AI has an energy problem. The data centres powering today’s large models are consuming electricity at a staggering rate, with tech companies building gigawatt facilities backed by dedicated nuclear power plants just to keep up.
Researchers are increasingly looking to the most energy-efficient information processor we know of- the brain, for a better way forward.
Why Silicon Keeps Hitting a Wall
Conventional computer chips achieve their power through brute uniformity — billions of identical transistors, all behaving the same way, fixed in rigid silicon once manufactured. The brain works in almost the opposite fashion. Biological neurons are wildly varied, producing a rich spectrum of firing patterns — isolated spikes, sustained bursts, rhythmic pulses — that allow them to encode and process information with extraordinary efficiency.
Most artificial neurons built for so-called neuromorphic chips, which attempt to mimic brain-like computing, have historically been pale imitations: simple, uniform, and limited. You end up needing millions of them just to accomplish modest tasks.
Printing Neurons From Ink
The Northwestern team’s approach is elegantly unconventional. They jet-printed an electronic ink — containing nanoscale flakes of molybdenum disulfide as a semiconductor and graphene as a conductor — onto a flexible polymer sheet.
The clever twist came from an accident of imperfection. There’s a stabilising polymer in the ink that researchers usually burn off completely after printing. This team only partially decomposed it, leaving behind irregular residue. When current was passed through the printed neurons, that residue broke down further in an uneven pattern, creating a tight, constricted channel through which current gets squeezed.
That constriction rapidly switches on and off, producing sharp voltage spikes — and not just simple on-off pulses, but the full range of complex firing patterns seen in real neurons. The result, from just two of these printed neurons and a handful of basic circuit components, was remarkably sophisticated signalling behaviour.
They Actually Talked to Real Brain Cells
To test whether their artificial neurons were genuinely convincing, the researchers hooked them up to slices of mouse cerebellum. The biological neurons fired in response. The synthetic signals were the right shape, the right timing, the right character to activate real neural circuits.
“You can see the living neurons respond to our artificial neuron,” said the research lead. It’s a remarkable demonstration — not just that the technology works on paper, but that it works in the context of actual biology.
This opens up potential applications beyond AI efficiency, including more sophisticated brain-computer interfaces and bioelectronic medicine, where devices need to speak the brain’s native electrical language.
The Bigger Picture
The more immediate ambition, though, is tackling AI’s runaway energy consumption. Today’s approach — packing ever more identical silicon components into ever larger data centres — is hitting physical and practical limits in terms of power and cooling. A computing architecture that mirrors the brain’s heterogeneous, adaptive, three-dimensional structure could, in theory, do far more with far less.
That future is still distant. The path from a lab demonstration to manufacturable chips to deployed AI hardware is long and littered with obstacles. But the underlying insight is compelling: the brain has already solved the energy efficiency problem that silicon is struggling with. Learning to build electronics that work more like it — flexible, varied, and adaptive — may be one of the most important engineering challenges of the coming decade.