Imagine a future where intelligence could be seamlessly integrated into our physical environment, operating autonomously for years on harvested energy. However, realizing this vision faces fundamental challenges in overcoming physics bottlenecks of miniaturized sensing, scaling intelligence across large deployments, and doing so in a sustainable way. This thesis presents research in reimagining computing from first principles to enable sustainable ambient intelligence at an entirely new level of granularity. First, we demonstrate how reframing spatial sensing as a learning problem enables detail perception at a fraction of resources - from sound-localization in centimeter-scale devices to depth perception for insect-scale robots using metamaterial-based frontends. Then, we present a real-time battery-free localization system that achieves GPS-like accuracy while consuming thousands of times less power, capable of tracking millions of micro-assets across cities by leveraging non-linearity and next-G cellular infrastructure. The thesis concludes with a glimpse of ongoing and future projects exploring spatial intelligence in LLMs and digital twins for healthcare.