Back to Home

Qwen’s Next Flagship Trains on Alibaba’s Own Chips

Alibaba is training its next-generation Qwen model on chips designed in-house by its T-Head semiconductor unit, CEO Eddie Wu told analysts on the company's quarterly earnings call, the clearest signal yet that China's most-downloaded AI model family is decoupling from Nvidia at the training stage, not just at inference.

The disclosure, reported by Nikkei Asia and AFP after the Thursday call, extends T-Head's role across both halves of AI computing: building models and serving them. Alibaba's Bailian platform, sold internationally as Model Studio, is already running inference on Chinese-made silicon, Wu said.

What Wu Said, Exactly

Wu framed the move as a cost and coordination play. Pairing homegrown chips with homegrown models, he argued, lowers computing costs and speeds up application development, while giving Alibaba a full-stack position inside China's domestic computing supply chain. The strategy links the two largest line items in Alibaba's AI budget: the Qwen model family and the hardware underneath its cloud.

  • Training: the next flagship Qwen is being trained on T-Head accelerators.
  • Inference: Bailian serves models to customers on Chinese-made chips.
  • Procurement: after 2027, "the vast majority of our new AI chip supply should consist primarily of our own domestic T-Head chips," Wu said, adding that Alibaba will still buy other domestic chips to cover internal inference demand.
  • Pipeline: T-Head's Zhenwu V900, unveiled in September, targets mass production in the first quarter of 2027.

The Rule That Forced the Pivot

The timing is not accidental. A new U.S. Commerce Department rule requires a license for data centers that rent large amounts of AI compute to China-headquartered companies, including facilities in Singapore, Malaysia and Indonesia, where Chinese labs had been renting Nvidia capacity by the rack. Existing operators received a 90-day transition window.

Wu said the restrictions will affect Alibaba's products and services in 2027 and create "very big uncertainty" for the global supply chain. TSMC and Samsung have already slowed some shipments to Chinese customers while licenses are processed, which turns a policy document into a procurement deadline for every Chinese AI lab.

For Alibaba specifically, the arithmetic is blunt: if the rental route closes in 2027, the only way to keep training frontier models at cadence is to own the accelerator. Wu's answer is that T-Head becomes the default supplier to Alibaba's own clusters, with domestic merchant silicon as a second source.

The Business Behind the Chips

Alibaba reported revenue of 271.5 billion yuan ($40.7 billion) for the quarter ended September 30, up 10% year over year, with its AI cloud business growing 53%. Those are the numbers that pay for fabs and tape-outs, and they explain why the chip story landed on an earnings call rather than at a developer conference.

The model side is doing the heavy lifting on adoption. Qwen has racked up more than 3 billion downloads in six months across 460-plus open-source models, with 300,000-plus derivatives built on top of them, according to the company. Hugging Face's August count put Google at 418 million downloads and Meta at 227 million over the comparable period.

  • Downloads: 3 billion-plus for Qwen in six months.
  • Open catalogue: more than 460 released models, 300,000-plus derivatives.
  • Closest rivals: Google at 418 million, Meta at 227 million downloads.
  • Flagship: Qwen4-Max, at 2.4 trillion parameters, ships closed-weight.

Nothing changes this week. The weights you pull from Hugging Face still target the same GPUs and the same inference stacks. What changes is upstream: future Qwen checkpoints will be co-designed with T-Head silicon and with the constraint that training has to happen on hardware Alibaba can actually buy at scale.

That constraint cuts both ways. Co-design can buy real efficiency, since the model team can tune kernels and memory layouts against a known accelerator. It can also narrow the search space, because architectural bets that suit T-Head hardware may not suit the Nvidia clusters most of the open-weight community runs on. Nobody on the call published a benchmark for the V900, so the efficiency claim is currently a promise, not a measurement.

There is a second, quieter risk: a training stack built on domestic chips inherits domestic tooling. Alibaba has spent years building compilers and frameworks to hide that gap, but Qwen's release cadence has been fast partly because the team could lean on a mature software ecosystem. Rebuilding that ecosystem is a multi-year project that runs on the same calendar as the export rules.

What to watch from here is less about the call and more about what shows up in public artifacts.

  • Benchmarks, not statements: the first credible V900 training-throughput numbers, ideally on a public workload.
  • The next Qwen checkpoint: whether it lands near Qwen4-Max quality and how its open-weight siblings change size or licence.
  • Merchant silicon orders: how much Alibaba buys from Cambricon, Huawei and others alongside T-Head.
  • 2027 procurement guidance: the first quarter where Alibaba reports capex with domestic chips as the majority line item.

The strategic point is simple. Export controls work by making frontier training expensive and slow. Alibaba just told investors it intends to make frontier training routine on hardware Washington cannot license or throttle. If the next Qwen lands at or near the current flagship, the controls lose most of their bite; if it slips a generation, the gap becomes visible in public benchmarks within a quarter. Either way, the era where a Chinese lab's roadmap depended on a purchase order from San Jose is closing.

Comments

No comments yet. Be the first to share your thoughts!