trillion and 10 trillion parameters.
That would represent a huge leap in model scale. Parameter count is not the only measure of AI capability, but at this level, the models would likely be designed for advanced reasoning, coding, multimodal understanding, enterprise automation, scientific research, and complex agent-based tasks.
Large models of this size demand vast amounts of compute power, fast memory, efficient interconnects, and optimized training systems. Alibaba’s Zhenwu V900 roadmap and 20 GW infrastructure plan appear closely tied to these AI model ambitions.
Alibaba is also exploring recursive self-improvement
Another notable part of Alibaba’s AI strategy is its work on recursive self-improvement, often described as a process where AI systems help improve future versions of themselves.
In practical terms, this can involve using AI to identify weaknesses, analyze real-world task feedback, design experiments, generate training data, assist with code development, and guide future training cycles.
If successful, this approach could accelerate AI development by reducing the amount of manual work needed to improve model quality. It could also create significant efficiencies as models become better at helping engineers and researchers build stronger successors.
Qwen-Image-2.1 strengthens Alibaba’s open AI portfolio
Alibaba has also released the model weights for Qwen-Image-2.1, a text-to-image generation and editing model. The visual generation system uses around 7 billion parameters across 32 single-stream DiT layers.
The model has gained attention for its image editing capabilities and has been ranked highly among open image editing models. This adds to Alibaba’s growing Qwen ecosystem, which now spans language models, image generation, editing tools, and broader multimodal AI research.
Alibaba’s AI strategy is becoming clearer
Alibaba’s latest announcements show a company moving aggressively across every layer of artificial intelligence.
It is developing custom AI chips, scaling interconnect technology, building full rack-level systems, expanding global compute capacity, and preparing trillion-parameter Qwen models. At the same time, it is investing in techniques such as recursive self-improvement and releasing advanced open AI models for image generation and editing.
The Zhenwu V900 may still be some time away, with launch expected in early 2027, but Alibaba is already positioning it as a central part of its future AI infrastructure. If the company can deliver on its performance, memory, and scaling claims, the chip could become a major force in large-scale AI training and cloud-based inference.
With demand for artificial intelligence continuing to rise, Alibaba’s long-term bet is clear: the future of AI will be shaped not only by smarter models, but also by the massive infrastructure needed to train and run them.
