Chips & Compute

Alibaba Trains Qwen on Its Own Zhenwu Chips as U.S. Export Controls Bite

Alibaba is training its next Qwen model entirely on in-house Zhenwu chips, opening that infrastructure to Chinese firms, and claims performance on par with Nvidia's China-market H20 in some inference tasks.

“China Accelerates AI Self-Reliance Amid U.S. Export Controls”: Alibaba Expands China’s AI Market With Its Own Chips
“China Accelerates AI Self-Reliance Amid U.S. Export Controls”: Alibaba Expands China’s AI Market With Its Own Chipsblickpixel / pixabay
By Sophie Lindqvist6 min read

Why it matters

  • Alibaba CEO Wu Yongming said on Nov 19, 2026 that Qwen training runs on T-Head's Zhenwu chips, with full Chinese-chip pretraining and post-training planned next year.
  • The Zhenwu 810E processed about 360 tokens per second on the 72-billion-parameter Qwen 2.5 in 8-bit inference, matching Nvidia's H20; at 16 bits the H20 led 565 to 414.
  • T-Head released its SAIL software suite free of charge at the World Artificial Intelligence Conference in Shanghai in July 2026, with source code available to overseas developers.
  • A data center in Shaoguan, Guangdong built with China Telecom in April links 10,000 Zhenwu chips, expandable to 100,000.
  • Mass production of the next-generation Zhenwu V900 begins in Q1 2027; Alibaba CEO Wu expects China's compute shortfall to peak in 2027–2028.

Alibaba is training its next-generation AI model, Qwen, on chips it designed itself, and it plans to run the entire training pipeline — pretraining and post-training — on Chinese-made silicon starting next year. The company's CEO, Wu Yongming, confirmed the shift during an earnings briefing on November 19, according to Nikkei Asia. The move is the clearest sign yet that U.S. export controls are pushing China's largest technology company to build a domestic AI computing market on its own terms.

The backdrop matters here. The United States has restricted not only exports of advanced AI chips to China but also Chinese companies' access to large-scale computing resources rented through overseas data centers. With U.S.-made chips getting harder to obtain even abroad, Alibaba is signaling that it intends to secure the computing power for its core models in-house. Wu's briefing turned that intention into a public commitment.

Where are Alibaba's own chips already running?

The first destination for the Zhenwu chips was Alibaba Cloud's data centers. When the Zhenwu 810E was unveiled in January, it was already running in several Alibaba Cloud installations, each linking 10,000 accelerators. That setup spreads the calculations needed to train large AI models — and to generate their answers — across many chips at once.

In August, Alibaba went further and opened infrastructure built around clusters of the successor chip, the Zhenwu M890, to cloud customers. The company says this infrastructure supports inference for large AI models with more than 2 trillion parameters — the variables a model learns during training. In plain terms: Alibaba isn't just powering its own workloads anymore; it's selling compute built on domestic silicon.

What comes next from T-Head?

Mass production of the next-generation Zhenwu V900 chip is scheduled to begin in the first quarter of next year. T-Head, Alibaba's semiconductor design subsidiary also known as Pingtouge, unveiled the V900 in September. The chip handles multiple computations simultaneously and serves both AI training and inference.

Alongside the V900, Alibaba plans to launch a server system in the same quarter that pairs the chip with its own semiconductors for inter-chip communication, data transfer, and storage. The V900 chips divide the computational workload, while the companion chips exchange results between processors and move and store data.

Alibaba intends to source most of the additional AI chips going into its data centers from T-Head. Wu predicted that China's shortfall in AI computing supply relative to demand will be most severe in 2027–2028, and T-Head's flagship chips will be prioritized for training the next-generation Qwen and improving later models. The company will also allocate T-Head chips to major partners training their own models, and Wu said Alibaba plans to sharply increase T-Head chip production next year to cover its own needs and its partners'.

How does Alibaba solve the CUDA lock-in problem?

Chips alone don't make a market. For partners to adopt Zhenwu silicon, the programs they already use have to run on it. The industry's most widely used tool has been Nvidia's CUDA, introduced in 2006. CUDA lets graphics processing units split up the calculations AI training requires, and Nvidia ships prewritten code for frequently used operations, saving developers from implementing the same functions themselves.

