At the 2026 Yungong Conference, held from September 22-24 in Hangzhou, Alibaba Cloud's Chief Technology Officer Fei-Fei Li delivered a keynote address outlining the company's strategic vision. Li emphasized that the large-scale deployment of AI agents hinges on three critical pillars: model capability, which dictates the quality of intelligent decision-making; comprehensive and timely contextual awareness, which provides reliable data for operational choices; and the Harness runtime environment, which ensures safe and secure execution for continuous self-improvement, correction, and constraint.
Li announced that the next step involves developing the first true 500,000-card wide-area supernode cluster, to be built in conjunction with the upcoming Xuantie V900 chip. This massive cluster would deliver approximately 1GW of computational power, achieve a bandwidth of 200PB/s, and maintain a communication latency of just 6 microseconds, all designed to support next-generation MoE models with trillions or more parameters.
To drive the scaled adoption of AI agents, Alibaba Cloud has set four major objectives. First, the company aims to transform probabilistic AI into reliable productivity. Since large models fundamentally predict tokens, their decisions carry inherent uncertainty, and the goal is to minimize this unpredictability in serious enterprise workflows. Second, Alibaba Cloud intends to shift from broad-brush usage to fine-grained efficiency. As token consumption continues to surge, the company's optimized agent runtime environment is designed to cut operational costs by up to 70%.
Third, Alibaba Cloud plans to elevate from point experiments to a comprehensive intelligent infrastructure. With an estimated 3 million small and medium-sized enterprises in China, each potentially deploying dozens to hundreds of agents, this could translate to tens of millions of agents operating daily on Alibaba Cloud's infrastructure. Fourth, the vision is to move from isolated agents to intelligent organizations, where single agents performing tasks evolve into multi-agent collaboration, akin to human organizational teamwork, with the goal of raising agent penetration in enterprise workflows to 50% or higher.
It should be noted that all conference transcripts are based on on-the-spot shorthand records and have not been reviewed by the speaker. This content is published for informational purposes only and does not constitute an endorsement of the views expressed or verification of the facts stated.
Comments