Alibaba's Zhenwu V900 triples the M890 in 2027, and Qwen 4 is in training (2026)

Alibaba opened its annual Apsara conference in Hangzhou on September 22 with two announcements: the Zhenwu V900, a training-and-inference accelerator from its chip-design unit T-Head, and a Qwen roadmap that runs to models several times the size of today's flagship. 1 Chief executive Eddie Wu presented both, and the V900 reaches mass production and commercial release in the first quarter of 2027. 2

Signal	Confirmed detail	What it means for you
What was announced	The Zhenwu V900 carries 216 GB of GPU memory and 1,200 GB/s of inter-chip bandwidth, handles FP8 and FP4 arithmetic, and delivers three times the throughput of the M890 it replaces. 1	A specification with a date, so procurement planning starts on paper.
What moved	T-Head's May roadmap placed the V900 in the third quarter of 2027, at the same memory and bandwidth figures. 3	The ship date came forward by two quarters.
The models	Qwen 4 is in training; the Qwen 4.5 and Qwen 5 series that follow are projected at 5 trillion to 10 trillion parameters, against 2.4 trillion for the current Qwen 3.8 Max. 1 2	A two-to-four-fold scale-up, with no release date attached.
The self-improvement claim	Over one month of automated runs, Qwen 3.8 Max completed 33 improvement cycles and rose from 40 to 45 on the benchmarking firm Artificial Analysis's intelligence index. 1	Alibaba ran the loop and scored the result.
The capacity target	Alibaba Cloud aims for more than 20 GW of global data-centre capacity by 2032, and says its AI supernodes reach commercial scale this quarter. 1 2	Wu named supply, not demand, as the limit on expansion. 2

What the chip changes, and when

The Zhenwu line's previous flagship, the M890, launched in May with 144 GB of memory and 800 GB/s of inter-chip bandwidth. 3 The V900 raises both, and T-Head reports the three-fold throughput gain against that predecessor alone. 1 The server built around it pairs the processor with T-Head's switch chip, network card and SSD controller, and holds supernode clusters of up to 500,000 cards. 1 In May, T-Head put Zhenwu shipments at 560,000 chips and more than 400 customers; Alibaba Cloud now counts more than 650. 1 3 Huawei, the closest domestic rival, announced computing architecture the previous week that it says scales to a million processors. 4

What runs today

The first production V900 parts arrive in 2027 and Qwen 4 has no release date, so the models Alibaba already hosts remain the ones to build on. 1 This conference added several of those: Qwen3.8-LiveTranslate for simultaneous interpretation, an updated Qwen-Audio-3.1 speech suite, and Qwen Intelligence, an agent platform that Alibaba is offering to smartphone makers. 1 Qwen-Image 3.1, an image model aimed at design and e-commerce work, is promised for later this year. 1

The second piece of evidence for the self-improvement claim sits outside the model's own scores. In a chip-design experiment, Alibaba says, Qwen 3.8 Max spent more than 60 hours generating production-grade bus modules through more than 10,000 calls to electronic-design-automation tools, cutting the module area by 42%. 1 Wu called the progress on models that improve themselves "meaningful". 2