Neuromorphic computing.
Twenty watts isn't enough.

Your brain runs on about 20 watts. A single rack of Nvidia's newest AI chips draws well over 100,000.

That gap has been the pitch for neuromorphic computing for almost 40 years: build chips that work more like the brain, and you get AI for a sliver of the power. The chips borrow two tricks. Their artificial neurons only "fire" when something changes, so they sit nearly idle most of the time instead of crunching every input. And they keep memory and computation in the same place, so they don't burn energy shuttling data back and forth, which is where most of a normal chip's power goes. A close cousin, in-memory and analog computing, chases the same savings without copying the brain as literally. The promise is AI that runs on microwatts in a sensor, or on a fraction of today's power bill in a data center.

Forty years in, that promise still hasn't turned into much of a business, or much of a market for neuromorphic computing stocks. BrainChip, the best-known listed pure play, booked US$1.2 million of revenue in the first half of 2026 while burning about US$11 million, and it has roughly a year of cash left. Meanwhile, ordinary AI chips keep getting more efficient, so the gap neuromorphic has to close shrinks a little with every generation.

Efficiency alone doesn't sell chips. A customer who can't get efficiency any other way does. These neuromorphic computing stocks are grouped by how close each company is to finding one: the listed edge pure plays, the memory companies underneath, the big-company lab, and the private companies chasing the data center and the battlefield.

These two are the listed bets on brain-inspired chips at the edge. Both still need more customers than they have.

BrainChip ASX: BRN

BrainChip sells Akida, a spiking neural network processor, as both chips and licensable designs. It's the purest listed neuromorphic company in the world, and it's still tiny.

The first production batch of its AKD1500 chip arrived this year, about 2,000 units, and yields came in lower than hoped. Its next chip, the AKD2500 on TSMC's 12-nanometer process, slipped to a December tape-out. First-half revenue was US$1.2 million against an operating cash burn of about US$11 million, which leaves roughly four quarters of funding. The share price is now below where it raised money last November.

This is a microcap racing its runway. The next raise is likely in 2027, and it'll come at a better price only if licensing deals and chip orders show up first.

GSI Technology NASDAQ: GSIT

GSI Technology makes compute-in-memory processors, chips that do the math inside the memory itself. It's not spiking neuromorphic, but it chases the same energy savings, and it has something BrainChip doesn't: an old memory-chip business with steady revenue and healthy margins that covers a lot of the bills.

Revenue is about $6 million a quarter, and defense customers made up nearly a third of shipments last quarter. It's running a U.S. Army research contract and a smart-city pilot, and its next chip, Plato, aimed at running language models under 10 watts, tapes out in March 2027. A strategic review ended in March with no sale, and it has $77 million in cash. The stock spiked twice in the past year and gave most of it back both times.

Keeping a neural network's weights right where the computation happens needs new kinds of memory. These two companies sell that memory today, mostly for ordinary uses, which makes them the steadiest neuromorphic computing stocks on the list.

Weebit Nano ASX: WBT

Weebit Nano licenses ReRAM, a memory that stores data by changing the resistance of a tiny cell. Texas Instruments and onsemi have licensed it, and it's one of the leading candidates for holding AI weights inside a chip.

Revenue more than tripled to A$15.3 million last fiscal year, and it has A$168 million in cash after raising about A$100 million this year. The company admits new licenses are taking longer than expected, and it hopes for its first royalties in 2027. The shares trade below its own raise price. Neuromorphic is an upside option here, not the business.

Everspin Technologies NASDAQ: MRAM

Everspin Technologies makes MRAM, magnetic memory that keeps its data with the power off and shrugs off radiation. That's why the military buys it: in April, Everspin signed a $40 million contract to make it for a U.S. Navy program.

Revenue hit a record $18.7 million last quarter. It also has a small engineering contract to develop MRAM for in-memory computing. The stock doubled this year, though it's well off its May peak. Everspin is a real memory business with a neuromorphic side door, and that's a much safer place to sit than a pure play.

IBM NYSE: IBM

IBM's NorthPole chip puts all its memory on the chip itself, borrowing the brain's trick of never moving data far. In IBM's tests it ran image recognition at 25 times the energy efficiency of comparable GPUs, and later ran a language model at more than 70 times the efficiency per token of a far more advanced GPU.

Last year IBM researchers described a full rack of 288 NorthPole cards running its Granite language models. It's still a research prototype, and it won't move IBM's stock. But it's the strongest evidence from any big company that brain-inspired design can compete in the data center, where the power bill is now the binding constraint.

Most of the real neuromorphic activity is private, and the money is splitting between data centers and defense.

Unconventional AI (Private)

Unconventional AI is Naveen Rao's new company. Rao sold two AI chip startups, Nervana to Intel and MosaicML to Databricks. This one is building analog chips that compute with the physics of coupled oscillators instead of digital logic.

It raised $475 million at a $4.5 billion valuation in December, led by Andreessen Horowitz and Lightspeed. It has no hardware yet, only simulations, and Rao has said a product is at least two years out. It's the market's biggest bet that brain-like efficiency matters most where the power bills are biggest.

Mythic (Private)

Mythic builds analog in-memory AI chips and nearly ran out of money a few years ago. It came back. In December it raised $125 million led by DCVC under a new CEO, and in January it bought Videantis, a German chip-design firm. Mythic is a survivor's bet: the technology was always interesting, and now it has the money to sell it.

Innatera (Private)

Innatera makes Pulsar, a spiking neural network microcontroller designed to sit in always-on devices like smart doorbells, wearables and sensors, running for months on a battery. Its first manufacturing customer, Joya, signed on last November. It's the closest thing to neuromorphic in an ordinary consumer product, if it can win more of them.

Prophesee (Private)

Prophesee makes event-based vision sensors, cameras that work like a retina and report only what changes in a scene instead of full frames. It went into court-supervised receivership in France in 2024, and a court approved its recovery plan this May.

Since then it has a new CEO, €20 million of fresh money and a new market: Mantara, a drone-detection system. A sensor that sees fast motion with almost no data or power turns out to be very useful for spotting small drones.

SpiNNcloud (Private)

SpiNNcloud, based in Dresden, builds supercomputers from SpiNNaker2 chips, a brain-inspired design from decades of European research. It has delivered systems to Sandia National Laboratories in the U.S. and to a research center in Leipzig. It's selling neuromorphic as infrastructure to government labs, which so far are the only buyers willing to pay for brain-scale machines.

Neuromorphic's first real customer is going to wear a uniform.

The early money already points there. Defense made up nearly a third of GSI's shipments. Everspin's biggest new contract is for a Navy program. Prophesee survived bankruptcy by pivoting to drone detection. SpiNNcloud's flagship U.S. customer is Sandia, a nuclear-weapons lab.

That's not an accident. A drone, a satellite or a soldier's sensor has a fixed battery, often no network connection, and a need to react in milliseconds. Those are exactly the conditions where brain-like chips beat ordinary ones, and where a buyer will pay more per chip to get them. A consumer gadget maker will always pick the cheaper, good-enough standard chip.

The commercial version of neuromorphic computing may well come later, in data centers starved for power. But for neuromorphic computing stocks, the business that funds it in the meantime will look a lot more like a defense contractor than a phone-chip supplier.