Alibaba unveils the chip it calls China's most powerful -- built entirely around the gap Washington's export controls left open

Alibaba CEO Eddie Wu unveiled the Zhenwu V900 at the company's Apsara Conference in Hangzhou on Sept. 22, calling it “the most powerful AI chip in China today” and claiming three times the performance of its predecessor, the Zhenwu M890. The chip is built by Alibaba's semiconductor subsidiary T-Head and is slated for mass production and commercial release in the first quarter of 2027 -- a keynote claim, not yet a shipped product any customer has benchmarked independently.

The specs Alibaba disclosed: 216 GB of HBM memory, 1.2 TB/s of inter-chip bandwidth, support for FP8 and FP4 precision, and an architecture Alibaba says can scale a single cluster up to 500,000 chips. Alongside the chip, Alibaba said it will expand its data-center capacity to more than 20 gigawatts by 2032, and laid out a roadmap for Qwen models scaling from today's flagship, Qwen 3.8 Max, at 2.4 trillion parameters, up toward 5 to 10 trillion parameters across the coming Qwen 4.5 and Qwen 5 generations.

That scale is the direct product of a policy choice made in Washington, not Hangzhou. US export controls keep Nvidia's most advanced accelerators out of China, and the V900 is Alibaba's answer to the resulting gap -- a chip designed to exist entirely outside the components export rules restrict. It follows the same logic behind Huawei's own decision to move its next Ascend chip up by nine months, even though Huawei still can't manufacture enough of its current generation to meet demand: both companies are racing to prove that a full domestic stack -- design, and eventually fabrication -- can substitute for what Washington won't sell.

Markets read the announcement as an Alibaba-specific win rather than a sector-wide one. Alibaba shares jumped 4% to $120.36 in the trading session after the keynote -- by one market analyst's account, the sharpest single-session move the stock had posted in weeks -- while Baidu, which also builds its own AI chips, moved just 0.3% to $92.45, essentially flat. A broader China-internet ETF (KWEB) rose only about 1%. “The flat Baidu reaction tells you the market is treating this as a company-specific win for Alibaba rather than a sector-wide upgrade,” one analyst wrote -- and the same commentary was explicit that the reaction is “a bet on a spec sheet, not shipments.”

“The flat Baidu reaction tells you the market is treating this as a company-specific win for Alibaba rather than a sector-wide upgrade.” — market analyst commentary on the V900 reaction

Alibaba has already committed tens of billions of dollars to the Qwen model family and the compute infrastructure behind it, and the Apsara Conference -- the company's annual flagship cloud-and-AI event in Hangzhou -- is where it has chosen to make that spending legible each year: a chip generation, a data-center target, and a model roadmap, announced together rather than piecemeal. T-Head, the in-house semiconductor unit behind the V900, is the same subsidiary tasked with proving that Alibaba Cloud's AI ambitions don't depend on a supplier Washington can cut off.

The chip announcement lands four days before Trump and Xi meet in Washington, where the US has been pitching a standing AI-incident dialogue built explicitly to leave existing chip export controls untouched. A domestically designed accelerator that Washington had no hand in restricting doesn't fit neatly into that conversation either way -- it's not a violation of anything, and it's not something Washington can loosen or tighten. It's a demonstration that the policy's leverage has a shelf life.

Independent chip analysts will want three things before treating the V900 as more than a slide: a real third-party benchmark against Nvidia's current China-legal offerings and against Huawei's Ascend line, confirmation of yield at the promised cluster scale, and evidence that Alibaba Cloud customers outside Alibaba's own Qwen team are actually training production workloads on it. Keynote specifications have a way of surviving contact with real deployments in rounder, less flattering numbers than the ones announced on stage.

None of it is independently verified yet. The “three times the performance” figure, the 500,000-chip cluster scaling, and the 20-gigawatt target are Alibaba's own claims, made in a keynote, about a chip that won't reach mass production for at least two more quarters. China's chip industry has a recent history of announced capability outrunning delivered capability -- Huawei's own roadmap acceleration, announced days earlier, exists precisely because its current-generation Ascend chip still can't be built in the volume demand requires. The V900's real test isn't the keynote; it's whether Alibaba Cloud customers are training on it a year from now.