Alibaba has rolled out its latest AI accelerator, the Zhenwu V900, designed by its T-Head division. According to the company, the new chip is three times faster than its predecessor—the M890, which dropped in May. The Zhenwu V900 is built for both training and inference of large language models (LLMs).

Specs and Target Use

The Zhenwu V900 packs 216GB of memory and inter-chip bandwidth of 1200GB/s. It supports FP8 and FP4 compute formats, letting users fine-tune resources for LLM training and deployment.

Mass Production and Supernode Server

Alibaba says mass production and commercial launch for the Zhenwu V900 are slated for Q1 2027. Alongside the chip, they also revealed a new server supernode powered by these accelerators.

Scaling Up

You can build clusters of up to 500,000 chips using these systems.