← back

Price

—

Travel

—

Access

—

Social

—

The rapid growth of foundation models and AI agents is turning AI infrastructure into a “token factory”. Our key objective is to deliver intelligence at the lowest cost, energy, and latency per token. This talk will discuss how end-to-end hardware–software co-design can improve the efficiency, scalability, and reliability of AI systems. I will first present our work on heterogeneous large-scale training and inference, including heterogeneous computing resource management, disaggregated prefill-decode serving, and automated kernel optimization for diverse hardware platforms. As AI moves beyond generating and reasoning over digital tokens toward perceiving, deciding, and acting in the physical world. I will then discuss the emerging infrastructure requirements of embodied intelligence. I will introduce our proposed embodied AI infrastructure that spans simulation, real-world data collection, reinforcement-learning training, cloud-edge execution, and robot-side inference. This talk will conclude with our perspective on AI infrastructure that can continuously evolve with models, agents, hardware, and real-world applications.

• Talks