Alibaba's next AI chip arrives sooner than it planned in May, and gives no figure for its speed

The Zhenwu V900 has 50 percent more memory than the chip before it and is due in volume in early 2027, two quarters ahead of Alibaba's May roadmap.

26 September 2026 · 3 min read · L2 Silicon · The Pulse, 22 September 2026

What happened

T-Head, Alibaba's chip design unit, unveiled the Zhenwu V900 at Alibaba's Apsara Conference in Hangzhou on 22 September. Alibaba says the chip, built for both training AI models and running them, has 216GB of memory and 1,200GB/s of bandwidth between chips, supports the low-precision FP8 and FP4 number formats, and is scheduled for mass production and commercial release in the first quarter of 2027. It also showed a server that puts the V900 together with its own networking and storage chips, and says a cluster of them can reach up to 500,000 cards. Alibaba says its Zhenwu chips already serve more than 650 customers.

Why it matters

The step up is easiest to see against the chip it replaces. Tom's Hardware reports that the Zhenwu M890 has 144GB of memory and 800GB/s between chips, so the V900 has 50 percent more of each, and that the M890 worked in groups of 128 chips. T-Head says more than 1,000 V900 chips can work together as a single system, TechNode reports. Tom's Hardware also reports that Alibaba's May roadmap placed the V900 in the third quarter of 2027, so it has moved up by two quarters.

Alibaba's chief executive, Eddie Wu, gave the reason in his conference speech: "Global shortages across the AI data center supply chain are currently limiting the speed at which we can scale our compute infrastructure." Huawei said a week earlier that its own next AI chip was ready three quarters early and now targets early 2027 as well, a Chinese outlet reported.

What would change it

Independent tests once the chip ships would show how fast it is. Alibaba says the V900 is far faster than the M890, but gave no computing-speed figure to check that against, so this desk does not repeat the multiple. The other test is delivery: whether mass production starts in the first quarter of 2027 as scheduled.

What we do not know

Tom's Hardware notes that Alibaba gave no figure for computing speed, manufacturing process, factory, power use or memory bandwidth. No price has been announced.

What this changes for you

Alibaba already runs its Qwen3.8 models and Moonshot's Kimi K3 on servers built from the older M890, TechNode reports. If the V900 arrives on time, the next chip under those services has half as much memory again, and comes from Alibaba's own design unit.

Sources

Everything above is written from these. Each line says what that document proves.

1. alizila.com: Alibaba's own account of the Apsara Conference, 24 September 2026: the V900's 216GB of memory, 1,200GB/s between chips, FP8 and FP4, mass production and commercial release in the first quarter of 2027, the new server and clusters of up to 500,000 cards, and more than 650 customers.
2. alizila.com: Eddie Wu's conference speech, 22 September 2026: the V900 unveiling and his sentence on global supply shortages limiting Alibaba's build-out.
3. technode.com: TechNode, 22 September 2026: more than 1,000 V900 chips working as one system, and Qwen3.8 and Kimi K3 running on M890 servers.
4. 163.com: NetEase (163.com), 17 September 2026, the source of this site's Reported item on Huawei's next AI chip moving forward to early 2027.
5. tomshardware.com: Tom's Hardware, 23 September 2026: the M890's 144GB and 800GB/s and groups of 128 chips, the May roadmap placing the V900 in the third quarter of 2027, and the missing computing, process, factory, power and bandwidth figures.

Keep reading