HANGZHOU, China — The chip Alibaba unveiled at its annual Apsara Cloud Conference on Monday has a name drawn from ancient mythology — Zhenwu, the dark warrior — and specifications that arrive as a direct answer to a question American policymakers have avoided for three years: what happens when export controls on advanced semiconductors are comprehensive enough to force China’s most capable companies to build their own?
T-Head, Alibaba’s chip design subsidiary, presented the Zhenwu V900 AI accelerator in Hangzhou, the eastern coastal city where the company was founded. Chief executive Eddie Wu described it as “the most powerful AI chip in China today.” The V900 delivers three times the performance of its predecessor, the Zhenwu M890, with 216 gigabytes of on-chip memory and 1,200 gigabytes per second of inter-chip bandwidth, specifications that allow it to scale into clusters of up to 500,000 units for frontier model training. Alibaba, which raised $10.2 billion in Hong Kong’s largest technology share sale earlier this year to fund this infrastructure push, said it expects to operate more than 20 gigawatts of computing capacity by 2032, according to Nikkei Asia.
The V900 is not yet in production. T-Head said mass manufacturing will begin in the first quarter of 2027. Performance claims for AI accelerators routinely diverge from measured results under real workloads, and Alibaba has not said which foundry will manufacture the chip or at what process node. Those are the questions that will determine whether the dark warrior lives up to the mythology. What is already clear is that the chip exists at all — a fact that carries weight given that US chip restrictions were designed, in part, to ensure it would not.
When the United States imposed sweeping export controls on advanced semiconductor manufacturing equipment and high-performance AI chips in 2022 and expanded them through 2024, the stated goal was to preserve a capability gap wide enough that China could not close it within a meaningful timeframe. The Zhenwu V900 is the third generation of chips T-Head has shipped since those restrictions took effect. The same dynamic played out at Huawei, where restrictions on Nvidia hardware produced the Atlas 960E SuperPoD, a 4,096-processor AI cluster built entirely from domestic components. China’s MIIT released its 15th Five-Year Plan two weeks ago mapping $532 billion in AI infrastructure investment built explicitly on domestic chips as export controls tighten.
Alongside the chip, Wu teased Alibaba’s next model frontier. Its most powerful current model, Qwen3.8-Max, carries 2.4 trillion parameters in a mixture-of-experts architecture with a one-million-token context window. Wu said Alibaba intends to train a new model at the 10-trillion-parameter scale, an ambition that would require exactly the kind of clustered training infrastructure the V900 is built to support.
For Alibaba, the announcement also serves a domestic commercial function. Chinese cloud providers have faced constraints on their ability to offer the most powerful AI services to domestic customers because those services require GPU-class hardware that restrictions have made difficult to obtain in quantity. A domestically produced chip that scales to 500,000 units gives Alibaba the capacity to sell AI training to Chinese companies that cannot route their workloads through data centers with access to Nvidia’s current generation. The US-China AI safety dialogue agreed Sunday between Treasury Secretary Bessent and Chinese Vice Premier He Lifeng explicitly left semiconductor export controls outside its scope, meaning the competitive dynamic the V900 is built to navigate will still be in place when the chip enters production next year.
There is a limit to what the announcement can resolve. The chip has been described by the company that built it as the most powerful in China, not the most powerful available anywhere. Nvidia’s latest data center processors remain ahead on publicly available specifications. Whether the 500,000-unit cluster architecture translates to the training efficiency that makes frontier models economically viable will not be known until V900 deployments are measured under real conditions. Wu did not say which models Alibaba plans to train on the chip first, or when results would be available. The 10-trillion-parameter model remains a roadmap, not a delivery date.
What is not in doubt is that China now has a company willing to put its name on a chip of this specification and commit it to mass production. Two years ago, that company was still several generations of export-restricted development away from making that claim.
