The Brain Runs on 20 Watts ― Where Neuromorphic Chips Stand
Why is a brain-like chip still not in the machine on your desk? We go and look at what is missing
The human brain runs on about 20 watts — the power of a single light bulb. Seeing, hearing, walking, talking, thinking: all of it on 20 watts. The computers that run today’s AI, on the other hand, burn so much electricity that people now talk about building a power station next to the data centre. So why not build a chip that works the way the brain does? People have been thinking that for more than thirty years, and brain-like chips have in fact been built, generation after generation. But there is not one in your phone or your laptop. Today we go and look at why not.
Today’s route
- Starting point Why the “brain-like chip” raised such hopes: the brain’s 20 watts and AI’s electricity bill
- First corner How far it has come in thirty years, keeping research chips and chips you can buy apart
- The place we most want to see The chip itself works, so what is missing? Three layers, checked one by one
- A detour What has quietly changed in the last two years
- The high point Drop the assumption that “an AI chip is a machine for computing fast” and the view changes
- End of the walk A doorway where you build one neuron with your own hands
What this walk wants to say A brain-like chip cannot win by trying to replace the GPU. But on one point — using no electricity while nothing is happening — nothing else can replace it. Products built around exactly that point have begun to appear. That is why this is the hour before dawn.
GPU Originally a chip for drawing the picture on your screen. Because it can do thousands of identical calculations at once, almost all of today’s AI runs on it. Fast, but hungry for electricity.
NPU A chip built only for AI calculations, designed to use less electricity than a GPU. Almost every current phone and laptop has one.
Neuromorphic chip (neurochip) A chip that computes by imitating how the brain’s nerve cells work. The hero of this walk. What is inside it is explained in section 3.
1 Starting point ― a 20-watt brain and a country’s worth of electricity
Two numbers set the size of the hope.
The first is on the brain’s side. The human brain is said to contain about 86 billion nerve cells (neurons), all connected to one another, seeing things and understanding words. This enormous machine draws about 20 watts, less than a laptop charger. And it never stops: while you sleep it keeps you breathing and wakes you at a noise. “Always on, at 20 watts” is the destination brain-like chips have been aiming at.
The second is on the machine’s side. According to a report by the International Energy Agency (IEA), the world’s data centres used about 415 terawatt-hours of electricity in 2024, said to be around 1.5 percent of global consumption. That is expected to rise to about 945 TWh by 2030 — roughly the total annual electricity consumption of Japan. Data-centre electricity has reportedly grown at about 12 percent a year since 2017, more than four times the growth of electricity use overall. AI’s limit is becoming the amount of electricity, not the speed of the calculation.
Could copying the brain close that gap? The first person to think so was Carver Mead. Professor Mead taught at the California Institute of Technology in the United States, wrote a textbook on chip design in the 1980s, and is one of the people who laid the foundations of how chips are designed today. At the end of the 1980s he coined the word “neuromorphic”, meaning “shaped like the nervous system”. What he said was this. Today’s computers split every voltage strictly into “high” or “low” and treat those as 1 and 0; nothing in between is allowed, and electricity is spent forcing every signal to one side or the other. The brain is not that strict: its signals are continuous, and it works fine with a little noise and scatter. That is why it gets by on so much less power. And the transistor, left to itself, actually behaves rather like a nerve cell: its current changes continuously with voltage. So why not stop forcing transistors into 1 and 0, and use them the way the brain uses neurons?
“Not 1 or 0” may make some readers think of quantum computers. That is a different story.
Today’s computers split voltage into high or low and treat it as 1 and 0. A brain-like chip does not split it: it uses the continuous value as it is. If a computer is a tap that is only ever fully open or fully closed, a brain-like chip also uses “half open” and “just a trickle”. Both are ordinary electronic circuits running on ordinary physics.
A quantum computer is something else entirely. It uses quantum properties so that a single element can be in a superposition — “both 1 and 0” — and computes with that. It is good at a particular class of mathematically hard problems: factoring large numbers (which matters for cryptography), simulating molecules and materials, certain kinds of optimisation and search — “calculations that would take an ordinary computer tens of thousands of years, done in a realistic time”. The idea of simply running everyday AI on a quantum machine to make it faster remains, for now, a research hypothesis, not a product.
