Alibaba Is Training Its Next Qwen on Its Own Chips

Human Verified

Share to: , and

A glowing teal computer chip on a dark circuit board

Alibaba’s next Qwen model is already training on the company’s own T-Head chips, chief executive Eddie Wu said on the earnings call this week. It is the clearest sign yet that China’s biggest cloud company plans to build its AI future on domestic silicon, not imported hardware.

The Chips Behind the Claim

T-Head has built a portfolio of proprietary chips spanning graphics, processors, storage and networking, Wu said. Supernode instances powered by the Zhenwu M890 chip launched in August and are now being commercialised at scale. Mass production of the next chip, the Zhenwu V900, starts in the first quarter of 2027.

The V900 delivers three times the computing performance of the M890, according to Wu. Supernode clusters powered by the chip will launch at commercial scale in the second quarter of 2027, and a third generation, the J900, follows in 2028. Domestic chips will increasingly support training, inference and agent workloads.

Wu said Qwen is already training its next-generation model on T-Head chips, and the Model Studio platform is already running inference on domestic chips. As T-Head’s capacity ramps up, he said, domestic compute will contribute a higher share of Alibaba’s total compute and become the primary engine of cloud growth.

Why the Timing Matters

The announcement lands days after Washington’s interim rule on remote access to advanced chips, which took effect on November 19. Wu acknowledged that the new export controls will affect next year’s products and services, and that supply-chain uncertainty remains. Building on domestic chips is Alibaba’s answer to that risk.

The numbers behind the pivot are large. AI-related products generated 17.9 billion yuan in quarterly revenue, nearly half of Alibaba Cloud’s external recurring revenue, and their annual run rate has passed $10 billion. Cloud revenue targets for 2030 remain $100 billion in external revenue at margins above 20 percent.

Our Read

Training a frontier model on your own chips is the hardest test of a domestic silicon strategy, and Alibaba is now running it in public. If Qwen’s next model trains well on T-Head hardware, the export-control risk shrinks. If it does not, the flagship model is tied to unproven silicon.

The Apple detail is worth watching too. Apple Intelligence, deeply integrated with Qwen, is set to launch for users in China in December. That puts Alibaba’s models in front of hundreds of millions of iPhone users, and makes the T-Head question a consumer story as well as an infrastructure one.

The market’s first verdict was cool: Alibaba’s shares fell after the call, even as the AI numbers impressed. The cloud business is now the company’s growth engine, but the risk is execution, and mass-producing the V900 on schedule matters more than any statement on a call.

Why It Matters

China’s AI self-sufficiency drive has mostly been about announcements and roadmaps. Alibaba training its next flagship model on domestic chips turns it into a live experiment with a deadline. The result will shape not just Alibaba’s cloud, but how far any Chinese lab can go without foreign silicon.

Background

Alibaba’s chip arm, T-Head, filed for a Hong Kong listing this month, with reports of a $2 billion to $3 billion raise. The same earnings call reported AI cloud revenue up 53 percent and group net income up 213 percent. Qwen4-Max, the current flagship, has 2.4 trillion backbone parameters.

Source: Alibaba earnings call transcript

Share to: , and

Faheem R. Khattak

Senior automation engineer and AI marketing journo covering LLMs, algorithms, analytics, core updates, Bing search, and technical SEO.

0 Comments

Leave a Reply

Optional Take Tag:

🛡️ Spam protection: Honeypot + Time-Trap Active

This site uses Akismet to reduce spam. Learn how your comment data is processed.