Alibaba’s Zhenwu V900: China’s Most Powerful AI Chip

Alibaba has officially introduced the Zhenwu V900, a processor that the company claims is currently the most powerful AI chip available in China. The new hardware delivers performance levels three times higher than its predecessor, the Zhenwu M890. This announcement was made by Alibaba CEO Eddie Wu during the Apsara cloud conference in Hangzhou on Tuesday. The chip, developed by Alibaba’s internal design unit T-Head, is engineered to handle both high-precision model training and ultra-low precision inference tasks efficiently.

The unveiling of the V900 was accompanied by two major strategic commitments. Alibaba outlined plans to develop artificial intelligence models with up to 10 trillion parameters. Additionally, the company set an infrastructure goal to achieve 20 gigawatts (GW) of data centre capacity globally by 2032. These targets underscore Alibaba’s aggressive expansion into the computational power sector.

Technical Specifications and Production Timeline

T-Head has released detailed specifications for the Zhenwu V900, highlighting significant advancements in memory and connectivity. The processor features 216GB of memory and supports an inter-chip interconnect bandwidth of 1,200 GB/s. According to reports from The Register, the chip natively supports floating-point formats ranging from FP32 down to FP4. T-Head has redesigned the Tensor Core arithmetic unit to enhance instruction precision specifically for FP8 and FP4 operations. The architecture also includes additional scaling factor formats and block size configurations under MXFP8 and MXFP4 standards.

Eddie Wu stated that a single cluster can accommodate up to 500,000 V900 chips. This configuration is designed to power frontier-level model training and inference workloads. While the previous generation M890 launched in May, mass production of the V900 is scheduled for the first quarter of 2027. Wu anticipates substantial growth in annual AI chip shipments as part of this rollout. Bloomberg reported that Alibaba intends to refresh its chip lineup annually.

Commercially, the Zhenwu family has already gained traction. Deliveries have reached more than 400 external customers, though Alibaba’s own communications suggest the number exceeds 650 clients. These customers span diverse sectors including automotive, language models, and embodied intelligence. Concurrently, Alibaba announced a roadmap for its next-generation Yitian server processors, which are optimized for agentic AI work and are also expected in 2027.

Scaling Qwen Models and Autonomous Learning

Alibaba is simultaneously advancing its large language model capabilities. The company is currently training Qwen 4, with subsequent versions named Qwen 4.5 and Qwen 5 planned to scale between five and 10 trillion parameters. For context, the current flagship model, Qwen 3.8 Max, contains 2.4 trillion parameters and was described as the company’s most capable model in August.

Wu framed the development of these larger models as a pathway toward artificial superintelligence. He explained that increased parameter counts enable the handling of longer and more complex tasks. Furthermore, the Qwen team is exploring recursive self-improvement mechanisms. In this system, a model identifies its own weaknesses, conducts experiments, and generates training data with minimal human intervention. Wu characterized this progress as meaningful.

The infrastructure push supports these ambitious model sizes. Wu set a target for Alibaba Cloud to surpass 20GW of global data centre capacity by 2032. Citigroup estimates that this level of capacity could generate over $160 billion in external revenue for the cloud division. By comparison, Cushman and Wakefield reports that 37.7GW of data centres are under construction in the United States alone. SpaceX, which held roughly 1.4GW of AI computing capacity earlier this year, aims to exceed 10GW by 2027.

Supply Chain Constraints and Financial Commitment

Despite the ambitious goals, supply chain limitations remain a critical bottleneck. Wu noted that global shortages across the AI data centre supply chain are restricting how quickly Alibaba can scale its compute resources. He emphasized that mid-to-long-term industry demand significantly outpaces current supply capabilities. Alibaba Cloud will begin deploying AI supernodes at commercial scale within this quarter, though specific locations for new data centres were not disclosed.

Financially, Alibaba has committed over $53 billion over three years to its AI expansion. This follows a follow-on share sale in August that raised approximately $10.2 billion. The heavy spending has impacted recent financial results, with quarterly profit dropping by 75% due to nearly $10 billion in expenditures. However, executives project that annualized revenue from AI products will reach $10 billion this quarter. The company expects to recoup its investment within three years, targeting $100 billion in combined cloud and AI revenue over five years.

Market reaction to the announcements was positive. Shares of Alibaba listed in Hong Kong rose 5.1% on Tuesday, reaching their highest point in a month. Tencent shares gained more than 7% following similar sector optimism.

Geopolitical Context and Strategic Positioning

The push for domestic chip design is heavily influenced by geopolitical factors. United States export controls prohibit Chinese firms from accessing Nvidia’s most advanced accelerators and largely restrict access to Taiwan Semiconductor Manufacturing Company (TSMC) facilities. It remains unclear whether Alibaba collaborates with SMIC, a domestic foundry facing its own capacity constraints. Huawei recently unveiled new chip technology at its Connect conference, indicating broader industry efforts to overcome these restrictions.

Neil Shah, vice president at Counterpoint Research, told the Associated Press that leveraging extra computing power helps China maintain strength in the local market. His colleague Parv Sharma highlighted that narrowing the technological gap depends on China’s ability to advance its own semiconductor foundries, not just chip design capabilities.

These developments occur against a backdrop of diplomatic engagement. Chinese President Xi Jinping arrived in Washington for a state visit and meeting with Donald Trump, where AI, trade, and tariffs were expected topics. Treasury Secretary Scott Bessent and Vice Premier He Lifeng recently agreed to establish a dialogue on AI to foster mutual understanding of goals and threats. Meanwhile, Washington has accused Alibaba and DeepSeek of illicitly accessing American models, charges Beijing denies. Anthropic previously accused Alibaba of distillation techniques in June.

Addressing concerns about the pace of development, Dario Amodei, Anthropic’s chief executive, recently called for slower model advancement. Wu responded directly to this sentiment:

> Last year, we asserted that the more capable AI becomes, the more powerful humanity will be. Today, I remain steadfast in that conviction.

Redefining Machine Intelligence

Throughout his keynote, Wu deliberately avoided the term “artificial intelligence,” preferring machine intelligence or machine thinking. He drew parallels between the current era and the industrial revolution. Just as steam and combustion engines were built to augment physical labor done by horses and labourers, machine power now drives 99.9% of the world’s physical work. Wu predicts a similar pattern will emerge in cognition.

He estimated that machines currently produce less than 3% of all thinking. However, he predicted they would eventually generate more than 1,000 times the thinking output of all humanity combined. Wu also suggested that groundbreaking applications of this technology have not yet materialized. He compared current AI coding tools to the light bulb, describing them as early, foundational applications rather than the transformative innovations that lie ahead.