# Alibaba Wants a 10-Trillion-Parameter AI Model, and It Just Built the Chip to Train It

By Rafiqul Islam Rabbi · AI · Published Thu, 24 Sep 2026 15:18:00 GMT
Source: The Current Tribune — https://currenttribune.com/article/alibaba-zhenwu-v900-chip-ai-model-trillion-parameters

Alibaba just told the world exactly how big it wants to go, and the number is almost hard to take seriously. At its annual cloud conference in Hangzhou this week, the company unveiled a new AI training chip and, in the same breath, said it’s building toward a model with five to ten trillion parameters. For context, that’s roughly two to four times the size of Alibaba’s current flagship, Qwen3.8 Max, which itself already ranks among the largest production models anywhere.

## A Chip Built to Feed a Very Hungry Model

The new silicon is called the Zhenwu V900, and Alibaba says it delivers roughly three times the performance of its predecessor, which only shipped four months ago. That’s an aggressive iteration cycle for custom AI silicon, and it tells you something about how fast the pressure is building inside Chinese cloud computing right now.

The V900 gets bigger memory and more bandwidth than the chip before it, and Alibaba is positioning it as a dual-purpose part that handles both training and inference rather than specializing in one or the other. The headline number, though, is scale: a single cluster built on the V900 can reportedly support up to 500,000 cards working in concert to train a frontier-class model. Mass production and commercial availability are pegged for the first quarter of 2027, which means this is very much a “trust us, it’s coming” announcement rather than a shipped product review.

### Why the Timing Matters

None of this is happening in a vacuum. Export controls have made it harder for Chinese firms to get their hands on Nvidia’s top-tier training chips, and that’s forced a wave of homegrown alternatives into the spotlight. Huawei’s Ascend line currently leads the domestic market, and Alibaba is also competing for attention against Cambricon and Baidu’s Kunlunxin chips. Building your own silicon used to be a nice-to-have for a company like Alibaba. Now it’s closer to a survival requirement if you want to keep scaling your AI ambitions without depending on hardware Washington might restrict further.

## The Trillion-Parameter Question

Here’s where it gets genuinely interesting. Qwen3.8 Max, Alibaba’s current top model, runs at around 2.4 trillion parameters. Jumping to a five-to-ten trillion parameter successor isn’t a routine version bump — it’s a different category of undertaking, one that requires enormous compute, enormous training data, and enormous patience from investors who are watching burn rates climb across the entire industry.

Alibaba CEO Eddie Wu framed the push in blunt terms, telling attendees the company expects “significant growth in annual AI chip shipment volumes” as demand for training capacity keeps outpacing what’s currently available. That’s corporate-speak for: we think everyone, including us, is going to need a lot more chips than they currently have.

### Where the Money Is Coming From

Alibaba isn’t shy about the price tag attached to this ambition. The company has already committed roughly $53 billion over three years to AI infrastructure, and a chunk of that is being backstopped by a $10.2 billion share offering out of Hong Kong specifically earmarked for AI investment. That’s a lot of capital chasing a model that doesn’t exist yet, built on a chip that isn’t shipping yet, in a market where DeepSeek and ByteDance are both racing to claim the same territory domestically.

Investors, for what it’s worth, liked what they heard. Alibaba’s Hong Kong-listed shares climbed 5.1% on the announcement, and the broader Hang Seng Tech Index ticked up 0.8% alongside it — a sign that the market is reading this less as a moonshot and more as a credible, funded roadmap.

## How This Fits the Bigger Picture

Step back and the pattern is unmistakable. Every major Chinese tech company is now simultaneously building its own chips and its own frontier models, largely because relying on outside hardware has become a strategic liability rather than just a cost consideration. Huawei is doing it. Baidu is doing it. Cambricon is trying to carve out its own lane. Alibaba’s move is notable mainly for the scale of what it’s promising and the speed at which it’s iterating on hardware — three generations of meaningful chip improvement in well under a year, by its own account.

It’s also a reminder that the “AI arms race” narrative isn’t just an American story about OpenAI, Anthropic, and Google slugging it out. There’s a parallel race happening inside China, with its own hardware constraints, its own capital structure, and its own timeline, and it’s advancing on a schedule that doesn’t wait for the rest of the world to catch its breath.

### The Skepticism Worth Holding Onto

A few caveats are worth keeping in mind before treating any of this as settled fact. Alibaba hasn’t shipped the V900 yet — commercial availability is more than a year out. The five-to-ten trillion parameter model has no confirmed release window at all. And “three times the performance” is a company claim, not an independently benchmarked one. Chip announcements at conferences like this tend to be aspirational marketing as much as engineering milestones, and the gap between a keynote demo and a generally available product can be wide.

That said, Alibaba has a track record of actually delivering on its cloud infrastructure promises, even when the timelines slip. Qwen has gone from a domestic curiosity to a model taken seriously by international developers in a relatively short window, and the company’s cloud division has consistently been profitable enough to fund exactly this kind of bet.

## What This Means

If Alibaba pulls this off, it puts a Chinese company in position to field one of the largest production AI models on the planet, trained on domestically designed silicon, at a moment when the entire industry is questioning whether bigger models are even the right direction to keep pushing. That tension — between the “just scale it up” camp and the “efficiency and smaller models” camp — isn’t going away, and Alibaba just planted its flag firmly in the first group.

For the AI chip market specifically, this is one more data point suggesting the days of Nvidia facing no serious competition are over, at least in markets where its top hardware isn’t fully available. Whether the Zhenwu V900 lives up to its billing won’t be clear until it actually ships in 2027. But the direction of travel — more parameters, more custom silicon, more capital thrown at the problem — isn’t in doubt, in China or anywhere else in this industry right now.