Neuromorphic Computing
The Chip That Thinks in Spikes, Not Streams
Every AI breakthrough of the last decade has been powered by hardware that is, at its core, doing something profoundly unnatural — and we're only now building chips that work the way brains actually do.
The Idea
Conventional processors — the ones running everything from your laptop to the data centres behind ChatGPT — operate on a clock. Billions of times per second, they ask: what's the next instruction? Data flows continuously, transistors switch on and off in lockstep, and enormous amounts of energy are spent just keeping the rhythm going, whether anything interesting is happening or not. Neuromorphic chips abandon this architecture entirely. Instead of continuous data streams, they process information through spikes — discrete, asynchronous electrical pulses, just like biological neurons. A neuron in your brain doesn't fire constantly; it fires when something meaningful crosses a threshold. Silence is free. This is the key insight: in a spiking neural network, the absence of a spike costs almost nothing. The result is a fundamentally different energy profile. Intel's research chip, Loihi 2, can perform certain inference tasks using a tiny fraction of the power of a conventional GPU — not because it's faster in clock-speed terms, but because it's quiet when it doesn't need to be loud. IBM's NorthPole chip, announced in late 2023, pushed this further: by embedding memory directly into the compute fabric, it dramatically cuts the energy wasted shuttling data back and forth — the so-called 'von Neumann bottleneck' that has haunted chip design for seventy years. Neuromorphic computing isn't trying to replace GPUs for training large models. It's targeting a different problem: running intelligence at the edge — in sensors, hearing aids, satellites, and autonomous systems — where power is scarce and response time matters more than raw throughput.
In the World
In 2017, a team at Stanford and the University of Zurich built a neuromorphic system called the Dynamic Vision Sensor — a camera that doesn't capture frames at fixed intervals but instead fires individual pixel-level spikes the moment light changes. Paired with a spiking neural network, it could track a fast-moving object with a reaction latency of around a millisecond, using a fraction of the power of a conventional vision system. The application that crystallised the stakes was drone obstacle avoidance. Traditional computer vision on a drone works by capturing frames, compressing them, sending them to a processor, running inference, and then acting — a pipeline with meaningful latency and a meaningful power bill. The neuromorphic approach collapsed that chain. The sensor and the processor were doing essentially the same thing the eye and the visual cortex do: responding only to change, and doing it almost instantaneously. Intel took this seriously enough to found a dedicated neuromorphic research community, the Intel Neuromorphic Research Community, and distribute Loihi chips to academic labs specifically to see what edge cases the architecture excels at. One group used it for olfaction — teaching the chip to recognise chemical signatures with startling accuracy using far fewer training examples than a conventional network would require. Another used it for robotic tactile sensing, where the irregularity and richness of touch data maps naturally onto spike-based processing. None of this is in your phone yet. But the trajectory is clear: as AI moves from the cloud to the device, the energy maths of conventional chips become increasingly untenable.
Why It Matters
The dominant story about AI hardware right now is about scale — more parameters, more GPUs, more power consumption. Data centres already account for a significant and growing share of global electricity use, and that figure is rising sharply as AI workloads expand. Neuromorphic computing represents a genuinely different bet: that the path to sustainable, ubiquitous intelligence isn't bigger clusters but smarter physics. This matters beyond energy statistics. If inference becomes cheap and local — running on a chip in a wearable, a medical device, a field sensor — the architecture of AI changes. Data doesn't need to travel to a server and back. Decisions happen at the point of perception. That shifts questions of privacy, latency, and access in ways that are hard to fully anticipate. It also reframes what we think 'thinking' looks like in silicon. The von Neumann architecture has been so dominant for so long that we've quietly assumed computing must look a certain way. Neuromorphic chips are a reminder that this was always a design choice, not a law of nature — and that biology solved the efficiency problem billions of years before we started asking it.
A Question to Ponder
If the hardware we use to run AI increasingly resembles the structure of biological brains, does that change what we should expect — or fear — from the intelligence it produces?
Get a new one of these every morning.
Start learning with Thinkable