Alibaba unveiled 'China's most powerful AI chip' and a plan for a 10-trillion-parameter model

China's push to build advanced AI without depending on American chips took its most concrete form yet this week, and it came from Alibaba.

At its Apsara technology conference in Hangzhou, the company unveiled a new AI accelerator, the Zhenwu V900, which it describes as the most powerful AI chip in China today. It also laid out a roadmap for future models far larger than anything it now runs.

The Chip

Alibaba says the Zhenwu V900 delivers three times the performance of its predecessor and carries 216 gigabytes of memory with high inter-chip bandwidth — the specifications that matter for training large models.

The more striking claim is about scale. Alibaba says a computing cluster built around the chip can link up to 500,000 of them for both training and inference. That is the kind of number that determines whether a country can train frontier models at all, and it is aimed squarely at the constraint US export controls were meant to impose. Mass production is slated for the first quarter of 2027.

The Model Roadmap

Alongside the chip, Alibaba set out where its Qwen models are going, and the trajectory is steep.

Its current flagship holds around 2.4 trillion parameters. The company said Qwen 4 is in training, and that Qwen 4.5 and Qwen 5 are projected to scale to between 5 and 10 trillion parameters — a multiple of today, and comparable in ambition to the largest models anywhere.

Why This Is A Statement, Not Just A Product

The point of announcing the chip and the model roadmap together is to describe a self-contained stack: Chinese hardware training Chinese models at frontier scale.

That is the outcome export controls were designed to prevent — not to slow Chinese AI in general, but specifically to keep it dependent on Western chips and therefore controllable. A credible domestic accelerator that can be clustered in the hundreds of thousands undercuts that leverage. This is China saying, in hardware, that it intends to route around the restriction.

The Reasons For Caution

Several qualifiers belong on this, and they are substantial.

These are Alibaba's own claims, made at its own conference, and 'most powerful in China' is a domestic comparison, not a claim to match Nvidia's latest. The chip does not ship until 2027, so its real-world performance and yield are unproven. And parameter counts of 5 to 10 trillion are targets for unreleased models, not delivered capability — a roadmap is a statement of intent, and intent is cheap to announce.

There is also the question the specifications do not answer: how efficiently 500,000 domestic chips can actually be made to work together, which is as much a software and networking problem as a silicon one, and where Western systems still hold an advantage.

But the direction is unmistakable, and it matches other recent moves — Chinese firms announcing domestic memory and logic chips built around, rather than stopped by, the controls.

What To Watch

Whether the Zhenwu V900 ships on schedule in early 2027 and performs as claimed once independent testers can measure it. Whether Alibaba actually trains a multi-trillion-parameter Qwen on domestic hardware, which would be the real proof. And how Washington responds to a major Chinese company openly building the self-sufficient AI stack the export controls were meant to deny it.