Opening
Alibaba Cloud announced a comprehensive upgrade to its full-stack technology system on May 20, 2026, at the Alibaba Cloud Summit, positioning itself for the Agentic AI era. The company introduced new products including the Qwen Cloud product website, the Pangu M890 self-developed AI chip integrated into the Panjiu AL128 supernode server, and the Qwen3.7-Max flagship model. This shift reflects a fundamental change in cloud computing's primary users: as AI agents operate 24 hours continuously with infinite AI and cloud demands, Alibaba Cloud is restructuring its entire technology stack from bottom-layer chips, Agentic Cloud infrastructure, models, to inference platforms. According to company executives, token-based AI revenue is poised to replace ECS (Elastic Compute Service) as Alibaba Cloud's largest product line, marking a transition from traditional cloud services to AI-driven consumption models. Over the past five months, Alibaba Cloud's daily average token revenue has grown approximately 15-fold, signaling the acceleration of this transformation.
Pangu AI Chip Series and Hardware Infrastructure
Alibaba Cloud released an aggressive chip roadmap centered on the Pangu M890, a next-generation training-and-inference unified AI chip with performance three times that of the previous-generation Pangu M810E. The Panjiu AL128 supernode server, powered by the M890 and equipped with the self-developed ICN Switch 1.0 interconnect chip, enables 128 AI chips to function as a single computing unit with peer-to-peer latency below 150 nanoseconds, addressing mass concurrent inference and large model training demands in agent scenarios.
Alibaba Cloud disclosed the Pangu chip series roadmap, committing to release one new generation annually over the next two years, with planned releases of the Pangu V900 and Pangu J900 chips offering increased computational capacity. To date, the Pangu series has shipped a cumulative 560,000 chips, serving over 400 customers across more than 20 industries.
Token Revenue Growth and Market Position
Alibaba Cloud holds the largest share in the large model MaaS (Model-as-a-Service) market. The company reported that token revenue experienced significant acceleration beginning this year, with the previous period characterized as merely a "prologue." According to company executives, daily average token revenue increased approximately 15-fold over the past five months, reflecting the rapid adoption of AI services. This growth trajectory indicates that token-based metrics are becoming the primary measurement unit for Alibaba Cloud's revenue expansion.
Cloud Product Redesign for Agent Workloads
Alibaba Cloud is fundamentally redesigning its cloud products to operate as agent-native systems. Traditional cloud products were designed with human operators in mind, but agent workloads exhibit characteristics incompatible with conventional cloud computing: irregular elasticity, short lifecycles, and instantaneous scaling. The company has undertaken Skill-ification, MCP (Model Context Protocol) transformation, and CLI (Command Line Interface) standardization of all cloud products, enabling agents to invoke cloud capabilities as standardized function calls.
This redesign philosophy prompted Alibaba Cloud to launch Qwen Cloud, a new product website separate from the main Alibaba Cloud portal. The website's homepage displays a single agent-readable prompt instruction. All model service capabilities are encapsulated as standardized Skills and CLI tools, allowing agents to parse the instruction, acquire full platform capabilities, and autonomously invoke required functions. According to company leadership, the core judgment underlying this initiative is that future cloud computing's primary users will be AI agents rather than human engineers, necessitating a fundamental shift in product architecture and interaction design.
Qwen3.7-Max Model Capabilities and Performance
Alibaba Cloud released Qwen3.7-Max as its latest flagship large language model. In the Arena global large model blind test rankings, Qwen3.7-Max ranks first among Chinese models, surpassing Kimi-K2.6, DeepSeek-v4-pro, and GLM-5.1, while approaching the performance levels of GPT, Claude, and Gemini's strongest models.
A production case study demonstrates the model's autonomous capability beyond standard benchmarking. On the Pangu M890 chip—a platform the model had never encountered during training—Qwen3.7-Max independently completed a production-grade AI compute kernel implementation and optimization task over 35 hours using only a task description, achieving 10 times the performance of the official reference version. This case exemplifies a fundamental shift in model design objectives: from optimizing for human preference alignment to optimizing for autonomous task completion. According to Alibaba's large model division leadership, Qwen3.7-Max was designed to serve as the intelligent core of agents, equipped with autonomous planning, continuous iteration, and cross-tool collaboration capabilities.
Alibaba Cloud has identified AI Coding (AI-driven programming) as a primary application domain. The company notes that AI Coding generates new applications while simultaneously modernizing legacy code accumulated over decades. Company executives highlighted that AI Coding targets software development and external outsourcing expenditures that were previously outside traditional cloud service revenue capture, representing a significant expansion of addressable market opportunity.
