Our algorithmic media landscape is driven by the miniaturisation of smartness, a process centred on embedding computational power into increasingly compact and efficient devices. At the core of this transformation lies the Graphics Processing Unit (GPU), a hardware originally designed for video games and graphics acceleration but now integral to AI and computational parallelisation. This article examines the materiality of GPUs and their miniaturisation process by connecting two communities at the opposite ends of the devices’ lifecycle. On one end are liquid-nitrogen overclockers in Taipei, Taiwan, who push computational limits through extreme cooling practices; on the other are urban miners in Agbogbloshie, Ghana, who recycle and repurpose discarded GPUs. By uncovering the material, environmental, and knowledge infrastructure of GPUs, this study reveals smartness as rooted in geological, thermal and sociotechnological entanglements, in contrast to the dominant narratives of AI around immateriality, universality and neutrality. Through three analytical sections – Process and Circulation, Embodied knowledge, and Prototypes – this article uncovers how miniaturisation shapes and is shaped by labour, a desire for optimisation, and stratified economic relations. It ultimately argues that the drive for computational efficiency is inseparable from exploitative dynamics, reinforcing vertical hierarchies between tech corporations and local communities.