Do they compete? No. The brain-like chip’s seat is “next to a sensor, on a battery, waiting all the time”. The quantum computer’s seat is “a large fixed installation solving specific hard problems”. There is no situation in which they contend for the same seat. The one thing they share is that both try to get past the limits of today’s computers with an element that works on a different principle.
What about power? The quantum chip itself, in the superconducting designs, has almost zero electrical resistance and draws very little. But to run it you need a refrigerator that cools the element to near absolute zero (close to minus 273 degrees) and the control electronics that drive the qubits. For a whole current-generation system, figures of somewhere between ten-odd and several tens of kilowatts are quoted as a rough guide — about the scale of one rack of GPUs. Not a data-centre’s worth, but not a battery-powered story either. And, importantly, quantum computers do not yet do “useful work” at a practical scale, so there is no meaningful way to compare “watt-hours per answer”.
To sum up: today’s computers are “1 or 0”; brain-like chips are “the continuous value between 1 and 0”; quantum computers are “both 1 and 0”. In terms of power, brain-like chips are “milliwatts, on a battery, next to a sensor”; GPUs are “tens of kilowatts per rack, a country’s worth per data centre”; quantum is “ten-odd kilowatts or more per installation, fixed in place, and not yet doing comparable work”. Quantum is not a candidate for low-power AI; it is a different road from the one this walk follows. Both are introduced as separate species in chapter 2 of The Cambrian Explosion of Chips.
The idea turned into numbers in the 2010s. From a research programme begun by the US Defense Advanced Research Projects Agency (DARPA) in 2008 came IBM’s TrueNorth chip in 2014: circuits equivalent to a million nerve cells on one chip, drawing about 70 milliwatts according to the announcement. It was introduced as “a brain the size of a postage stamp”. Intel announced its Loihi chip in 2017 and a second generation in 2021. Market researchers have forecast that the market for these chips would grow by tens of percent a year over the coming decade, and those forecasts are still being updated.
2 First corner ― research chips kept growing; things you can buy have only just arrived
So what happened over those years? Line up the main chips and a clear pattern appears.
| Year | Chip | Scale / character | Where it stands |
|---|---|---|---|
| 2014 | IBM TrueNorth | 1M neurons, about 70 mW | Research. Never became a product |
| 2017 | Intel Loihi | About 130k neurons | Provided to researchers |
| 2021 | Intel Loihi 2 | Up to 1M neurons | Provided to researchers |
| 2024 | Intel Hala Point | 1,152 Loihi 2 chips, 1.15 billion neurons, up to 2.6 kW | Installed at a US national laboratory. Research |
| 2021– | BrainChip Akida | Under 1 µW on standby | Product. Volume shipments announced June 2026 |
| 2025 | Innatera Pulsar | Under 1 mW, for wearables | Product |
| 2020– | SynSense Speck and others | Camera and chip in one, under 1 mW | Product |
Figures as announced by each company. “Where it stands” is our reading of public information at the time of writing and may change.
The upper rows (IBM and Intel) keep setting records for scale. Hala Point, in 2024, packed more than a thousand times TrueNorth’s neurons into one system. But every one of them is “for researchers” or “installed at a national laboratory”; none is sold as a product. The lower rows (BrainChip, Innatera, SynSense) are products, but they are orders of magnitude smaller, and they aim at one narrow place: right next to a sensor, on almost no power, waiting all the time.
Ten years ago the hope was that the upper rows would become products and the lower rows would spread. What actually happened is that the upper rows stayed in research while only their scale grew, and the lower rows have only now begun to appear as products.
Two reasons, both noted at the time, explain why things did not go as hoped. First, the words “like the brain” ran ahead of the technology. TrueNorth and Loihi both run on a very simplified model of how a nerve cell behaves. When they were introduced as “learning like the brain”, that learning was a different, still immature method from the one today’s AI uses. Second, the performance comparisons were not fair. Figures such as “a thousand times more efficient than a GPU” were, it has been pointed out, usually measured on tasks the brain-like chip is good at, against a GPU that had not been optimised for them. On tasks today’s AI is good at — recognising what is in a photograph, say — the brain-like chips are said to have lost on accuracy, and the efficiency gap shrank.
