Virtually all current Artificial Intelligence (AI) applications are designed to run on sequential (von Neumann) computer architectures. As a result, current systems do not scale up. As knowledge is added to these systems, a point is reached where their performance quickly degrades. The performance of a von Neumann machine is limited by the bandwidth between memory and processor (the von Neumann bottleneck). The bottleneck is avoided by distributing the processing power across the memory of the computer. In this scheme the memory becomes the processor (a "smart memory"). This paper highlights the relationship between three representative AI application domains, namely knowledge representation, rule-based expert systems, and vision, and their parallel hardware realizations. Three machines, covering a wide range of fundamental properties of parallel processors, namely module granularity, concurrency control, and communication geometry, are reviewed: the Connection Machine (a fine-grained SIMD hypercube), DADO (a medium-grained MIMD/SIMD/MSIMD tree-machine), and the Butterfly (a coarse-grained MIMD Butterflyswitch machine).