The exponential scaling of Artificial Intelligence (AI) necessitates unprecedented increases in computational power. This soaring demand, paired with the waning momentum of Moore’s Law, leads to surging energy costs and carbon footprints, raising serious sustainability concerns. While optimizing traditional computing methods remains key to addressing the current computational power crunch in the short term, exploring innovative post-Moore computing paradigms presents a transformative pathway toward sustainable AI development.

In this talk, I will present a dual-pathway research program for achieving truly efficient AI. The first pathway focuses on near-term solutions by optimizing traditional digital computing methods, specifically through the development of self-adaptive computer architectures. These architectures introduce strong hardware flexibility to elegantly handle computational irregularities inherent in AI workloads—the core bottleneck in achieving both computational efficiency and algorithmic expressivity. The second pathway holds a long-term ambition, aiming to unlock the untapped but massive computational potential from nature, specifically through the development of DS-Machine. DS-Machine effectively enables nature-powered computing through embodying Dynamical Systems as processors, releasing revolutionary computational efficiency for AI and beyond. The talk will conclude with a discussion on the deployment of this dual-pathway research in promoting real-world scientific discovery, with applications in power grid management and nuclear fusion ignition.