Alibaba showed Zhenwu V900 — a 2027 chip nobody outside can score yet
T-Head unveiled the Zhenwu V900 at Apsara on September 22, claiming about three times the M890, 216 GB of memory, and clusters of up to 500,000 cards. Alibaba says mass production and commercial release are in the first quarter of 2027. Reuters relayed the company calling it China's most powerful AI chip.
What happened
On September 22, 2026, at Apsara — Alibaba Cloud’s annual conference in Hangzhou — introduced the , a training and inference accelerator. Alibaba’s own note says the V900 delivers three times the performance of the , which came out in May, with 216 GB of memory, 1,200 GB/s of bandwidth between chips, and native support for and . The same note says an upgraded supernode server, tying the V900 to Alibaba’s switch, smart network card, and SSD controller, can support a cluster of up to 500,000 cards. CNBC, Reuters, and Bloomberg carried those company figures. Alibaba’s note does not publish a peak figure, so this story leaves that number out.
Reuters relayed Alibaba calling the V900 China’s most powerful AI chip. Chief executive Eddie Wu is the person on stage for that line. It is Alibaba talking, not an independent ranking against or .
The date that matters is the calendar. Alibaba says mass production and commercial release land in the first quarter of 2027. Until then this is a stage product with a ship date, not a box a lab can rent and benchmark. Alibaba also says Zhenwu chips already serve more than 650 customers, across cars, finance, models, energy, and factories. That covers the family, especially M890-era deployments. It does not prove the V900 is in the wild.
The same morning was bigger than one chip. Reuters reported Wu’s plan for Qwen models in the 5 trillion to 10 trillion parameter range, and a target for Alibaba Cloud to operate more than 20 gigawatts of data-center capacity by 2032. Hong Kong-listed shares rose 5.1% in Reuters’ account. That is market color, not silicon proof. The press note also sketches 2027 Yitian server CPUs, including a first in-house microarchitecture. That is a roadmap slide beside the accelerator, not evidence the V900 already runs.
A same-day check of public Geekbench listings, recorded for this Brief, found zero Zhenwu results. No fresh MLPerf submission showed up either. So there is no honest public score to set next to Nvidia or Ascend.
Why it matters
Every English wire will lead with “three times faster” and “China’s most powerful.” The useful read is simpler. A large Chinese cloud company is building its own AI chips because top Nvidia parts are hard to buy, and it is promising real hardware next year. That is strategically serious. It is also still soft evidence until someone outside Alibaba can rent a V900 and publish numbers.
What to watch is whether the first-quarter 2027 date sticks, whether Alibaba Cloud instances actually expose the chip to outsiders, and whether any public benchmark appears after silicon ships. Until then, treat the ratio, the memory, and the half-million-card cluster as company claims with a calendar attached.
The memory story running beside this one is CXMT’s fifth-generation DRAM, which is phone memory, and Samsung’s reported HBM4 yield. The capacity bind on the other big Chinese accelerator path is Huawei’s Ascend queue.