BUSINESS / ECONOMY
Alibaba says next-gen Qwen trained on in-house chips amid intensive roll-out of Chinese AI models
Published: Nov 19, 2026 10:47 PM
Concept picture of AI Photo: VCG

Concept picture of AI Photo: VCG


Alibaba said on Thursday at its fiscal 2027 second-quarter earnings call that its next-generation Qwen artificial intelligence (AI) model is being trained on in-house chips. The company also outlined its third-generation GPU roadmap and revealed that AI-related annual revenue will top $10 billion.

At the earnings call, Alibaba Group CEO Wu Yongming outlined a three-generation GPU roadmap for T-Head, the group's wholly-owned semiconductor and AI chip design subsidiary. The Zhenwu M890 will be fully launched this year, while the V900 and the J900 will see production supply chains advance and enter mass production in 2027.

Wu said the infrastructure based on self-developed domestic computing is becoming the supporting base for Alibaba's full-stack AI technology. It will gradually carry workloads for the group and Alibaba Cloud customers in model training, inference and agentic tasks, and become a foundation for AI revenue growth. He also said that with capacity expansion in 2027-28, the pace of long-term model iteration will be "more controllable."

Alibaba said its total revenue in the third quarter rose 10 percent year-on-year. In particular, AI-related product revenue continued to show strong momentum, reaching 17.92 billion yuan ($2.67 billion) in the third quarter and delivering the 13th consecutive quarter of triple-digit year-on-year growth. Revenue from AI cloud and compute services rose 53 percent year-on-year, according to the company's financial results.

In the past week, other major Chinese AI model developers also announced new products or updates, with competition shifting from parameter scale alone toward cost, post-training optimization and domestic computing adaptation.

On November 12, Baidu released the Ernie 5.5 Flash Preview foundation model, which has 280 billion total parameters, activates 14 billion parameters per token, and supports a context window of up to 1 million tokens. Baidu also launched a large-model trading platform. 

On the application side, Baidu released two models and made them available on Baidu AI Cloud platform through API services for enterprise customers and partners. Baidu said it uses a sparse architecture, unlike the traditional one, to balance performance and cost, and that video generation costs are 50 percent lower than the industry average.

Tencent unveiled AI business updates on Wednesday during its third-quarter financial results release, as AI demand drove cloud service revenue up 35 percent year-on-year.

Tencent's AI agent WorkBuddy topped 10 million daily active users and 100 trillion daily tokens, making it the Chinese AI product with the highest daily large language model (LLM) token consumption, according to Tencent. 

On Monday, Tencent released the Hy4 flash LLM, with 224 billion total parameters, and only 12 billion activated parameters. It supports multimodality and a 1-million-token context. Pricing is 0.5 yuan per million input tokens and 2 yuan per million output tokens, or 0.02 yuan on cache hits. 

Zhipu AI on Tuesday released the GLM-5.3 model. Unlike the usual route of stacking parameters and changing architectures, the firm kept the same base model and attributed the improvement mainly to post-training. Official data showed GLM-5.3's coding capability improved 50 percent over GLM-5.2, with better performance on cybersecurity-related tasks. 

Zhipu also said its model as a service (MaaS) platform has more than 10 million developers. Industry insiders said that this figure means that approximately all Chinese developers use an AI model of Zhipu. 

According to Databricks data as of November 2, GLM-series models accounted for more than 60 percent of token usage among open-weight models.

The wave of AI model releases and updates shows that Chinese AI model developers are simultaneously advancing model iteration, lowering inference costs and driving application adoption. 

Global Times