Alibaba unveils AI chip and plans to double data center capacity by 2032

Alibaba’s aggressive expansion into AI infrastructure signals a shift in competition among China’s tech giants to control the compute capacity underpinning the industry’s growth. The company’s new homegrown chip and data center roadmap could reshape how Chinese firms compete with Western peers for dominance in AI hardware and models.

• Alibaba introduced Zhenwu V900 chip delivering three times the performance of its May-released predecessor, the Zhenwu M890.
• The company plans to expand Alibaba Cloud’s global data center capacity to exceed 20 gigawatts by 2032, roughly double current industry commitments.
• Qwen 4 model is in training with planned Qwen 4.5 and Qwen 5 series.

• 3x Zhenwu V900 performance gain over the Zhenwu M890 released in May
• 20 GW Target global data center capacity by 2032, compared to current footprint
• 650+ Customers already using Zhenwu AI chips across automotive, finance, and energy sectors
• Q1 2027 Scheduled mass production and commercial release date for Zhenwu V900

Alibaba shares rose approximately 3 percent in Hong Kong on Tuesday (September 22) after the company unveiled a portfolio of AI infrastructure investments at its Apsara Conference in Hangzhou, according to reporting by CNBC. The announcements—spanning proprietary chips, cloud capacity expansion, and foundation model development—position the Chinese technology giant as a major player in the global competition for AI compute resources alongside Western peers racing to build out data center networks.

Zhenwu V900 targets commercial deployment in early 2027

Alibaba’s chip design unit T-Head unveiled the Zhenwu V900, an AI training and inference processor designed to handle demanding machine learning workloads. The processor supports multiple data precisions including FP8 and FP4 formats, allowing operators to reduce inference costs while maintaining performance across both high-precision model training and ultra-low-precision inference tasks. T-Head has already deployed its existing Zhenwu chip family to over 650 customers across automotive, finance, large language models, embodied intelligence, energy, and manufacturing sectors.

The V900 is scheduled for mass production and commercial release during the first quarter of 2027.

Data center expansion to 20 gigawatts reflects trillion-parameter model ambitions

Alibaba committed to expanding Alibaba Cloud’s global data center capacity to surpass 20 gigawatts by 2032, a scale that reflects the company’s expectation of exponential growth in AI compute demand. Eddie Wu, chief executive of Alibaba Group, framed the expansion within a broader thesis about machine intelligence development: the total volume of machine thinking currently represents less than 3 percent of all human thinking, and if that proportion eventually scales to 1,000 times human capacity, “machine thinking still has an enormous growth runway.”

The 20-gigawatt target places Alibaba’s ambitions well ahead of comparable Western commitments. Nvidia earlier this month outlined plans to work with Australian partners to support up to 2 gigawatts of AI infrastructure in the country by 2027, while Meta in July announced a 1-gigawatt data center facility in Alberta, Canada expected to cost roughly 9 billion dollars and take up to three years to complete. Huawei, Alibaba’s domestic peer, last week unveiled new AI infrastructure it said can scale to as many as one million processors, underscoring the intensity of infrastructure competition among Chinese technology firms.

Qwen 4 training underway as model series scales toward 10 trillion parameters

Alibaba’s next-generation Qwen 4 foundation model is currently in training, with planned Qwen 4.5 and Qwen 5 series.

The company also detailed progress in what it calls Recursive Self-Improvement (RSI) driven by empirical feedback mechanisms. Over one month of fully automated training runs spanning pipeline design, data validation, iterative experimentation, and error diagnosis, Qwen 3.8-Max completed 33 iterative cycles, with the updated model boosting its Artificial Analysis score from 40 to 45. In a chip design experiment, the model underwent over 60 hours of self-improvement across the entire design lifecycle, making more than 10,000 electronic design automation tool calls to produce production-grade chip bus modules—reducing chip area by 42 percent with no compromise in performance.

Alibaba also released updates across multimodal models, including Qwen 3.8-LiveTranslate for simultaneous interpretation that reduced latency nearly 20 percent from 2.8 to 2.3 seconds, and Qwen-Audio-3.1-TTS-Next, a next-generation audio generation model for creating cinematic soundscapes. The company launched Qwen Intelligence, a business-facing agentic solution optimized for smartphones that provides phone makers with access to a Qwen-powered agent platform for next-generation AI phones capable of handling complex cross-application tasks.

The truly groundbreaking products of the Machine Intelligence era have not yet arrived.

Eddie Wu, CEO of Alibaba Group

The BlockWest read. Alibaba’s data center expansion timeline extends to 2032—a decade out—raising questions about how the company will finance and operationalize 20 gigawatts of capacity when demand patterns for large language models remain uncertain and power infrastructure constraints persist in many regions. The chip roadmap positions Alibaba to reduce reliance on external semiconductor suppliers, but the V900’s commercial viability depends on adoption velocity among the 650 existing customers and new market entrants.

Market participants should monitor whether Alibaba’s data center buildout meets its 2032 timeline and at what capital cost, particularly given power availability constraints in China and overseas. The commercial release of the Zhenwu V900 in the first quarter of 2027 will provide the first concrete test of whether the chip gains traction beyond Alibaba’s internal use and its existing customer base.

BlockWest is a news publication. Nothing here is investment advice. Read our disclaimer and editorial policy.