Alibaba Group unveiled a new artificial-intelligence accelerator that it says delivers three times the performance of its predecessor, while outlining plans for substantially larger AI models and more than 20 gigawatts of data-center capacity by 2032.
The Zhenwu V900, developed by Alibaba’s semiconductor unit T-Head, was introduced Tuesday at the company’s annual Apsara Conference in Hangzhou. Alibaba says the processor is designed for both AI training and inference and can be deployed in clusters containing as many as 500,000 accelerator cards.
The announcement expands Alibaba’s effort to control more of the AI technology stack — from semiconductors and data centers to foundation models, cloud infrastructure and AI agents — as Chinese technology companies seek greater access to computing capacity amid U.S. restrictions on advanced chips.
Zhenwu V900 Triples Previous Performance
Alibaba says the V900 delivers three times the performance of the Zhenwu M890, which was released in May.
The accelerator includes 216 GB of memory and 1,200 GB per second of inter-chip bandwidth, while supporting multiple data formats, including FP8 and FP4. Alibaba said commercial availability and mass production are scheduled for the first quarter of 2027.
The company also unveiled an upgraded supernode architecture combining the V900 with proprietary networking, storage and controller technology. That system can scale to clusters containing up to 500,000 accelerator cards, according to Alibaba.
Alibaba describes the V900 as China’s most powerful AI accelerator. That characterization is the company’s claim rather than an independently established industry benchmark.
Alibaba Targets Models With Up to 10 Trillion Parameters
Alongside the hardware announcement, Chief Executive Eddie Wu said Alibaba intends to train a next-generation model containing between 5 trillion and 10 trillion parameters.
That would represent a substantial increase from its current flagship Qwen model. Reuters reported that Alibaba’s Qwen 3.8 Max contains about 2.4 trillion parameters, making the proposed system roughly two to four times larger by parameter count.
Parameter count alone does not determine the capability or quality of an AI model, but the target illustrates the scale of computing infrastructure Alibaba is preparing to deploy.
More Than $53 Billion for AI Infrastructure
Alibaba has committed more than 380 billion yuan, or roughly $53 billion, to cloud and AI infrastructure over three years, according to Bloomberg. The company is now planning for Alibaba Cloud’s data-center capacity to exceed 20 gigawatts by 2032.
Wu described the strategy as a full-stack approach encompassing AI models, chips and data centers. Reuters reported that Alibaba shares rose 5.1% to a one-month high following Tuesday’s announcements.
China and U.S. AI Strategies Diverge Over Pace
The announcement comes during a broader debate over how quickly frontier AI should advance.
In the United States, executives including Anthropic CEO Dario Amodei have recently called for slowing the pace of improvements to the most advanced AI models while companies strengthen evaluation and safety systems. OpenAI CEO Sam Altman and SpaceXAI CEO Elon Musk have also supported greater attention to slowing frontier development, according to Bloomberg.
Alibaba, meanwhile, is outlining plans for substantially larger models, proprietary processors and a major expansion of computing infrastructure. The contrast does not amount to a unified U.S. policy of slowing AI: the Trump administration has publicly opposed an industrywide slowdown and emphasized maintaining U.S. leadership over China.
The competing approaches are emerging as artificial intelligence becomes an increasingly important part of the economic and technology relationship between the world’s two largest economies. U.S. and Chinese officials agreed this week to establish a dialogue on AI ahead of talks between Presidents Donald Trump and Xi Jinping.
For Alibaba, the V900 is therefore more than a new processor. Together with the planned 5-to-10-trillion-parameter model and 20-gigawatt infrastructure target, it represents an effort to build an increasingly self-contained AI ecosystem spanning chips, computing infrastructure, models and applications.
