Neuromorphic computing is an approach to building computer hardware that imitates the physical style of the brain rather than just the math of neural networks. The term, coined by Caltech’s Carver Mead in the 1980s, covers chips whose design choices - how they store data, when they compute, how they communicate - are borrowed from biology.
Three ideas recur. First, computation is event-driven: like real neurons, the silicon units do work only when they “spike,” so an idle network burns almost no energy. Second, memory sits next to processing instead of in a separate bank, avoiding the constant data shuttling - the von Neumann bottleneck - that dominates the energy budget of ordinary computers. Third, the system is massively parallel and asynchronous, with no single global clock driving every step. IBM’s TrueNorth, Intel’s Loihi, and Manchester’s SpiNNaker are the best-known examples.
This is a different bet from the mainstream of modern AI. The deep-learning boom runs on GPUs, which are fast, dense, clocked arithmetic engines - very unlike a brain. Neuromorphic chips trade that raw throughput for extreme energy efficiency on sparse, spiking workloads, and for the ability to learn continuously on the device.
Why business readers should care: the running cost of AI is increasingly an energy story, and large models already strain power grids and data-center budgets. Neuromorphic computing is one of the long-shot bets on a fundamentally more efficient substrate - promising for always-on sensing at the edge, though still mostly a research field rather than a mass-market product.