Alibaba Group has unveiled an ambitious full-stack artificial intelligence roadmap at its annual flagship technology event, the Apsara Conference, held in Hangzhou. Spanning proprietary AI chips, cloud infrastructure, foundational large language models, and agentic ecosystems, the tech giant is making aggressive moves to solidify its leadership in the global intelligence economy. 

Scaling the Qwen Frontier and Recursive Self-Improvement

At the core of Alibaba's model strategy is the continuous evolution of its Qwen foundation family. While the next-generation Qwen 4 is currently undergoing intensive training, the company outlined its long-term trajectory toward the Qwen 4.5 and Qwen 5 series, which are projected to scale up to an unprecedented 5 to 10 trillion parameters. 

A major highlight of the conference was Alibaba’s breakthrough in Recursive Self-Improvement (RSI). Demonstrating empirical success, models like Qwen3.8-Max have autonomously executed cycles of pipeline design, data validation, and error diagnosis to optimize their own performance. Remarkably, Alibaba also utilized AI models to assist in chip design; through thousands of automated Electronic Design Automation (EDA) tool calls, the AI successfully optimized chip bus modules, shrinking the physical chip area by 42% without sacrificing computing performance. 

Furthermore, Alibaba is heavily expanding its multimodal and application-specific ecosystem. Updates include the real-time translation tool Qwen 3.8-Live Translate, audio generation models, and upcoming solutions like Qwen-Image 3.1 tailored for high-speed creative and e-commerce workflows. For mobile hardware manufacturers, the newly introduced Qwen Intelligence platform enables the creation of cross-application AI assistants on smartphones. 

Proprietary Silicon: The Zhenwu V900 Processor

To power these massive workloads independently, Alibaba's semiconductor subsidiary, T-Head, introduced its most powerful AI processor to date: the Zhenwu V900. Offering triple the performance of its predecessor (the M890), the V900 features 216GB of high-speed memory and massive inter-chip bandwidth. Engineered to natively support various precisions down to FP4, a single cluster built on these processors can scale up to 500,000 chips. Mass production and commercial rollout are slated for the first quarter of 2027. 

Alibaba also laid out a clear roadmap for its Yitian Central Processing Unit (CPU) series. The upcoming Yitian 750, scheduled for release by 2029, will feature proprietary core architecture designed to interface directly with Zhenwu AI accelerators, ensuring tightly integrated hardware performance across servers and supernode clusters. 

Agentic Cloud and Massive Infrastructure Expansion

Addressing the operational demands of the AI agent era, Alibaba Cloud introduced its comprehensive Agentic Cloud architecture. Structured across model, execution, and context tiers, the framework optimizes storage efficiency, reduces token consumption by up to 67%, and integrates high-throughput data lakehouse solutions like OpenLake. Network latency and bottlenecks are heavily mitigated by Alibaba's proprietary HPN 8.0 Pro networking architecture, supporting 100-petabit scales.

To sustain this exponential growth, Alibaba Group CEO Eddie Wu announced a massive infrastructure target: Alibaba Cloud aims to expand its global data center capacity to exceed 20 gigawatts (GW) by 2032. Emphasizing that mid-to-long-term industry demand heavily outstrips current global supply chains, Wu framed the initiative as a crucial step toward preparing for an era where machine intelligence and autonomous thinking will vastly scale past human capacity.