That convenience became a trap. As the volume of CUDA-based code has grown, companies wanting to switch to other chipmakers have faced the burden of adapting everything they've built. Reducing that switching cost is why T-Head released SAIL, its dedicated software suite for Zhenwu chips, free of charge at the World Artificial Intelligence Conference in Shanghai in July. The release covered the entire software stack needed to run AI programs on Zhenwu silicon, and the source code was made available to overseas developers too.

T-Head Vice President Gao Hui said developers using SAIL can reuse existing code on Zhenwu chips by modifying only part of it. For partners that have already trained models on other hardware, that cuts the amount of new code they need to write. Alibaba also plans to improve compatibility between Zhenwu chips and widely used AI development tools.

A quick comparison of the two ecosystems:

  • Nvidia CUDA: released in 2006; supports parallel GPU computing and ships code for common operations; switching away means rewriting CUDA-based programs.
  • T-Head SAIL: released free at the Shanghai AI conference in July 2026; runs AI programs on Zhenwu chips with source code open to overseas developers; existing code can be reused with partial modifications, and compatibility with major AI tools is on the roadmap.

Can a single Zhenwu server run DeepSeek?

Training large models takes more than adapted software — it takes infrastructure that runs thousands of chips together. The flagship example is the installation Alibaba built with Chinese telecom operator China Telecom at a data center in Shaoguan, Guangdong province, in April. The two companies linked 10,000 Zhenwu chips there to support training and inference for models with hundreds of billions of parameters.

China Telecom operates the infrastructure, and the partners named healthcare and advanced materials research among the fields that will receive computing resources. They also left room to expand the Shaoguan installation to as many as 100,000 chips to meet future demand.

The site supports outside models as well. Alibaba Cloud built a common programming interface compatible with major AI development tools so developers can migrate existing programs to Zhenwu servers. Each server carries 1.5 terabytes of memory for AI computing, and Alibaba says that's enough to run both its own Qwen 3.5 and Chinese AI company DeepSeek's V3.2 on a single server without reducing their parameter counts. In Shaoguan, companies in the electronic information and biopharmaceutical sectors use the infrastructure for model training, inference, and tuning models to their own operations.

How does Zhenwu stack up against Nvidia's H20?

For companies planning to run services in Shaoguan, chip speed matters most. What caught the overseas semiconductor industry's attention: the Zhenwu 810E performed similarly to Nvidia's H20 — the chip designed for the Chinese market — on some inference tasks.

Alibaba Cloud's test linked two chips of each type in separate setups and had the 72-billion-parameter Qwen 2.5 generate answers of the same length. In a dialogue test using 8-bit numerical representations, both chips processed about 360 tokens per second. At 16-bit precision, the gap opened: the H20 processed 565 tokens per second while the Zhenwu 810E managed 414.

When Alibaba introduces successor chips, it benchmarks them against its own previous generation. In May, the company said the M890 delivers three times the performance of the 810E, and at the V900 unveiling in September it again cited a threefold improvement. The specifics Alibaba has disclosed concern memory and chip-to-chip bandwidth:

  • Memory per chip rose from 144 gigabytes on the M890 to 216 GB on the V900.
  • Maximum data exchanged between chips per second rose from 800 GB to 1,200 GB.

There are limits to the disclosures. The V900 presentation materials omitted the model name and test conditions behind the claimed "threefold performance" gain, and Alibaba gave no specific figures on how much faster the next-generation Qwen would train.

The gaps in the benchmark data mean outside buyers will judge Zhenwu on what they can verify — the Shaoguan installation, the SAIL toolkit, and the H20 comparison. Whether Alibaba can scale production fast enough to cover its own 2027–2028 crunch and its partners' demand is now the question that decides how far this domestic market grows.

Source: economy.ac

Share this article:

More from Sophie Lindqvist

Sophie Lindqvist

Show full bio

Staff writer covering regulation, policy and their impact on artificial intelligence and machine learning at AI In Context.

823 articles

Related articles

  1. Alibaba unveils Zhenwu V900 AI chip, targets 20 GW cloud buildout by 2032
  2. Huawei and Alibaba Chip Blitz Arms Xi With Leverage Ahead of Trump Summit
  3. Smuggled Chips and Server Loopholes: How China Gets Nvidia GPUs
  4. Chip Scare Drove Alibaba Down 41% This Year. PDD Offers a Cheaper Bet
  5. US export rule forces Southeast Asian data centers to get licenses for AI chips

« Previous articleNext article »