Alibaba has unveiled a new AI accelerator, the Zhenwu V900, which it calls the most powerful AI chip in China today. The company also laid out a roadmap for Qwen models far larger than anything it currently runs. The announcements were made this week at its Apsara technology conference in Hangzhou and mark China’s most concrete step yet towards building advanced AI without relying on American chips.
Zhenwu V900: Specifications and Scale
Alibaba says the Zhenwu V900 offers three times the performance of its predecessor and comes with 216 gigabytes of memory and high inter-chip bandwidth, specifications that matter most for training large models.
The bigger claim concerns scale. According to Alibaba, a computing cluster built around the chip can connect up to 500,000 units for both training and inference. Numbers of that order decide whether a country can train frontier models at all, and the claim is aimed directly at the constraint US export controls were meant to impose. Mass production is scheduled for the first quarter of 2027.
Qwen Roadmap: Up to 10 Trillion Parameters
Alongside the chip, Alibaba outlined the direction of its Qwen models. Its current flagship has around 2.4 trillion parameters. The company said Qwen 4 is being trained, while Qwen 4.5 and Qwen 5 are projected to scale to between 5 and 10 trillion parameters. That is several times today’s size and comparable in ambition to the largest models anywhere.
Why the Announcement Matters
Announcing the chip and the model roadmap together is meant to present a self-contained stack: Chinese hardware training Chinese models at frontier scale.
That is the outcome export controls were designed to prevent. Their aim was not to slow Chinese AI in general but to keep it dependent on Western chips and therefore controllable. A credible domestic accelerator that can be clustered in the hundreds of thousands weakens that leverage. In effect, China is signalling through hardware that it intends to route around the restrictions.
Reasons for Caution
Several significant qualifiers apply. These are Alibaba’s own claims, made at its own conference, and “most powerful in China” is a domestic comparison rather than a claim to match Nvidia’s latest chips. The chip does not ship until 2027, so its real-world performance and yield remain unproven. The 5 to 10 trillion parameter figures are targets for unreleased models, not delivered capability. A roadmap states intent, and intent is cheap to announce.
The specifications also leave one question open: how efficiently 500,000 domestic chips can be made to work together. That is as much a software and networking problem as a silicon one, and it is an area where Western systems still hold an advantage.
The direction is still clear, and it fits other recent moves, with Chinese firms announcing domestic memory and logic chips built around the export controls rather than stopped by them.
What to Watch
Three things stand out. The first is whether the Zhenwu V900 ships on schedule in early 2027 and performs as claimed once independent testers can measure it. The second is whether Alibaba actually trains a multi-trillion-parameter Qwen model on domestic hardware, which would be the real proof. The third is how Washington responds to a major Chinese company openly building the self-sufficient AI stack that export controls were meant to deny it.