Breaking

Alibaba Unveils New AI Chip and Targets More Than 20 GW of Data-Center Capacity by 2032

September 21, 2026 at 23:21 EDT

Infra & Chips
Foundation Models

Alibaba Group used the opening day of the 2026 Apsara Conference in Hangzhou to present a larger AI-infrastructure push: a new T-Head accelerator described as China's most powerful AI chip and a goal of operating more than 20 gigawatts of data-center capacity by 2032. The plan ties Alibaba's chip roadmap, cloud build-out and Qwen model strategy into a more vertically integrated bid for the next phase of AI demand.

Sep. 22, 2026 · Alibaba

Alibaba Builds Its AI Stack Around Power

A new T-Head accelerator and a more than 20 GW data-center target by 2032 put chips, energy and cloud capacity at the center of Alibaba's AI strategy.

Capacity Goal, Drawn To Scale

20+ GWtarget by 2032

Each block represents 1 GW of planned data-center power capacity. The number shows why Alibaba's AI roadmap is now as much about electricity and infrastructure as it is about model performance.

Performance Roadmap

1xM890

3xV900 target

The earlier roadmap targeted roughly three times M890 performance for V900.

144 GB

M890 on-chip memory

560,000+

Zhenwu chips shipped before the new push

650+

external M890 customers cited later

The Full-Stack Bet

T-Head chips→Interconnects→Alibaba Cloud→Qwen workloads

The strategy is to co-design hardware, networking, cloud services and models instead of treating accelerators as standalone parts.

Why it matters

Proprietary silicon could lower inference costs and reduce dependence on foreign GPUs for Alibaba's cloud and Qwen workloads.

What remains open

Public pricing, power use, independent benchmarks and broad software compatibility remain limited.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.