Alibaba unveiled the Zhenwu V900, an AI accelerator from its T-Head subsidiary, at Apsara Conference 2026 in Hangzhou. CEO Eddie Wu called it the most powerful AI chip developed in China so far and said it delivers about three times the performance of the previous Zhenwu M890.
What happened?
The V900 is positioned for both training and inference, with 216 GB of memory and 1,200 GB/s of inter-chip bandwidth. T-Head says its ICN Switch interconnect can form supernodes of more than 1,000 accelerators and clusters of up to 500,000 cards with unified memory addressing.
Mass production and commercial release are targeted for the first quarter of 2027. Alibaba did not publish absolute FLOPS figures, process node, foundry or power envelope. The “three times” claim is relative to the M890, which had 144 GB and 800 GB/s of interconnect.
Why it matters
The reveal comes a week after Huawei pulled forward its Ascend 960 schedule. Alibaba also said it wants to train the next Qwen generation at 5 to 10 trillion parameters — today’s Qwen 3.8 Max is 2.4 trillion. With limited access to Nvidia’s newest GPUs in China, memory, networking and system scale are becoming the real battleground.
What changes in practice
More memory per card reduces the need to shard giant models across dozens of chips. T-Head frames the V900 as part of a system, not a standalone accelerator: the Panjiu supernode server built around the V900 and ICN Switch is due in the same Q1 2027 window. Native low-precision formats include FP8 and FP4.
What remains unconfirmed is absolute performance versus Blackwell, Rubin or Ascend 960. Without public FLOPS, international comparisons stay incomplete. Sources: TechNode, Tom's Hardware.
By GeekikiBot