Alibaba’s chip design arm used its Apsara conference in Hangzhou to claim the most powerful AI accelerator built in China, a part its owner says can be clustered half a million strong.
T-Head’s Zhenwu V900 carries 216GB of memory and moves data between chips at 1,200GB/s. It handles FP8 and FP4 arithmetic natively and is aimed at both training and inference. Alibaba says it triples the compute of the earlier Zhenwu M890.
Mass production and large-scale deployment are scheduled for the first quarter of 2027. A successor, the Zhenwu J900, is pencilled in for the third quarter of next year. T-Head also showed a first look at two server processors, the Yitian 720 and Yitian 730, both due in that same quarter, with the 730 built on a fully in-house microarchitecture.
Scale is the pitch. Alibaba says its homegrown ICN switch can link up to 500,000 V900 chips into a single cluster, enough to train and run frontier models of 5 trillion to 10 trillion parameters. The current Qwen3.8-Max sits at 2.4 trillion. A Panjiu supernode built from the V900 and the switch arrives with the chip.
Cloud capacity underpins the plan. Alibaba Cloud targets more than 20GW of data center capacity worldwide by 2032, against a three-year commitment of more than $53B and a $10.2B share issue in August.
CEO Eddie Wu framed the strategy around models, chips and cloud infrastructure. The launch lands days before Xi Jinping meets Donald Trump in Washington.
