Devices That Hum on Their Own Could Give Brain-Like Computers a Hardware Body

The brain runs on roughly the wattage of a dim light bulb. A large AI model can draw the output of a small power plant. That gap is the quiet motivation behind a whole field of research, and behind a review just published in Chemical Reviews by Gonzalo Rivera-Sierra, Roberto Fenollosa and Juan Bisquert at the Instituto de Tecnología Química, a joint center of Spain's CSIC and the Universitat Politècnica de València.

The mismatch is partly architectural. A conventional computer keeps memory and processing in separate places and moves data between them constantly. Every trip costs energy, and for the enormous models behind modern AI those trips dominate the power bill. The brain does not work this way. A neuron stores, processes and adapts in the same place, and computation emerges from its physical behavior rather than from shuttling numbers around.

"Neuromorphic" computing is the effort to build hardware that behaves more like that. The obstacle has always been the physical parts. You can approximate brain-like behavior in software running on ordinary chips, but you inherit the same energy-hungry architecture underneath. To get the efficiency, many researchers argue, the physics of the device itself has to do the work, and that is where the review focuses.

Its subject is self-oscillating devices: electronic components that generate their own electrical signals instead of merely responding to an external clock. According to the authors' institutional summary, these components can reproduce several signatures of neural behavior: they generate pulses, respond to stimuli, and synchronize with one another. Those are close cousins of what neurons do when they spike and fall into rhythm across a network, behaviors thought to be central to how brains compute.

The appeal is that a single device can carry several jobs at once. "The computation emerges directly from the device's physical and dynamic properties," Bisquert said in the summary. "Rather than separating memory and processing like traditional computers, oscillators combine storage, processing, and adaptation in one element." Collapse those functions into one physical element and you cut the constant data movement that drives up energy use, which is the whole point.

The review sketches where such hardware might eventually land: artificial intelligence, robotics, pattern recognition, signal processing, and systems that learn and adapt in real time. Those are early destinations, not shipping products.

This article sits at the materials-and-device-physics frontier rather than in software or model design, closer to chemistry than to the algorithms most people mean by "AI." Whether self-oscillating hardware becomes a genuine route to efficient neuromorphic computing, or remains one promising idea among several, is exactly the kind of question a review is meant to frame rather than settle.

Still, the framing is timely. As AI's appetite for power keeps climbing, the search for hardware that computes the way the brain does (cheaply, with physics doing the heavy lifting) is only getting more urgent. Devices that hum along on their own are one of the more intriguing places that search has led.