None of this means the technology failed. It means the first lap finally showed what it is good at and what it is not.
3 The place we most want to see ― the chip is done; what is missing?
This is the place we most wanted to see on today’s walk. We use the same ruler as on the IoT walk and look at three layers. But first, what is inside a brain-like chip — because without that, you cannot see what is missing.
Today’s AI (a neural network) repeats one calculation — multiply lots of numbers, add them up, pass them on — everywhere, every time, to a fixed beat. To look at one photograph, or one frame of video, it runs that calculation from start to finish, all of it, every frame.
A nerve cell in the brain does not work like that. When it receives signals from around it, it accumulates them, little by little. What it has accumulated slowly leaks away over time. And only at the moment the accumulated amount crosses a certain level does it send out a single short signal. That short signal is called a spike. Once it has fired, the accumulated amount goes back to zero.
So inside a brain-like chip, information is carried not by “how big the number is” but by “whether a signal fired or not” and “when it fired”. A network built this way is called a spiking neural network (SNN). In this article we will just call it the spike method.
The good part: while nothing is happening, no nerve cell fires, which means almost no electricity is used. The bad part: today’s AI training methods and tools cannot be used on it as they are.
To sum up: the brain-like chip has cleared the technology layer. It is stuck at the two layers above it. On the IoT walk, IoT was stuck above its technology layer too, at “people who turn collected data into decisions” and “a design in which the money comes back”. The brain-like chip is stuck at “tools and people who translate today’s AI into the spike method”. In both cases, what is missing is translation.
4 A detour ― what has quietly changed in the last two years
We said “stuck”, but the view has moved in the last couple of years. Here are the tailwinds and the headwinds, side by side.
Tailwind ― products, at last
In June 2026 BrainChip announced volume shipments of a chip called the AKD1500, running continuous inference on under one watt, and said several customers were already receiving production parts. Innatera released Pulsar in 2025. What is interesting is what is inside it: not only a spike-method circuit, but a circuit for today’s AI and a small microcontroller, all on the same chip. It abandons the ideal of “do everything by the spike method” in favour of “use the spike method only for the part that waits, and hand over to ordinary AI once something happens”. SynSense has put a special camera, in which only the pixels that see movement send a signal, together with a spike-method circuit, into one component.
What these share is that they have stopped competing on scale and narrowed to one point: next to a sensor, on almost no electricity, waiting for something to happen. They have moved straight toward what the first lap showed they were good at.
Tailwind ― electricity has become a question of money
Once data-centre electricity reaches the scale of a country’s consumption, efficiency stops being a nice-to-have and becomes a matter of capital expenditure and utility bills. A chip IBM announced in 2023 called NorthPole does not use the spike method at all, but borrows one design idea from the brain — keep the place where you remember things next to the place where you compute — and has drawn attention in exactly this context. The brain-like way of thinking is seeping out beyond the spike-method chip as a single product, into how chips in general are designed.
Headwinds ― two
First, NPUs keep getting stronger. The efficiency of circuits that run today’s AI as it is improves every year, narrowing the room in which a brain-like chip can make a difference. Second, the big players have not committed. Intel continues to provide Loihi to researchers but has named no date for a product. IBM never turned TrueNorth’s successor into one. As long as the two companies leading on scale stay inside “research”, money will keep flowing away from the software and people layer.
5 The high point ― dropping the assumption that “an AI chip is a machine for computing fast”
This is the spot with the best view on today’s walk.
Much of the reason brain-like chips are not the main act comes from being compared on the GPU’s ground. Accuracy at recognising photographs, answers per second, calculations per watt — every one of these is a ruler the GPU defined. Measured by it, the brain-like chip lands in a losing position: slightly less accurate, fewer tools, efficiency that depends on conditions.
But the reason the brain gets by on 20 watts is not that it computes fast. It is that when nothing is happening, it does nothing.
