Terrestrial ecosystems are essential to planetary health, sustaining biodiversity, forest resilience, water and energy regulation, and the long-term stability of human and natural systems. We aim to build AI-powered digital twins of Earth’s terrestrial ecosystem by enabling scalable, long-term simulation of ecosystem dynamics across the globe. While process-based ecosystem simulation models have been developed for decades, their high computational cost limits the scalability for large-scale and high-resolution simulations under different scenarios. We develop DeepED, a family of global-scale AI emulators for process-based ecosystem models, that learns to efficiently approximate their key ecosystem dynamics such as vegetation height, biomass, etc. The framework can be used with different AI model backbones and we also develop an uncertainty-aware version based on generative AI that can explicitly report statistical confidence levels of its emulation results. Finally, we also develop physics-guide AI models to integrate field observations over the globe to further refine the simulation quality. Together, these models provide a foundation for scalable terrestrial ecosystem digital twins. To encourage broader community progress, we also release an AI-ready global benchmark dataset and evaluation framework for developing, comparing, and improving ecosystem forecasting at scale.