At the Apsara Conference in Hangzhou on 22 September, Alibaba Group chief executive Eddie Wu announced plans to train an artificial intelligence model containing between 5 trillion and 10 trillion parameters, laying out a strategy that spans models, chips and data centres, Reuters reported.
Key points
- Alibaba plans models of 5 trillion to 10 trillion parameters, roughly two to four times the size of its current 2.4 trillion parameter flagship
- The Zhenwu V900 chip, from Alibaba’s T-Head semiconductor unit, is scheduled for mass production in the first quarter of 2027
- Alibaba Cloud targets global data centre capacity above 20 gigawatts by 2032
- The Qwen team reports progress on recursive self-improvement, where models identify limitations and synthesise data to evolve
Model scale and the Qwen roadmap
Alibaba said its next-generation model Qwen 4 is currently in training, and the upcoming Qwen 4.5 and Qwen 5 series are projected to scale up to 5 trillion to 10 trillion parameters, Reuters reported. The current flagship, Qwen 3.8 Max, contains 2.4 trillion parameters, so the planned model would be approximately two to four times that size. Parameters are the variables a model learns during training and serve as a rough gauge of its size.
According to Wu, Alibaba’s own M890 AI supernode already runs inference for models exceeding 2 trillion parameters, a feat he said only a handful of companies worldwide can claim. The Qwen team has achieved what Wu called “meaningful progress” on recursive self-improvement, where models spot their own weaknesses, devise experiments and generate data to fuel self-directed evolution.
Zhenwu V900 and Alibaba’s silicon push
Alibaba also unveiled the Zhenwu V900, a next-generation AI processor from its T-Head semiconductor division, which Wu described as the most powerful AI chip in China. Wu said the Zhenwu V900 delivers triple the performance of the earlier M890, and a single cluster built on it can hold up to 500,000 cards for frontier model training and inference. The M890 was released in May.
Mass production and commercial release of the Zhenwu V900 are slated for the first quarter of 2027, Alibaba said. Wu said the company expects “significant growth” in annual AI chip shipments. The announcement arrives while Chinese companies scramble to develop home-grown substitutes for Nvidia processors under U.S. export controls.
Data centre targets and supply constraints
Wu set a goal for Alibaba Cloud’s worldwide data centre capacity to exceed 20 gigawatts by 2032. He said AI demand stays “exceptionally robust” and is driving Alibaba Cloud’s revenue higher, though he noted that worldwide shortfalls in the AI data centre supply network are constraining the pace of expansion.
The industry’s mid-to-long-term demand far outpaces our supply capabilities.
Eddie Wu, Chief Executive, Alibaba Group
Alibaba Cloud will begin bringing its AI supernodes online at commercial scale this quarter, Wu said. He cast the moment as the start of a “Machine Intelligence” era on par with the Industrial Revolution, saying machines will one day generate over 1,000 times the “thinking” of all humanity, up from under 3 percent today. He compared AI-assisted programming to the light bulb in the age of electricity — an initial use case, not the defining invention — and added that the truly groundbreaking products of the Machine Intelligence era have yet to appear.