GPUs and NPUs march to a drum called the clock. Every beat, everyone takes a step together. Whether or not there is anything to look at, the drum keeps beating and everyone keeps moving. The brain is different. A nerve cell fires once when what it has accumulated crosses its threshold, and the rest of the time it does almost nothing. It is the difference between an army on the march and a sentry. If nothing comes, the sentry just stands there, and calls out only when something does.
That difference matters in situations where, most of the time, nothing is happening. And most of the world’s sensors are placed in exactly such situations. A microphone listening to a factory machine’s vibration. A camera watching a car park at night. An accelerometer waiting for an elderly person to fall. A smart speaker waiting to be spoken to. For more than 99 percent of the day, these keep saying “nothing unusual”. Computing “nothing unusual” dozens of times a second, to the beat of a drum, is a waste of battery. Post one sentry, and wake the main force — ordinary AI — only when something happens.
Seen this way, the brain-like chip’s place changes. Not a replacement for the GPU, but the circuit that lets the GPU sleep. Not the thinking brain, but the spinal cord that pulls your hand back before you have thought. The part in charge of reflexes, in front of the “thinking AI”. In chapter 8 of The Cambrian Explosion of Chips we wrote that an AI with a body needs a loop that “senses its surroundings, decides in an instant, and drives its motors”. The “senses” and “in an instant” parts of that loop are sentry work, not drum work. Pulsar putting an ordinary-AI circuit next to its spike-method circuit can be read as exactly this “sentry and main force” arrangement, built inside one chip.
Drop the assumption that “an AI chip is a machine for computing fast”, and say instead that “an AI chip can also be a machine that wakes only when it should”, and for the first time you can see where the brain-like chip is truly good.
6 End of the walk ― building one neuron with your own hands
Everything above can be understood by reading. But whether “a circuit that does nothing when nothing is happening” really saves electricity, and what it means for “when a signal fired” to carry information — these two things will not get into your body until you build one. So this walk ends by building a single nerve cell with your own hands.
“Building” is nothing grand. The Basys 3 board used in the FPGA course has push-buttons and LEDs on it. We put one nerve cell between a button and an LED. Since this is a walk, we only sketch here what game you should be able to play; actually building the circuit and taking the measurements is the job of another corner (The Anything Lab, in preparation).
Press the button once. Nothing happens. Press it two or three times, slowly: still nothing. But hammer the button fast and, at some point, the LED flashes once and goes out. Keep hammering and it flashes again and again at a steady interval.
That is a nerve cell. Each press pours a little water into a “bucket” inside. The bucket has a small hole, and the water slowly leaks out. Press slowly and it leaks faster than it fills, so it never overflows. Press fast and the next pour arrives before the last has leaked, so eventually it overflows. At the moment it overflows the LED flashes once, and the bucket is empty again.
The Basys 3 also has a seven-segment display, so if you show the bucket’s water level as a number, you can watch it rise with each press, fall as you wait, and drop to zero with a flash when it crosses the line.
Use the board’s slide switches to change the size of the hole (how fast it leaks) and the overflow level (the threshold). Make the hole big and you have to hammer very fast before it flashes. Make it small and even leisurely pressing eventually flashes. Raise the threshold and it takes longer.
Here you notice something. What decides whether it flashes is not how many times you pressed but the interval between presses. Ten presses in one second flash; ten presses in ten seconds do not. This circuit is looking not at the size of a number but at a pattern in time. That is what section 3 meant by “information is carried by when the signal fired”.
Build a second nerve cell and use two buttons. Now connect the two cells so that the LED flashes only when both buttons are pressed at almost the same moment. Hammer the left alone: nothing. Hammer the right alone: nothing. Press left and right together: it flashes. This is a “coincidence detector”, the most basic neural circuit there is.
This is the most interesting part of the experiment. Your brain uses exactly this circuit to work out which direction a sound came from. When a sound comes from the right, it reaches your right ear a few hundred millionths of a second before your left. Inside the brain is a row of nerve cells that fire only when the signals from both ears arrive at exactly the same moment, and which cell fires tells you the direction. The two nerve cells you built on the board are the smallest unit of that mechanism.
