What changed
Alibaba has unveiled the Zhenwu V900, a new AI accelerator that the company says provides about three times the performance of its previous Zhenwu M890, while also outlining an ambition to train a model with as many as 10 trillion parameters.The announcement came at Alibaba’s annual conference in Hangzhou as Chinese technology companies continue building alternatives to the most advanced US accelerators. Alibaba says the V900 is currently its most powerful AI chip and is designed for both training and inference inside its data centres and cloud business.
The 10-trillion-parameter target is even more ambitious than the chip announcement. Alibaba’s current Qwen3.8-Max is reported at 2.4 trillion parameters, while other Chinese model developers are also pushing the scale of open and closed systems.
Why the market is watching
Parameter count alone does not determine model quality. Architecture, training data, compute efficiency, context handling and inference methods can matter as much as raw scale. But the target signals how much computing Alibaba believes will be required for the new of models.The company is also planning to expand its data-centre capacity to as much as 20 gigawatts by 2032. That figure shows the physical side of the AI race. More capable models require more accelerators, more networking, more electricity and more cooling capacity.
Alibaba’s strategy is So broader than designing a chip. It is trying to control a complete stack: accelerators, cloud infrastructure, models and the services that consume them. That can reduce dependence on external hardware while giving Alibaba a larger role in the domestic AI supply chain.
The timing also matters because US-China technology restrictions remain a major factor in the availability of high-end AI processors. Chinese companies have been forced to improve domestic alternatives while improving software around whatever hardware they can obtain.
The V900 will need to prove its value in real workloads, but Alibaba’s announcement makes the direction clear. China’s AI competition is increasingly being fought not only at the model layer, but across the entire computing stack beneath it.
The technology is moving quickly, but the commercial test remains familiar. Products have to work consistently, fit existing systems and justify their cost.
Alibaba’s plan also shows why model development and infrastructure development can no longer be treated as separate stories. A 10-trillion-parameter target is meaningful only if the company can assemble enough compute, memory, networking and electricity to train and serve such a system. The hardware roadmap and the model roadmap So reinforce each other.
The real test will be whether the hardware can support the ambitious model roadmap at commercially useful cost.
Alibaba’s plan also shows the importance of memory and interconnect technology. A large model is not useful simply because a company can claim a parameter count. The system has to move data between compute units fast enough to keep the processors busy. That is why AI infrastructure increasingly depends on packaging, high-bandwidth memory and advanced networking alongside the accelerator itself.
The V900 announcement So belongs to a much larger story about computing independence and the infrastructure required to sustain it.
Alibaba’s 20-gigawatt data-centre ambition puts the scale of the hardware challenge into perspective. Training a frontier model is not simply a matter of buying more accelerators. It requires enough electrical capacity to run them, networking to connect them and cooling to keep them operating. If the model target expands from trillions of parameters toward even larger systems, infrastructure planning can become a bottleneck before chip design does. Alibaba is effectively betting that it can build those layers together.
The infrastructure requirement is what makes the announcement more significant than a parameter-count headline. A model at that scale needs memory bandwidth, networking, storage, cooling and power and accelerators. Alibaba’s chip and data-centre plans are So parts of the same strategy. The company is attempting to reduce dependence on external hardware while building enough computing capacity to keep its model ambitions moving.