Introduction
The modern artificial intelligence boom rests on an awkward secret: the smarter our machines become, the more catastrophically inefficient they are. A large language model learns by consuming oceans of data, and every gigabyte exacts a price in electricity. Technology companies now speak openly of building gigawatt data centers powered by dedicated nuclear reactors, while the water needed to cool them strains regional supplies. The human brain, by contrast, runs on roughly the power of a dim light bulb. By Hersam’s estimate it is five orders of magnitude more energy efficient than a digital computer, and it never stops learning.
This gap has long tantalized engineers. If the brain is the most efficient computer we know, why not build hardware that works the way it does? For half a century that ambition, known as neuromorphic engineering, has produced clever imitations of neural behavior but never anything that a living neuron would mistake for one of its own.
In April 2026, a team at Northwestern University crossed that line. Led by materials scientist Mark C. Hersam, the researchers printed artificial neurons that do not merely imitate the brain but communicate with it. When their devices were applied to slices of living mouse brain tissue, real neurons listened, and fired in response. Published in Nature Nanotechnology, the work marks a shift from mimicry to dialogue, and it carries implications for both the future of medicine and the energy crisis at the heart of AI.
The Von Neumann Wall: Why Silicon Hit a Ceiling
To understand the achievement, it helps to understand what silicon cannot do.
For sixty years, computing has advanced by a single relentless strategy: shrink the transistor, multiply it, and pack billions of identical copies onto a flat, rigid chip. This architecture is extraordinarily good at executing precise instructions in sequence, but it separates memory from processing, forcing data to shuttle back and forth across a bottleneck that the brain simply does not have. Every shuttle costs energy.
“Silicon achieves complexity by having billions of identical devices,” Hersam explains. “Everything is the same, rigid and fixed once it’s fabricated. The brain is the opposite. It’s heterogeneous, dynamic and three-dimensional.”
The contrast is fundamental. A processor is built once and then frozen, whereas the brain is assembled from many specialized kinds of neurons, arranged in soft three-dimensional webs that rewire themselves continuously as they learn. To move toward that, Hersam argues, the field needs not just new circuit designs but new materials and new ways of building electronics altogether.
The Long Search for an Artificial Neuron
The dream is older than the silicon crisis. In 1952, Alan Hodgkin and Andrew Huxley produced their Nobel-winning mathematical description of how a neuron fires, capturing the precise shape and timing of the electrical spike that carries thought. Their equations handed engineers a target: reproduce that spike in hardware.
In the late 1980s, Carver Mead founded the field of neuromorphic engineering on exactly this premise, designing circuits that emulated neural signaling. The arrival of practical memristors, devices whose resistance depends on their own history, gave the effort a powerful new building block. First theorized by Leon Chua in 1971 and physically realized by a team at Hewlett-Packard in 2008, the memristor behaves a little like a synapse, remembering the current that has passed through it.
Yet a stubborn problem remained. Most artificial neurons could produce only crude, simple signals, a single pulse where a real neuron commands a rich vocabulary. To perform anything sophisticated, engineers had to wire many devices together, and every added device burned more power, defeating the very efficiency the enterprise was meant to deliver. Worse, the timing was always wrong. As Hersam puts it, neurons built from organic materials spiked too slowly, while those built from metal oxides spiked too fast. Neither could meet a living cell on its own terms.
A Neuron You Can Print
Hersam’s team approached the problem from the material upward. Their artificial neurons are not carved from a crystal but printed from ink.
The inks are made of nanoscale flakes of two now-famous two-dimensional materials: molybdenum disulfide (MoS2), a semiconductor, and graphene, a superb electrical conductor. Using a technique called aerosol jet printing, the team deposited these inks onto flexible polymer surfaces, placing material only where it was needed. The result is cheap, low-waste, and bendable, the antithesis of a rigid silicon wafer.
The cleverest move, however, was to embrace a defect. The printable inks contain a stabilizing polymer that earlier researchers treated as a nuisance and washed away, because it interfered with electrical performance. Hersam’s team did the opposite. They left the polymer in and partially decomposed it; then, by driving current through the device, they pushed that decomposition further. Crucially, it happened unevenly.
“This decomposition occurs in a spatially inhomogeneous manner, leading to formation of a conductive filament, such that all the current is constricted into a narrow region in space,” Hersam says.
That narrow filament gives the device a sudden, threshold-triggered electrical response, a thermally activated “snap-back” that behaves uncannily like a neuron reaching its firing point. And because the behavior is so rich, a single printed neuron can generate not one signal but many: lone spikes, continuous tonic firing, and the staccato bursts that real neurons use to encode information. The researchers call this range multi-order complexity. Where silicon needs a crowd of identical components to approximate one neuron, a single printed device can do the work of many, which is precisely where the energy savings begin.
