Alibaba unveiled its in-house AI chip, the Zhenwu V900, at the Apsara Conference in Hangzhou on September 22 (local time).
Key points: Alibaba says the V900 delivers three times the performance of its predecessor, the M890, released in May, and that up to 500,000 chips can be linked together to train cutting-edge models.
The chip was developed by T-Head, Alibaba’s semiconductor subsidiary, and will enter mass production and go on sale in the first quarter of next year.
Alibaba has introduced an in-house AI chip aimed at Nvidia. Wu Yongming, Alibaba’s chief executive, unveiled the Zhenwu V900 AI accelerator at the Apsara Conference, the company’s annual developer conference held in Hangzhou, China, on September 22 (local time), calling it the most powerful AI chip in China. Developed by T-Head, Alibaba’s semiconductor subsidiary, the product will enter mass production and go on sale in the first quarter of next year.
According to the company, the V900 delivers three times the performance of its predecessor, the Zhenwu M890, which was released in May. It is equipped with 216 gigabytes of high-bandwidth memory (HBM), while data can be transferred between chips at a rate of 1.2 terabytes per second. Up to 500,000 chips can be connected as a single cluster for training cutting-edge models.
Rather than relying on the performance of individual chips, the approach links hundreds of thousands of chips to increase overall performance. This is a strategy also adopted by Chinese semiconductor companies such as Huawei and Moore Threads. It is a response to the U.S. blockade on exports of Nvidia’s advanced chips to China.
T-Head said it has shipped 560,000 Zhenwu-series chips to date, which are being used by more than 400 external customers. Bloomberg reported that Alibaba plans to release new AI chips every year and is also pursuing a listing for T-Head.
Data centers totaling 20 gigawatts
Wu said Alibaba plans to increase the global capacity of Alibaba Cloud’s data centers to more than 20 gigawatts by 2032, citing the exponential growth in demand for AI computing. Twenty gigawatts represents roughly the amount of electricity used by 20 nuclear power plants.
Although this would be among the largest such capacities for a Chinese company, it is not overwhelmingly large compared with SoftBank’s plan to build a 10-gigawatt facility in Ohio alone.
Alibaba has committed 380 billion yuan (approximately $53 billion) to AI over three years and raised an additional $10.2 billion through a rights offering on the Hong Kong stock market in August. Last quarter, its profits fell by more than 75% following capital expenditures of nearly $10 billion.
The company has said that AI-related revenue will reach an annualized $10 billion in the third quarter of this year and that it could recoup its investment within three years. It also aims to increase cloud and AI revenue fivefold to $100 billion within five years.
Wu also said the pace of computing expansion is being constrained by shortages across the AI data-center supply chain. AI supernodes—server racks densely packed with accelerators—will begin operating at commercial scale in the third quarter of this year.
Models with 10 trillion parameters
Alibaba also announced its model plans. The company is training Qwen 4, while the subsequent Qwen 4.5 and Qwen 5 series will have between 5 trillion and 10 trillion parameters. Parameters are internal values adjusted during training and serve as a measure of a model’s size.
The company’s current top model, Qwen 3.8 Max, has 2.4 trillion parameters. Alibaba said this is an unusually large size among open-weight models and is intended to handle longer and more complex tasks.
The announcement came one day before a planned summit between U.S. President Donald Trump and Chinese President Xi Jinping. AI is expected to be a major agenda item at the meeting, while Nvidia CEO Jensen Huang and Microsoft CEO Satya Nadella are scheduled to attend a state dinner at the White House. Earlier, U.S. Treasury Secretary Scott Bessent and Chinese Vice Premier He Lifeng agreed to establish a channel for dialogue aimed at building a common understanding of AI goals and threats.
The United States has alleged that Alibaba, DeepSeek and others improperly accessed advanced U.S. models to develop their own products. China has rejected the claims as unfounded and warned that it would retaliate if the United States targeted Chinese AI companies.
The two countries also differ over the pace of development. Debate has continued in the U.S. industry since Dario Amodei, CEO of Anthropic, proposed on September 12 that development should be slowed to address AI risks. Wu said he still holds the conviction he expressed last year that the more powerful AI becomes, the more powerful humanity will become. Alibaba’s shares rose more than 5% at one point on the Hong Kong stock market on September 22.