Every time someone asks a question of a large language model, accepts a connection suggestion on a social platform, or follows a traffic rerouting from a navigation app, energy is being consumed at a scale most people never consider. The hardware running these interactions sits in vast data centers, drawing power around the clock. Understanding why that is, and why a seemingly minor laboratory finding might change it, requires going back to basics: how AI hardware actually works, and how far it still is from the organ it tries to imitate.
The Energy Problem Hidden Inside Every AI Interaction
The processors that power modern AI are GPUs, and each one can consume up to 1,000 watts. That figure puts a single GPU on par with a vacuum cleaner, a dishwasher, or a stove. The difference is that data center GPUs do not switch off. They run continuously, processing trillions of operations per second, and data centers contain thousands of them operating in parallel.
For comparison, a modern smartphone uses less than 1 watt.
This is not simply a matter of scale. It reflects a deeper architectural mismatch. GPUs are built to run software that simulates artificial neural networks, and that simulation requires moving enormous quantities of data through billions of transistors. The process is inherently wasteful because the hardware was never designed to think the way a brain thinks. It was designed to execute binary logic quickly, and neural network simulation is being layered on top of that.
The human brain, by contrast, is roughly one million times more energy efficient than current AI hardware at comparable tasks. That gap is not a minor engineering shortcoming. It is a signal that the entire computational approach may need rethinking.
What the Brain Does That Transistors Cannot
Biological neurons operate through a specific mechanism. When the voltage in a neuron’s cell body rises enough to cross a threshold, it triggers a self-propagating pulse called an action potential, which travels down a long projection called an axon. At the end of the axon sits a synapse, an electrochemical junction connecting to the next neuron’s branching projections, called dendrites. If enough of these pulses arrive within a given time window, the receiving neuron fires its own action potential. The system is event-driven, sparse, and extraordinarily efficient.
Artificial neural networks borrow the conceptual logic of this process but implement it through software running on transistors. The gap between the biological original and the electronic approximation is enormous, and it shows up directly in energy consumption.
Neuromorphic engineering is the field trying to close that gap. The goal is to build electronic components that behave like neurons and synapses at the hardware level, not simulate them in software. Progress has been made, but the dominant approach has a serious limitation: implementing a single artificial neuron or synapse requires interconnecting dozens, and sometimes even hundreds, of standard transistors. That demands significant chip area, introduces delays when signals must be processed sequentially, and makes scaling to systems that could compete with state-of-the-art GPUs extremely difficult.
Some research has explored entirely new experimental devices, but these have not yet reached the reliability needed for large systems. The field appeared to be stuck between two imperfect options.
A Single Transistor, Behaving Like a Brain Cell
Here is what most coverage of AI hardware misses: the solution may have been present all along, inside one of the most ordinary components in electronics.
Researchers discovered that a standard CMOS transistor, the basic building block of virtually all digital logic, can behave like a neuron or a synapse on its own, without being combined with dozens of other components. The discovery was accidental. It emerged from observing what happens when a normally overlooked part of the transistor, the bulk terminal, is not simply grounded but is instead connected through a resistance.
Under standard operation, increasing voltage at the drain produces a current that rises steadily and falls along the same path. But when resistance is added to the bulk terminal, something different happens. Electron-hole pairs that would normally be swept away begin to accumulate. Once the bulk voltage reaches a certain point, a hidden bipolar-junction transistor embedded within the MOSFET activates, causing current to spike suddenly. The current then remains elevated until the drain voltage drops past a separate threshold. The activation and relaxation points are different from each other, a property called hysteresis, which means current flows only for a defined period before the device returns to rest.
This behavior closely matches what an artificial neuron needs: a sudden, nonlinear spike when a voltage threshold is crossed, followed by automatic relaxation. A related configuration produces the adjustable, stable conductance that an artificial synapse requires. Each behavior, neuron and synapse, is achieved with a single device.
Why This Matters Beyond the Laboratory
The implications extend well past hardware engineering. AI’s energy footprint is already a significant concern, and it grows as models become larger and more widely deployed. A computing architecture that approaches the brain’s efficiency, rather than remaining six orders of magnitude away from it, would change the economics and environmental profile of AI infrastructure in ways that are difficult to overstate.
There is also a scaling question. Current neuromorphic approaches are constrained by the area and complexity required to build each neuron and synapse from many transistors. A single-device implementation removes that constraint and opens a path toward systems large enough to be genuinely competitive with GPU-based hardware.
This is what makes the accidental nature of the discovery worth noting. The component was not exotic or newly invented. It was an ordinary transistor, operating in a regime that had simply not been examined carefully. The insight was not in the material but in the observation.
In Short
A standard transistor, long overlooked in a specific operating mode, can behave like a biological neuron or synapse on its own, without requiring the complex multi-transistor circuits that have limited neuromorphic computing until now. Current AI hardware consumes up to 1,000 watts per GPU and is roughly one million times less energy efficient than the human brain at comparable tasks. This discovery does not solve that gap immediately, but it points toward a hardware architecture that could make AI far less energy-intensive, and far more scalable, than anything built so far.
Based on reporting from IEEE Spectrum AI.