The Conversation: When a Printed Neuron Made a Living One Fire
A convincing imitation is one thing. Persuading a living neuron to respond is another. For that test, Hersam’s team turned to neurobiologist Indira M. Raman, whose laboratory studies the cerebellum, and applied the artificial signals to slices of mouse cerebellar tissue.
The target was the Purkinje neuron, one of the largest and most electrically expressive cells in the brain. The printed neurons’ spikes, it turned out, matched the biological ones in both timing and duration, and they reliably activated the living cells, triggering neural circuits in much the way natural activity does.
“You can see the living neurons respond to our artificial neuron,” Hersam says. “We’ve demonstrated signals that are not only the right timescale but also the right spike shape to interact directly with living neurons.”
This is the heart of the breakthrough. Earlier artificial neurons spoke in an accent the brain could not parse, too slow or too fast, the wrong shape entirely. Northwestern’s printed neurons landed inside a temporal window never before achieved, close enough to biology that the boundary between the synthetic and the living briefly dissolved. The machine spoke, and the brain answered.
Why It Matters: Brains, Machines, and the Energy of Thought
The consequences run in two directions at once.
The first is medical. A device that generates authentic neural signals on a soft, flexible substrate is a natural candidate for the interface between electronics and the nervous system. The researchers envision brain-machine interfaces and neuroprosthetics: implants that might one day help restore lost hearing, vision, or movement by speaking directly to the neurons that remain. Flexibility is not a cosmetic detail here, because living tissue is soft and rigid silicon is not.
The second is computational. By packing more behavior into each device and matching the brain’s own efficiency, printed neuromorphic hardware points toward AI systems that could perform demanding tasks at a tiny fraction of today’s energy cost. Hersam frames the stakes in stark terms.
“To meet the energy demands of AI, tech companies are building gigawatt data centers powered by dedicated nuclear power plants,” he warns, adding that it is hard to imagine a next-generation data center requiring a hundred nuclear power plants. Heat and water compound the problem, since the centers are cooled with water that AI is now straining. However you look at it, he argues, AI needs fundamentally more efficient hardware, and the brain is living proof that such efficiency is physically possible.
Conclusion
For most of its history, the project of building a brain-like machine has been an exercise in imitation, in coaxing silicon to behave a little more like a neuron. The Northwestern result reframes the goal. The achievement is not that a device mimicked a neuron more convincingly, but that it earned a reply from a living one.
That distinction matters. A technology that can both compute in the brain’s efficient idiom and converse with the brain’s own cells sits at the intersection of two of the century’s hardest problems: the unsustainable hunger of artificial intelligence, and the long-deferred dream of repairing the nervous system itself. A printed flake of molybdenum disulfide will not solve either overnight. But it has shown that the gap between the synthetic and the living is not unbridgeable, and that the bridge may turn out to be printable, flexible, and cheap.
The question is no longer whether a machine can imitate a neuron. It is what we will choose to say, now that the brain has started to listen.
References
-
Northwestern University. “Artificial neurons successfully communicate with living brain cells.” ScienceDaily, 18 April 2026 - https://www.sciencedaily.com/releases/2026/04/260417225020.htm
-
Hadke, S. S., Klingler, C. N., Brown, S. T., et al. (2026). “Printed MoS2 memristive nanosheet networks for spiking neurons with multi-order complexity.” Nature Nanotechnology. - https://doi.org/10.1038/s41565-026-02149-6
-
Northwestern Now. “Printed neurons communicate with living brain cells.” 2026 - https://news.northwestern.edu/stories/2026/4/printed-neurons-communicate-with-living-brain-cells
-
EurekAlert. “Printed neurons communicate with living brain cells.” 2026 - https://www.eurekalert.org/news-releases/1123883
-
Neuroscience News. “Printable Artificial Neurons That ‘Talk’ to Living Brain Cells.” 2026 - https://neurosciencenews.com/printed-artificial-neurons-brain-communication-30529/
-
Hodgkin, A. L., Huxley, A. F. (1952). “A quantitative description of membrane current and its application to conduction and excitation in nerve.” The Journal of Physiology, 117(4), 500-544. - https://doi.org/10.1113/jphysiol.1952.sp004764
-
Chua, L. O. (1971). “Memristor: The Missing Circuit Element.” IEEE Transactions on Circuit Theory, 18(5), 507-519. - https://doi.org/10.1109/TCT.1971.1083337
-
Mead, C. (1990). “Neuromorphic Electronic Systems.” Proceedings of the IEEE, 78(10), 1629-1636. - https://doi.org/10.1109/5.58356
-
Strukov, D. B., Snider, G. S., Stewart, D. R., Williams, R. S. (2008). “The missing memristor found.” Nature, 453, 80-83. - https://doi.org/10.1038/nature06932
