With the rapid adoption of generative AI, data centers are facing major challenges in power supply, cooling, and site availability. Simply continuing to expand large, centralized data centers is no longer sustainable, and their limits are becoming clear. In this session, we explain the real-world vision of decentralized data centers based on the characteristics of AI workloads—centralized training and distributed inference—and the concept of integrating power and networking through a watt–bit collaboration perspective. We also introduce a practical approach to building AI infrastructure that is fast to deploy, high-density, and scalable, using IIJ’s modular edge data center, DX edge Cool Cube, as a concrete example.