Alibaba T-Head Launches Zhenwu M890 AI Chip, Unveils 128-Chip Server
Alibaba's semiconductor unit T-Head launched the Zhenwu M890 artificial-intelligence chip for training and inference, saying it offers 144 gigabytes of memory and a threefold performance gain over its predecessor. The company also detailed the Panjiu AL128 server, which uses a custom interconnect chip to link 128 M890 chips in a single rack, underscoring Alibaba's push to build larger domestic AI computing systems.
Alibaba plans annual upgrades to the Zhenwu line, with the V900 scheduled for 2027 and the J900 for 2028. It says the M890 is designed for agentic AI workloads that require large context windows, and that more than 560,000 units have already been shipped to domestic enterprises.
From the sources (11 posts)
@poezhao0605I’ve covered Alibaba Cloud for 11 years. This is the first time I can recall T-Head, Alibaba’s chip unit, taking the main stage at the Alibaba Cloud Summit. Zhenwu AI chip: 560,000 units shipped. 400+ customers. 20+ industries. The signal i
@poezhao0605T-Head revealed a four-generation Zhenwu chip roadmap through 2028. 810E (2024) → M890 (2026 Q2) → V900 (2027 Q3) → J900 (2028 Q3). Memory scales from 96GB to 216GB. Each generation claims 3x performance gains on a proprietary parallel co
@cnbcAlibaba reveals more powerful Zhenwu AI chip, new LLM
@businessAlibaba added a new processor to its expanding AI technology stack, enhancing a push to cover every aspect of artificial intelligence development
@mtsliveSITUATION DETECTED: Alibaba’s chipmaking unit T-Head has announced its latest Zhenwu M890 PPU chip for training and inference workloads, per Bloomberg.
@teortaxestexI really like T-Head style. Just big blocky heat sinks but more importantly… 800 GB/s is not HBM speed, it's die-to-die of their NVLink alternative, ICN Link. So a bit behind NVLink 4 (Hopper). They claim support of a 128-chip ICN node (Hop
@techmemeAlibaba's T-Head unveils the Zhenwu M890 AI chip for training and inference, saying it is particularly suited for agentic tasks, and plans annual upgrades (@luzdingyu / Bloomberg) (Visit Techmeme dot com for the link and full context!)
@cointelegraph🇨🇳 NEW: China's Alibaba unveils its new Zhenwu M890 AI chip built for both training and inferencing, with plans to release annual upgrades at a pace matching Nvidia.
@ftr_investors$BABA Alibaba unveiled its new Zhenwu M890 AI chip, which it says is 3x more powerful than its previous generation, alongside its new Qwen 3.7-Max LLM focused on coding, reasoning and AI agents 🤖🇨🇳
@reutersAlibaba unveils new AI chip in push for domestic alternatives
@wesrothAlibaba semiconductor unit, T-Head, launched the Zhenwu M890 AI chip for training/inference, boasting 144GB of memory and a 3x performance leap over its predecessor. Alibaba plans to match Nvidia’s brutal annual release cycle, scheduling