In the past two years, the development path of AI has changed. The capabilities of large models are soaring, and inference speeds are also being optimized. Global capital and institutions are betting on this trend. However, behind this wave of centralization, another path is being explored—DeAI (Decentralized AI Training and Inference). This architecture targets two hidden issues in current AI development: first, the blind trust problem caused by over-reliance on centralized models; second, the scalability vulnerabilities brought by single points of failure. From a Web3 perspective, DeAI could be a key to breaking AI monopolies and building an open ecosystem. Whether this decentralized approach can become the mainstream direction in the future is worth ongoing attention.