Alibaba has lifted the kimono on its Zhenwu V900 AI accelerator, claiming three times the performance while remaining remarkably coy and cute about the numbers to back it up.
According to Tom’s Hardware, Alibaba Group chief executive Eddie Wu showed off the chip at the company’s 2026 Apsara Conference in Hangzhou, calling it “the most powerful AI chip in China today.”
The V900 comes from Alibaba chip-design outfit T-Head and targets AI training and inference. It packs 216 GB of GPU memory and 1,200 GB/s of inter-chip bandwidth, with native FP8 and FP4 support. Alibaba claims the silicon delivers three times the performance of the Zhenwu M890 released in May. Unfortunately, anyone hoping for a dull thing like a FLOPS figure will have to keep hoping.
The company has disclosed no process node, foundry, power consumption or memory-bandwidth figure for the V900. That makes its headline performance claim somewhat difficult to poke with a calculator. Its predecessor, the M890, carried 144 GB of memory and offered 800 GB/s of inter-chip bandwidth. M890 systems could scale to 128 chips per supernode.
T-Head says more than 1,000 V900 accelerators can operate as a single system, while its upgraded supernode architecture can scale to clusters containing as many as 500,000 cards. That is a lot of silicon to put behind a performance figure nobody has been allowed to see.
Alibaba says the V900 will enter mass production and commercial release during the first quarter of 2027. That is considerably earlier than the third-quarter 2027 slot shown on the company’s May roadmap.
As of May, T-Head had shipped more than 560,000 Zhenwu chips to more than 400 external customers. Alibaba now says the chips serve more than 650 customers across industries including cars, finance, large language models, energy, and manufacturing. The outfit expects to refresh the Zhenwu line annually and predicts substantial growth in AI-chip shipments.
Huawei’s Ascend 960PR, due in the third quarter of 2027, is slated for 192 GB of memory, 2.4 TB/s of memory bandwidth and a 2.2 TB/s scale-up interconnect.
The V900 forms part of a much larger plan to feed increasingly enormous Qwen models. Alibaba intends to train future models containing between five trillion and 10 trillion parameters. Its current Qwen3.8-Max has 2.4 trillion parameters, while Moonshot AI’s Kimi K3 uses a 2.8 trillion-parameter base.
Qwen 4 is already in training, while Alibaba expects the Qwen 4.5 and Qwen 5 families to push model scale further. The company said Qwen3.8-Max completed 33 iterative cycles during a month of automated work, raising its Artificial Analysis score from 40 to 45. Alibaba ran a separate chip-design experiment involving more than 10,000 EDA tool calls across more than 60 hours of testing. It claims the exercise reduced chip area by 42 per cent without sacrificing performance.
All those increasingly hungry models will need somewhere to live. Alibaba Cloud wants data-centre capacity of more than 20 GW by 2032, a target which will require rather more than optimistic PowerPoint slides and a few racks in Hangzhou.
Wu conceded that shortages in the AI supply chain “are currently limiting the speed at which we can scale our compute infrastructure.”
For now, the V900 remains months away from commercial hardware, while its promised five trillion to 10 trillion-parameter models have yet to emerge from training.
