Alibaba unveils Zhenwu V900 AI chip and 5–10T Qwen roadmap at Apsara

At Apsara Conference in Hangzhou, Alibaba announced the Zhenwu V900 AI training and inference processor, said Qwen 4 is in training with Qwen 4.5 and Qwen 5 aimed at 5–10 trillion parameters, and set a 20 GW Alibaba Cloud capacity target by 2032.

CHINA desk — Alibaba is pairing a domestic AI accelerator roadmap with multi-trillion-parameter Qwen plans while U.S. export controls still shape chip supply.

T-Head, Alibaba’s chip design unit, unveiled the Zhenwu V900. The press release calls it an AI processor for training and inference. Training means teaching a new model. Inference means running a trained model so it can answer. The page says the chip has 216 GB of memory — it calls that GPU memory, the memory on the accelerator itself — and 1,200 GB/s of bandwidth between chips. GB/s is gigabytes per second, how fast data can move from one chip to the next. The same paragraph says the chip natively supports FP8 and FP4. Those are thinner number formats, 8 bits and 4 bits. The company says that mix lets the chip train at higher precision and run inference more cheaply. File the memory, the bandwidth, and the FP8 / FP4 line as Alibaba’s. This desk did not time a job.

The performance line is a company comparison, not a score this desk ran. Alibaba says the V900 delivers three times the performance of the Zhenwu M890, which the page says was released in May. Three times is T-Head’s comparison with its own previous chip. It is not a comparison with Nvidia, and it is not an outside benchmark. The page says mass production and commercial release are scheduled for the first quarter of 2027. That date is a plan. The V900 is not a chip you can buy on 22 Sep 2026.

Alibaba also described a supernode server that puts the V900 together with three other T-Head parts: an ICN Switch, a Panmai SmartNIC, and a Zhenyue SSD controller. A supernode, in this release, is one large machine that ties many accelerators together so they can work as a cluster. A switch moves data between chips. A SmartNIC is a network card with its own processor. An SSD is a fast storage drive, and Zhenyue is the name of the chip that controls that drive. The company says this server can support a cluster of up to 500,000 cards. Up to 500,000 is Alibaba’s scale claim. This desk did not count a rack.

On models, Alibaba said Qwen 4 is currently in training. Qwen is Alibaba’s family of large language models. The company said the later Qwen 4.5 and Qwen 5 series are projected to scale to 5 to 10 trillion parameters. A parameter is one learned number inside a model. Five to ten trillion is the size the company is aiming at for models that are not out yet. The headline’s 5–10T means that range. Qwen 4 in training is not a model this desk can call. Qwen 4.5 and Qwen 5 are a roadmap, not a release.

Eddie Wu, CEO of Alibaba Group, said the target is for the global data center capacity operated by Alibaba Cloud to surpass 20 GW by 2032. A gigawatt, written GW, is a billion watts. Here it is how much electricity those data centers would be built to draw. Surpass 20 GW by 2032 is his target. It is not a reading of how much capacity Alibaba Cloud has today. This desk did not meter a data hall.

The press release says Zhenwu AI chips have been serving more than 650 customers, in industries it names as automobiles, finance, large language models, embodied intelligence, energy, and manufacturing. Embodied intelligence, here, means AI that drives a physical machine, such as a robot or a car. More than 650 is the company’s count of customers for the Zhenwu line. It is not a count of V900 chips already sold. This desk did not audit the list.

Leave the 3× line as T-Head’s claim against the M890. Leave Q1 2027 as the commercial-release schedule on the page. The press release does not print an independent benchmark, and this filing does not invent one. U.S. rules that limit which advanced chips can be sold into China are the backdrop for why a domestic accelerator roadmap matters. This card does not report a new export-license decision.

Plain English for the rest of the card: T-Head = Alibaba’s chip design unit, also called 平头哥. Zhenwu V900 = the new AI chip for training and inference. M890 = the previous Zhenwu chip, which the page says was released in May. 216 GB = the memory on the chip. 1,200 GB/s = how fast data can move between chips. FP8 / FP4 = 8-bit and 4-bit number formats. supernode = a large server that ties many accelerators into one cluster. ICN Switch = T-Head’s chip-to-chip switch. Panmai SmartNIC = T-Head’s smart network card. Zhenyue = T-Head’s SSD controller. 500,000 cards = Alibaba’s claimed cluster size, not a count this desk made. parameter = one learned number inside a model. 5–10T = 5 to 10 trillion parameters, the Qwen 4.5 / Qwen 5 size target. 20 GW = 20 gigawatts of data-center power capacity, Eddie Wu’s 2032 target. Q1 2027 = when the company says the V900 reaches mass production and commercial release.

PRIMARY here: Alibaba Cloud’s 22 Sep 2026 press release, “Alibaba Unveils Roadmap on Full-Stack AI Strategy from Chips, Cloud Infrastructure, Models to Agents,” datelined Hangzhou — Tier A PRIMARY, the company’s own record. TechNode’s same-day article is the news account of the T-Head unveil and the source of the booth photograph, not a second company newsroom. The Apsara roadmap, the Zhenwu V900 training-and-inference processor, the 216 GB of memory, the 1,200 GB/s inter-chip bandwidth, the FP8 and FP4 line, the 3× claim versus the Zhenwu M890 released in May, the Q1 2027 mass-production and commercial-release schedule, the supernode server with the ICN Switch, Panmai SmartNIC, and Zhenyue SSD, the up-to-500,000-card cluster, Qwen 4 in training, the Qwen 4.5 and Qwen 5 projection of 5 to 10 trillion parameters, Eddie Wu’s 20 GW-by-2032 target, and the more-than-650-customer line are Alibaba’s. NOT claimed: an independent chip benchmark, that the V900 is shipping today, that 3× is a comparison with Nvidia, a current data-center wattage, a U.S. export-license outcome, that Qwen 4.5 or Qwen 5 is available to use, that this desk sat in the hall or ran the chip, a stock tip, or investment advice. Distinct from the already-filed cxmt-g5-dram-mass-production, qwen-image-2.1, and xiaomi-mimo-v2-6.