September 22, 2026
Alibaba triples the capacity of its AI accelerator: V900 has 216 GB of memory versus 144 GB for M890
Zhenwu V900 is three times more powerful than Alibaba's previous chip, but it will only be available to buy from the first quarter of 2027.
China’s most powerful domestically developed AI accelerator has already been built in hardware. You can’t buy it yet.
The Zhenwu V900 accelerator was made by Alibaba’s semiconductor division T-Head and unveiled at the Apsara conference on September 22. Mass production and commercial launch are scheduled for the first quarter of 2027.
It is a bid for an in-house stack for Alibaba’s own models. The previous chip in the line was used to run China’s first language model with more than 2 trillion parameters, while Zhenwu chips already serve more than 650 customers in automotive, finance, energy, and manufacturing.
Before and now. In May 2026, the M890 carried 144 GB of HBM and 800 GB/s of inter-chip bandwidth. The V900 carries 216 GB and 1200 GB/s, natively supports FP8 and FP4, and delivers three times the performance. The updated supernode combines the V900 with the ICN Switch, Panmai network card, and Zhenyue SSD controller, and scales into a cluster of up to 500,000 cards.
The money is already committed: 380 billion yuan (about $52 billion) for cloud and AI over three years. CEO Eddie Wu is targeting more than 20 GW of Alibaba Cloud data-center capacity by 2032, and also acknowledged that the industry’s medium- and long-term demand far exceeds supply capacity.
What is available today. The hardware is not for sale, but Alibaba’s models are. The Coding Plan in Alibaba Cloud Model Studio costs $50 per month on the Pro plan and provides 6000 requests per 5 hours, 45,000 per week, and 90,000 per month. It connects through an OpenAI-compatible or Anthropic-compatible endpoint with a key like sk-sp-xxxxx, and works with Claude Code, Cursor, Cline, and four other clients.
The model lineup is growing alongside it. Qwen 4 is already training, with Qwen 4.5 and Qwen 5 on the roadmap at 5–10 trillion parameters, versus 2.4 trillion for the current flagship Qwen 3.8 Max. Over a month, it completed 33 self-improvement cycles and raised its Artificial Analysis score from 40 to 45.
At Apsara, Eddie Wu himself compared AI coding to the light bulb in the age of electricity: an early application, not a breakthrough product. In his words, truly breakthrough products for the era of machine intelligence have yet to appear.
The Panjiu Supernode server platform on V900 is promised for the first quarter of 2027, with the next-generation Zhenwu J900 due in March 2028.
Source
