On September 22, 2026, Alibaba Group unveiled its new-generation AI chip, the Zhenwu V900, at the Apsara Conference in Hangzhou, along with the training roadmap for the next generation of the Qwen large models. Group CEO Eddie Wu made the announcement personally in his keynote, marking Alibaba's biggest bet yet on in-house AI chips.
What exactly is the Zhenwu V900
The V900 was developed by T-Head, Alibaba's semiconductor arm. According to the company, it is the most powerful AI chip ever designed in China, delivering roughly three times the compute of the previous-generation Zhenwu M890, and built for the two most compute-hungry workloads: large-model training and inference. A single V900 cluster can scale up to 500,000 chips to support training and running frontier models. Alibaba plans volume production and commercial availability in Q1 2027.
A larger compute blueprint was announced on the same stage: Alibaba Cloud said it would push global data-center compute past 20GW by 2032, and that AI super nodes would enter commercial-scale deployment starting this quarter. T-Head also disclosed that cumulative shipments of the Zhenwu line have exceeded 560,000 chips serving hundreds of customers — a figure that counts existing chips; the V900 itself has not yet gone on sale.
Why this year of all years
The direct driver is models demanding more compute. Alibaba confirmed at the event that Qwen4 is currently in training, and the Qwen4.5 and Qwen5 series are expected to reach 5 to 10 trillion parameters; per Reuters, the current flagship Qwen3.8-Max has 2.4 trillion parameters, so the next generation would be two to four times larger. The bigger the parameter count, the harder the compute math for training and inference — making in-house chips a necessity.
The deeper context is on the supply side. U.S. export controls have limited Chinese companies' access to NVIDIA's most advanced GPUs, and domestic cloud providers are all hunting for alternatives: demand for Huawei's Ascend processors has already outstripped supply. By presenting the chip and model roadmaps on the same stage, Alibaba effectively declared that its vertical integration "from chips to models" has taken shape.
The market reacted first
On the day of the announcement, Alibaba's U.S. shares rose about 3% in early trading, while its Hong Kong shares climbed more than 5%, hitting a near one-month high. What investors bought was not a chip but the completeness of the AI story: the group has invested over $53 billion in AI over the past few years, and Alibaba Cloud's AI-related product revenue has posted triple-digit growth for 12 consecutive quarters.
How to read it: three unanswered questions
First, process and foundry. Alibaba did not disclose who will manufacture the V900 or what process node it uses, nor did it provide benchmark details backing the "3x performance" claim. Under export controls, access to advanced process nodes is itself the biggest uncertainty.
Second, the mass-production timeline. A Q1 2027 production target leaves a window of more than a year, during which neither NVIDIA's nor Huawei's product lines will stand still.
Third, the software ecosystem. The other half of the compute story is the software stack: CUDA's lock-in effect remains strong, and whether the Zhenwu delivers its claimed efficiency in real training workloads will depend on subsequent third-party testing.
The V900 is Alibaba's key step in turning "compute self-reliance" from slogan into product roadmap. But whether it is a game-changing variable or just conference-stage numbers will only become clear after volume production in 2027.
