Alibaba's Zhenwu V900 Claims 3x Uplift; 10T-Param Model Next

For AI engineers and researchers, Alibaba's Hangzhou announcements put real numbers on China's accelerating frontier: a claimed 3x performance leap from the Zhenwu M890 to V900, and a training target of 5–10 trillion parameters for the next Qwen model. The practical question is whether the V900 can maintain throughput and memory bandwidth at scale for trillion-parameter training runs, not just marketing benchmarks.

On September 22, 2026, at its annual flagship conference in Hangzhou, Alibaba rolled out the Zhenwu V900 AI chip, which CEO Eddie Wu called "the most powerful AI chip in China today." The company says the V900 delivers three times the performance of its previous-generation Zhenwu M890. The announcement came one day before Chinese leader Xi Jinping was scheduled to arrive in Washington on September 23 for a state visit with U.S. President Donald Trump, where AI competition, trade, and tariffs are expected to be major themes. The chip is aimed squarely at frontier AI training and inference. Alibaba said the Zhenwu chips are deployed across its data centres, providing compute to both its own model teams and external cloud customers. A 3x generational uplift, if it holds up in real-world workloads, would materially reduce the cost and time required to train trillion-parameter models, strengthen Alibaba Cloud's position in China's domestic AI infrastructure market, and offer a homegrown alternative as U.S.-China technology rivalry pushes Chinese companies toward domestic silicon. The company did not publish detailed benchmarks, power draw, or interconnect specifications, so the performance claim remains a vendor statement rather than an independently verified result.

“

For comparison, Chinese AI company Moonshot says its Kimi K3, released in July 2026, is the world's largest open model at 2.8 trillion parameters.

Beyond silicon, Alibaba outlined an ambitious model roadmap. It plans to train a new AI model at the scale of five to 10 trillion parameters. That is a leap from its current Qwen3.8-Max, the most powerful model in its Qwen series at 2.4 trillion parameters. For comparison, Chinese AI company Moonshot says its Kimi K3, released in July 2026, is the world's largest open model at 2.8 trillion parameters. A 5-10 trillion parameter model would be more than three times the size of Moonshot's current open leader, though parameter count alone does not determine quality; architecture, data quality, and post-training alignment matter just as much in practical capability.

Alibaba also cited "exponentially rising" demand for AI computing. The company expects to have more than 20 gigawatts of AI computing capacity by 2032. That target is enormous: the source reporting notes SpaceX has approximately 1.4 gigawatts of AI compute, which would make Alibaba's 2032 ambition roughly 14 times larger, though the two are not directly comparable across cloud versus dedicated training clusters. Such capacity expansion signals that Alibaba sees multi-year structural demand from both its own frontier model development and cloud customers building and serving AI applications.

The timing of the announcement is deliberate. U.S. AI leaders, including Anthropic CEO Dario Amodei, have been warning that China's AI development poses a threat and calling for a slowdown. With Xi and Trump scheduled to meet, Alibaba's product cadence becomes a geopolitical signal: China can continue advancing AI capabilities despite external pressure. The meeting may produce new restrictions, tariffs, or a competitive framework, but for now, Alibaba is demonstrating technical momentum that it hopes will shape both domestic policy support and the global AI race narrative.