Alibaba claims new Zhenwu chip triples performance and can cluster 500,000 units

Alibaba claims new Zhenwu chip triples performance and can cluster 500,000 units

The company previews the Zhenwu V900 chip, Qwen models with 5–10 trillion parameters — up to four times the current flagship Qwen 3.8 Max — and a target of 20 gigawatts of data center capacity in Alibaba Cloud by 2032.

Alibaba used its annual Apsara conference in Hangzhou on Tuesday, September 22, to wrap three big promises into a single message: a next-generation AI chip, a plan for a leap in model scale, and a long-term data center capacity target. The company said, according to Reuters, that it is developing an AI model up to four times larger than its current flagship, and unveiled a chip it itself describes as "China's most powerful AI chip" (Reuters via MSN).

The shares rose — but by how much is unclear. Reuters reports that the Hong Kong-listed shares climbed 5.1 percent on Tuesday. Dow Jones Newswires/WSJ puts the gain at 3.1 percent, against an increase of 1.4 percent in the Hang Seng Tech index (MarketScreener/Dow Jones), and a third, syndicated account says "around 3 percent" (MSN). The difference may reflect different measurement points in the trading day, but the sources conflict, and it is worth keeping in mind.

The chip: Zhenwu V900, delivered in 2027 — if all goes to plan

The new Zhenwu V900 chip is developed by T-Head, Alibaba's chip division. Alibaba says it triples performance compared with its predecessor, the M890, which was launched in May, and that the chips can be linked into clusters of up to 500,000 units for training frontier models — figures given by CEO Eddie Wu, according to Bloomberg (Bloomberg).

Mass production and commercial launch are set for the first quarter of 2027. Wu said, according to Reuters, that the company expects "significant growth" in annual AI chip shipments. A secondary source states that Alibaba's existing Zhenwu chips are already used by more than 650 customers, but that figure is not supported by the other accounts and should be read with that caveat.

The most important caveat, however, concerns the performance claim itself: "China's most powerful AI chip" and the tripling of performance are Alibaba's own characterizations, not independent measurements. No third-party benchmark figures appear in the source material.

The model plan: 5–10 trillion parameters as a target

Wu said, according to Reuters, that the Qwen team plans to train a new model with 5 to 10 trillion parameters, to handle "more complex tasks with longer time horizons" in the pursuit of artificial superintelligence — that is, an intelligence surpassing human capability.

To put the number in context: the current flagship, Qwen 3.8 Max, has 2.4 trillion parameters, so the target amounts to a fourfold increase at the upper end. Alibaba says, according to Reuters, that Qwen 4 is now in training, while Qwen 4.5 and Qwen 5 are expected to reach this size range.

This is a plan, not a demonstrated capability. The sources themselves point out that parameter count is a rough measure of size and capacity, not a guarantee of how good the models will be. Wu's talk of superintelligence is framing from a commercial launch event — and should be read as such.

Infrastructure: 20 gigawatts by 2032

The third promise is the most concrete and also the most capital-intensive. Wu set a target, according to Bloomberg, of 20 gigawatts of data center capacity for Alibaba Cloud by 2032, citing "exponentially rising demand for AI" (Bloomberg via Yahoo Finance). Bloomberg notes at the same time that Alibaba has already pledged more than $53 billion over three years to AI buildout — meaning the 20 GW target is a continuation and escalation of an existing effort, not a new starting point.

Wu described customer demand for AI as "exceptionally robust" and said it is accelerating Alibaba Cloud's revenue growth — but also that supply chain constraints are slowing the pace of buildout, Reuters writes. He further claimed that the company's M890 AI supernode already handles inference for models with more than 2 trillion parameters, a capability he said only "a handful of companies globally" possess (Reuters via MSN). This too is a self-serving claim without independent verification in the sources.

Why now

The announcements come in a charged environment. The US has tightened export restrictions on Nvidia's advanced AI chips to China, making domestic alternatives such as T-Head's chips strategically important for Chinese companies. At the same time, competition over Chinese AI infrastructure is intensifying: Huawei made its own large-scale claims about roughly a million processors just a week earlier.

What is actually verified — and what is not

The distinction matters here:

Verified in the source material: That Alibaba made the announcements at Apsara on September 22; that the company itself states the specifications and targets; that the shares rose on Tuesday (the size varies between sources).

Not verified: The tripling of performance, the "China's most powerful" title, the parameter targets, the 20 GW capacity plan as achievable, and the claim about the M890 supernode's inference capability. All of these are Alibaba's own figures, relayed by Reuters and Bloomberg without independent verification. That is normal for corporate announcements, but it means readers should treat the numbers as intentions and marketing, not results.

What to watch

The nearest test is the first quarter of 2027: will the Zhenwu V900 actually enter mass production and commercial delivery, with specifications close to those stated? Then: do the parameter targets materialize in Qwen 4.5 and Qwen 5, and how do the models actually perform on independent benchmarks — regardless of parameter count? Finally, a smaller but instructive point: even the market reaction, the most "objective" part of the story, is unclear, with sources reporting anything from around 3 to 5.1 percent. Clearer consensus on the share move — and Alibaba's own primary materials from the conference — would complete the picture.