Alibaba Unveils Zhenwu V900 AI Chip, Qwen 4 Foundation Models, and 20GW Datacenter Plan

At its annual flagship Apsara Conference 2026 in Hangzhou on September 22-23, Alibaba Group unveiled an expansive full-stack artificial intelligence strategy headlined by the Zhenwu V900, its most powerful proprietary AI accelerator to date. Presented by CEO Eddie Wu, Alibaba confirmed the chip features 216GB of high-bandwidth memory with cluster scalability reaching 500,000 cards, while formally announcing the training of the Qwen 4 foundation model series, on-device Qwen Intelligence, and a commitment to expand Alibaba Cloud data center capacity beyond 20 Gigawatts by 2032.

Zhenwu V900 Architecture and 500,000-Card Cluster Scalability

Developed by Alibaba chip design subsidiary T-Head (PingTouGe), the Zhenwu V900 is engineered to provide a sovereign, high-throughput compute foundation for training and inferencing trillion-parameter models. The silicon incorporates 216GB of ultra-dense high-bandwidth memory (HBM) and features proprietary interconnect fabrics engineered to link up to 500,000 accelerator cards within a single unified computing domain. Slated for commercial mass production in the first quarter of 2027, the accelerator establishes a domestic alternative capable of supporting sustained frontier AI workloads.

Hardware & Infrastructure Specifications

• Zhenwu V900 Accelerator: Proprietary AI processor developed by T-Head packing 216GB of integrated HBM capacity.
• Massive Cluster Fabric: Engineered to scale up to 500,000 accelerators per unified supercomputing cluster with low-latency optical fabrics.
• Production Roadmap: Sampling to key domestic cloud customers ahead of full commercial volume production in Q1 2027.
• On-Device Qwen Intelligence: Lightweight end-to-end inference stack enabling mobile chipsets to execute multi-step agentic reasoning locally.
• Global 20GW Commitment: Alibaba Cloud pledged to expand its global operational datacenter power capacity beyond 20 Gigawatts by 2032.

Qwen 4 Foundation Models and 20GW Hyperscale Cloud Infrastructure

On the software and foundation intelligence layer, Alibaba announced that its next-generation Qwen 4 family (led by Qwen 4 Max) has entered active pre-training. Alibaba outlined an ambitious AI roadmap centered on Recursive Self-Improvement (RSI), projecting models scaling toward 10 trillion parameters as part of its path toward Artificial Superintelligence (ASI). To meet this exponential compute demand, Alibaba Cloud is investing heavily in next-generation liquid-cooled modular data centers across Asia, Europe, and the Middle East.

Product Profile Sheet