Alibaba Group is planning to train a new artificial intelligence model with between 5 trillion and 10 trillion parameters as part of a broader push across AI models, chips and data-centre infrastructure. Alibaba CEO Eddie Wu announced the plan at Alibaba Cloud’s annual Apsara Conference in Hangzhou on September 22, outlining the company’s longer-term strategy for developing AI capabilities and infrastructure.
Alibaba said its next-generation Qwen 4 model is currently in training, while the upcoming Qwen 4.5 and Qwen 5 series are projected to scale to between 5 trillion and 10 trillion parameters. Parameters are the variables that an AI model learns during training and are commonly used as an indicator of model size.
The planned model would be significantly larger than Alibaba’s current flagship Qwen 3.8 Max, which the company says has 2.4 trillion parameters. Based on those figures, the planned 5 trillion to 10 trillion-parameter models would be roughly two to four times larger in parameter count.
Wu said Alibaba’s Qwen team is continuing research into model architecture and data optimisation, with the objective of enabling AI systems to handle more complex and longer-horizon tasks. He also said the company is working towards artificial superintelligence, or ASI.
Another area highlighted by Alibaba is recursive self-improvement. Wu said the Qwen team has made meaningful progress in this area, describing it as a process in which AI models identify their limitations, design experiments, synthesise data and evaluate results as part of a continuing cycle of improvement.
Alongside its model roadmap, Alibaba unveiled the Zhenwu V900, a next-generation AI chip developed by its T-Head semiconductor unit. Wu described the V900 as the most powerful AI chip in China and said it delivers three times the performance of its predecessor, the M890.
Alibaba said a single cluster based on the V900 can support up to 500,000 cards for frontier AI model training and inference. The company expects the chip to enter mass production and commercial release in the first quarter of 2027. The previous M890 chip was released in May 2026.
Alibaba is also developing infrastructure to support the growing computing requirements of large AI models. The company plans to increase Alibaba Cloud’s global data-centre capacity to more than 20 gigawatts by 2032.
Wu said Alibaba Cloud is already bringing its AI supernodes online at commercial scale. Alibaba’s proprietary M890 AI supernode can handle inference for models above 2 trillion parameters, according to Wu, who said only a handful of companies globally currently possess that capability.
The company’s strategy reflects its effort to build AI capabilities across multiple layers rather than relying solely on model development. Alibaba has been investing in proprietary AI chips and cloud infrastructure alongside its Qwen model family. In May 2026, the company said its T-Head AI chips had reached production at scale and that AI-related products were becoming an increasingly important part of Alibaba Cloud’s business.
The expansion also comes as Chinese technology companies work to develop domestic AI computing infrastructure amid restrictions on access to some advanced US-made processors. Alibaba’s development of its own AI chips is therefore part of a wider effort within China to build domestic alternatives across the AI hardware and software stack.
Alibaba expects demand for AI computing to remain strong. Wu said customer demand for AI was “exceptionally robust” and that shortages across the AI data-centre supply chain were limiting how quickly Alibaba Cloud could expand. He added that the company expects industry demand in the medium to long term to exceed its current supply capabilities.
With Qwen 4 already in training, larger Qwen models planned for the future, a new generation of proprietary AI chips and a target of more than 20 GW of global data-centre capacity by 2032, Alibaba is positioning AI models, computing hardware and cloud infrastructure as connected parts of its long-term technology strategy.