To decide “did they arrive together?”, today’s AI turns both signals into numbers, multiplies, adds and compares. The nerve cell just overflows when water arrives in both at once. Which uses less electricity is obvious.
Finally, measure the board’s current. When you are not pressing the button — when the nerve cell is “doing nothing” — by how much has the current dropped?
The answer is “hardly at all”. The FPGA marches to its clock, and even when the nerve cell is doing nothing, the drum keeps beating. “No electricity when nothing is happening” is not achieved just by writing the nerve cell’s formula. Only when you add a circuit of your own that stops the clock while the button is not pressed does the current fall.
In other words, the brain-like chip’s low power comes not from “imitating a nerve cell” but from the design decision to stop the drum when nothing is happening. What section 5 called “drum and sentry” shows up here as a number. And you learn in your hands how much trouble stopping the drum is (how do you detect the first press after you have stopped it, for instance).
The last one is a sketch for turning this walk’s claim itself into a number. Build the same “did they arrive together?” circuit two ways. One is the ordinary way: read both buttons every beat, multiply and add the recent values, and compare (the drum method). The other is the two buckets above (the sentry method). Put the same number of each on the board — a hundred of each, say — and read the board’s current with a USB ammeter.
The prediction is this. The drum method works every beat whether or not you touch the buttons, so a hundred of them move the meter by tens of milliamps. The sentry method has nothing changing inside unless a button is pressed, so a hundred of them barely move the meter at all. Press a button ten times a second and it still barely moves. Drop the drum method’s clock to a tenth and the extra current should fall to near a tenth too, which shows at the same time that the electricity is proportional to “how many times something switched”.
Why a hundred? Because the electricity of one nerve cell is buried in what the board itself draws while idle (a hundred-odd milliamps). Working out how to make it visible is part of the lab’s job. And even with the sentry method, “the clock wiring itself” keeps beating the drum; to stop that too, you have to add a circuit that halts the clock while no button is pressed. That is where the “stop the drum” design from thing four gets checked in numbers.
Do these five things and what you could not get from reading stays in your hands. What it means to compute by “when it arrived”. Why two nerve cells can do the same job as the brain. And that the real source of low power is not the nerve cell’s formula but the design of the power and the clock. The brain-like chip’s promise and its trap are both there, in how one LED flashes and how an ammeter’s needle moves.
This experiment will not tell you whether brain-like chips make money. But someone who has measured the difference between the drum and the sentry with their own hands reads the next piece of news about this technology in an entirely different way. In section 3 we said there are not enough people who can translate. This is the shortest road to becoming one of them.
End of the walk Doorways you can try with your own hands
- To the lab Actually building and measuring things one to five on a Basys 3 is the job of The Anything Lab, Experiment 01 “Drum and Sentry” (in preparation). The walk sketches; the lab builds. That is the division of labour
- A road to finish first Foundations 02: FPGA, as far as blinking an LED (the first exercise). The lab’s experiment starts from there
- The road beyond Build the same nerve cell out of a capacitor (the bucket), a resistor (the hole) and a comparator (the overflow detector). Extending Foundations 04: Analog IC Design, you meet a problem digital never had: neighbouring nerve cells whose thresholds refuse to match. Beyond that lies submitting a small array of nerve cells to Tiny Tapeout
- The numbers under our feet Among the 3,625 designs in Chip Fabrication by the Numbers, how many are nerve-cell or spike-method designs, and how is that changing? A count is under consideration
Appendix: the limits of this walk
- What we did not look at. Quantum computers are covered only as far as explaining how they differ from brain-like chips. Brain-implanted devices (Neuralink, Synchron and others) and biochips that compute with cultured living nerve cells were out of scope; both will get walks of their own
- Conditions on the numbers. “N times a GPU” efficiency figures are measured on tasks and under conditions each company chose. We quote them, but independent like-for-like comparisons are scarce. Market-size forecasts vary so widely in definition and value between research firms that we quote no specific amounts. The power consumption of quantum computers varies greatly by method and generation; the figures here are a rough guide for whole systems of the current superconducting type
- The date. This is the view in September 2026. Shipment status, the big players’ product plans and the consolidation of development tools can all change within months. Before dawn, the scenery changes fast