Alibaba unveils Zhenwu V900 AI chip with more on-package memory than NVIDIA H200, and plans to deploy 20GW of compute by 2032 while developing 4-10 trillion parameter Qwen models.

Alibaba dominated the Apsara Conference in Hangzhou, China today, unveiling ambitious compute deployment plans, detailed specifications for its upcoming Zhenwu V900 chip, and announcing parameters for its next-generation Qwen AI models.

The Zhenwu V900 accelerator, slated for Q1 2027 launch, will feature 216GB of on-package memory—exceeding NVIDIA's H200 GPU at 141GB. The chip is expected to leverage CXMT's HBM3E solution based on aligned volume production timelines. Chip-to-chip interconnect speeds will reach approximately 1.2 TB/s via Alibaba's ICN Switch fabric, delivering roughly 3x the performance of its Zhenwu M890 predecessor while offering native FP8/FP4 support. The V900 will achieve peak performance of around 1.8 PFLOPS at FP16, compared to the M890's 0.6 PFLOPS.

Alibaba's ICN Switch enables up to 1,000 Zhenwu V900 chips to operate as a single logical accelerator. The company claims each V900 cluster can scale to 500,000 chips, yielding an aggregate memory capacity of 108 petabytes. Alibaba is also introducing a rack-scale solution integrating Yitian CPUs, Zhenwu V900 GPUs, ICN interconnects, Pangu network interface controllers, and Zhenyue storage controllers.

Alibaba targets 20 GW of global data-center capacity by 2032, up from undisclosed current levels. Morgan Stanley estimated in September 2025 that Alibaba held approximately 5 GW of capacity at year-end 2026, implying annual additions of 2–3 GW are needed to reach the 2032 target.

Upcoming Qwen 4.5 and Qwen 5.0 models will span 4 to 10 trillion parameters. The company is pursuing recursive self-improvement (RSI), where AI systems help build, code, or train subsequent versions of themselves, unlocking economies of scale.

Alibaba released Qwen-Image-2.1, a text-to-image generation and editing model whose visual stack uses approximately 7 billion parameters across 32 single-stream diffusion transformer layers. The model currently ranks at the top of Arena's open-source image editing models.