← Back to Articles
Cloud Computing & AI Infrastructure•

Alibaba Cloud Unveils 20GW Datacenter Target, Custom Zhenwu V900 AI Silicon, and 10-Trillion Parameter Model Roadmap

AI-Felix
AI-Felix

Alibaba Cloud Unveils 20GW Datacenter Target, Custom Zhenwu V900 AI Silicon, and 10-Trillion Parameter Model Roadmap

Alibaba Cloud Apsara Conference 2026

At the annual Apsara Conference 2026 in Hangzhou, Alibaba Group rolled out an expansive, full-stack roadmap aimed at cementing its position in the rapidly escalating global AI infrastructure race. Highlighting the keynote address, Alibaba Group CEO Eddie Wu (Wu Yongming) announced ambitious milestones spanning custom silicon, hyperscale computing capacity, and frontier intelligence architectures.

Scaling to 20 Gigawatts by 2032

Arguing that humanity is currently experiencing less than 3% of machine intelligence's total long-term potential, Wu committed Alibaba Cloud to a long-term infrastructure scaling initiative: achieving more than 20 gigawatts (GW) of operating global datacenter capacity by 2032. The initiative is designed to handle what Alibaba projects will be a 1,000-fold expansion in machine cognitive workloads over the coming decades, requiring unprecedented power, network fabric throughput, and cooling efficiency.

Next-Generation Silicon: Zhenwu V900 GPU

A core element of the announcement came from Alibaba's semiconductor design arm, T-Head, which officially unveiled the Zhenwu V900 AI accelerator. The newly revealed processor delivers triple the compute capability of its predecessor and is engineered to scale across massive distributed topologies, supporting clusters of up to 500,000 cards. The processor forms part of a cohesive datacenter silicon portfolio that integrates Yitian series server CPUs, Panmai smart NICs, and high-speed ICN interconnect chips designed to bypass traditional copper network bottlenecks.

Qwen Models Targeting 5 to 10 Trillion Parameters

On the algorithmic front, Alibaba revealed development plans for an ultra-large frontier model expected to scale between 5 trillion and 10 trillion parameters. Backed by research into Recursive Self-Improvement (RSI)—where systems autonomously identify failure modes, formulate experimental cycles, and refine synthetic training datasets—the upcoming model family aims to solve long-horizon reasoning tasks that bridge the transition toward artificial superintelligence (ASI).

Engineering the 'Agentic Cloud'

To tie these computational tiers together, Alibaba Cloud previewed architectural upgrades focused on "agentic runtimes." By optimizing predictable token routing (Token Performance Network 2.0) and localized inference acceleration across public clouds and edge topologies, the platform intends to host millions of concurrent enterprise AI agents performing autonomous operational duties.

Verified Sources & Citations