Alibaba Group has unveiled its full-stack AI strategy at the Apsara Conference, spanning foundation models, proprietary silicon, cloud infrastructure and enterprise agent deployment.
At its annual Apsara Conference, Alibaba announced a major update to its full-stack artificial intelligence strategy, unveiling next-generation Qwen foundation models proprietary computing chips, an agentic cloud architecture and mobile platform solutions designed to scale AI across global industries.
Under Chief Executive Officer Eddie Wu, Alibaba Cloud outlined a long-term infrastructure target to expand its global data center capacity beyond 20 gigawatts by 2032 to support rising compute demand. The corporate strategy aims to integrate every layer of the AI ecosystem from physical silicon and cloud networking to foundation models and consumer device integration.
Alibaba revealed that its next-generation Qwen 4 model family is currently under active development, with future iterations Qwen 4.5 and Qwen 5 planned to scale between 5 trillion and 10 trillion parameters.
Demonstrating advancements in recursive self-improvement, Alibaba highlighted how its Qwen3.8-Max model completed 33 automated improvement cycles over a month, boosting its performance scores without human intervention.
In a practical chip design experiment, the AI model spent over 60 hours executing self-improvement protocols across the engineering lifecycle, making more than 10,000 electronic design automation calls to produce production-grade chip bus modules that reduced physical chip surface area by 42 percent without compromising performance.
For end-user devices, Alibaba introduced Qwen Intelligence, a business platform that enables smartphone manufacturers to embed AI agents capable of handling complex, cross-application tasks on mobile phones.
On the hardware front, Alibaba’s chip design division, T-Head, introduced the Zhenwu V900 AI processor for large-scale model training and inference. Slated for mass production and commercial launch in the first quarter of 2027, the Zhenwu V900 delivers three times the performance of its predecessor featuring 216 GB of GPU memory and 1,200 GB/s inter-chip bandwidth while supporting FP8 and FP4 data precisions.
To support massive computing clusters, T-Head integrated the Zhenwu V900 into an upgraded supernode server architecture alongside the ICN Switch, Panmai SmartNIC, and Zhenyue SSD controller chip.
The integrated platform is engineered to support supernode clusters containing up to 500,000 cards. Additionally, Alibaba published its roadmap for the next-generation Yitian 720 and Yitian 730 server CPUs, which are scheduled for market release in 2027.
Alibaba Cloud introduced structural updates to its cloud platform, organizing its infrastructure around an “agentic cloud” framework split into three operational pillars: model, harness and context.
The restructured cloud architecture centers on an AI Native Cloud layer that incorporates significant updates to Alibaba Cloud’s Platform for AI, including a new Cloud Parallel File Storage (CPFS) system offering hundred-terabyte-per-second throughput that cuts AI storage costs by up to 69 percent, alongside HPN 8.0 Pro networking capable of supporting 100 petabits of bandwidth across 130,000 high-speed ports.
Building on this core, the Agent Native Cloud layer introduces AgentCore, an enterprise platform designed to build, run, and manage autonomous AI agents throughout their operational lifecycle while utilizing the Agent Security Center for compliance and safety oversight.
Finally, the Context Engine integrates Agent Context to connect enterprise documents, chat logs, business databases, and multimodal information, providing AI agents with real-time operational context and long-term memory while reducing overall token consumption by up to 67 percent.
In tandem, Alibaba upgraded its OpenLake multimodal data lakehouse platform which unifies structured, unstructured, vector and streaming data. The unified architecture reduces query response times by 40 percent and lowers total operational costs by 38 percent compared to legacy data management systems, establishing a comprehensive technological framework for global enterprise AI deployment.
