Alibaba Unveils Zhenwu V900 AI Chip in Hangzhou

Alibaba used the opening day of its Apsara Conference in Hangzhou to introduce the Zhenwu V900, an AI accelerator that chief executive Eddie Wu described as the most powerful AI chip China has today. The company puts it at three times the performance of the Zhenwu M890 it replaces. That figure is Alibaba's own; no independent benchmark has been published.

What the chip claim actually covers

The reporting available so far gives a product name, a predecessor and a multiplier. It does not give clock speeds, memory configuration, a manufacturing partner or a ship date. Who fabricates the V900, and how many will land in Alibaba's own data centers, is not disclosed.

The comparison point matters as much as the number. Three times the M890 is a measurement against Alibaba's previous silicon, not against imported accelerators, so it settles nothing about how the V900 stacks up outside the company's own product line.

Models, chips and cloud as one stack

Wu framed AI models, AI chips and the AI cloud as the three cornerstones of the machine intelligence era, and restated the company's pursuit of artificial superintelligence. Chairman Joe Tsai said Alibaba is "firmly investing in building full-stack AI" so that the technology moves from breakthrough to value creation. Both are statements of intent rather than dated commitments.

A model of 5 trillion to 10 trillion parameters

Alongside the silicon, Alibaba signaled a coming model in the 5 trillion to 10 trillion parameter range. Its current flagship, Qwen3.8-Max, runs at 2.4 trillion parameters; Moonshot's Kimi K3, released in July, sits at 2.8 trillion. No release window was given for the larger model.

Parameter count is a weak proxy for model quality, but it is a direct claim on hardware. That is the simplest reading of why the chip and the model plan shared a stage.

Why the 20 gigawatt figure is a power number

The chip sits inside a build-out plan: Alibaba Cloud aims for more than 20 gigawatts of global data center capacity by 2032. Gigawatts measure electrical draw, not compute, so the target describes how much power the sites can pull rather than how much work they can do. The capital behind that target is not stated in the publicly readable portions of the reporting.

Timing, export controls and supply

U.S. export restrictions keep advanced AI chips and chipmaking equipment out of China, which turns in-house accelerators into a supply problem rather than a branding exercise for Chinese cloud providers. The launch also lands in the days before Xi Jinping's state visit to Washington to meet President Trump. Whether that timing was deliberate cannot be read off the announcements themselves.

What we could not verify

Only one of the three sources checked was readable in full; the Nikkei Asia and South China Morning Post reports stopped after the lede behind a paywall. Investment totals, the manufacturing partner, V900 availability and any market reaction therefore remain unconfirmed here.