Top Neuromorphic Computing Stocks 2026: Thinking on Microwatts

Here's the dirty secret of the AI boom: most of the electricity a GPU burns goes to moving data back and forth between the processor and the memory, over a copper wire, billions of times a second.^[1]

Your brain doesn't do that. Memory and compute live in the same place (the synapse), neurons only fire when something changes, and the whole thing runs on about 20 watts. Roughly a dim lightbulb.

Neuromorphic computing is the attempt to copy that design in silicon. It consists of two big ideas:

1. Put the memory where the math happens ("compute-in-memory"), so data stops commuting.
2. Only compute when there's an event (a "spiking neural network"), so idle circuits draw close to nothing.^[2]

A GPU is a factory that runs every light and every machine 24/7, just in case. A neuromorphic chip is a night watchman who only wakes up when a window breaks.

On one hand, neuromorphic has been "five years away" for fifteen years. Intel's Loihi 2 and IBM's NorthPole are impressive research chips that you still can't buy at Digi-Key. The software is hard, training spiking networks is harder, and every vendor has its own toolchain.

On the other, 2026 is the year real products started shipping. BrainChip is delivering production AKD1500 silicon. Innatera's Pulsar microcontroller is designed into actual smoke detectors and wearables. GSI Technology's Gemini-II is running a 12-billion-parameter vision-language model at about 30 watts on a drone. And a two-month-old startup called Unconventional AI raised a $475 million seed round to build brain-inspired chips for the data center.

The edge (sensors, wearables, drones, always-on devices) is where neuromorphic wins first and cleanly, because the enemy there is the battery rather than NVIDIA. The data center is the bigger prize, but it's a 3-5 year bet that depends on chips that don't exist yet.^[3]

We slice this theme by neuromorphic exposure:

• Pure-plays live or die on this. Tiny, volatile, and where the leverage is.
• Integrated giants have neuromorphic labs, but the stock moves on everything else.
• Picks-and-shovels sell the memory and foundry capacity every neuromorphic chip needs, whoever wins.
• Private bellwethers you can't buy yet, but they tell you where the money and talent are going.

These are the companies where "neuromorphic" isn't a lab or a slide, it's the whole business. Both are small, both burn cash, both have spent years being told they're early. If the thesis works, these are the ones that move 5x. If it doesn't, they're the ones that go to zero.

BrainChip ASX: BRN

BrainChip is the company that has been shouting "neuromorphic is real" longer than anyone, and in 2026 it finally has silicon to back it up.

The Akida AKD1500 went into commercial production shipments on June 30, 2026, built on GlobalFoundries' 22nm FD-SOI process. It draws under 300 milliwatts in PCIe mode, sits next to a sensor as a co-processor, and only spends energy when an event arrives. In July it showed up in an M.2 form factor for fanless industrial boxes. In August, BrainChip released a free open-source bundle that lets IBM Spectrum Symphony (a workload scheduler enterprises already use) route small inference jobs to Akida instead of defaulting to a GPU. That's the first time BrainChip has pointed at the data center rather than the edge.

Before you buy, know this: the half-year to June 2026 showed revenue of US$1.22 million (up 19%) against a net loss of US$12 million. This is a licensing-model company with a microcap balance sheet, and it has funded itself through dilutive facilities for years. The LDA Capital facility wrapped up this period, so where the next raise comes from is an open question. The catalysts to watch are the AKD2500 development milestone (still tracking for late 2026) and an internal demo of its generative-AI platform (TENNs, a state-space-model architecture) by year-end.

GSI Technology NASDAQ: GSIT

GSI Technology is the sneaky one. On paper it's a 30-year-old SRAM chip company. Underneath, it's spent a decade building the Associative Processing Unit, a true compute-in-memory architecture where the math happens inside the memory array itself. Same principle as a synapse: nothing moves, so nothing burns.

The proof point that changed the story came from Cornell, whose researchers found the first-gen Gemini-I matched an NVIDIA A6000 on a retrieval task while using 98% less energy. GSI raised $50 million right after, and ended calendar 2025 with $70.7 million in cash. Since then, Gemini-II landed a proof-of-concept backed by the U.S. Department of War (with Israeli partner G2 Tech) for an autonomous perimeter-security system, posted a 3-second time-to-first-token running the Gemma-3 12B vision-language model at roughly 30 watts, and won Phase I of a smart-city deployment in Hsinchu County, Taiwan.

Unlike BrainChip, GSI has a real revenue base (quarterly sales around $6 million, growing double digits) paying the bills while the APU matures. Management is guiding to a revenue ramp in 2027 and taping out a next-gen chip called Plato, which is meant to cut power to 10 watts and shrink the die to a quarter of Gemini-II's size. The risk: the stock has a habit of selling off on good news, because every announcement so far has been a pilot, not a purchase order.

The giants have the deepest neuromorphic research on the planet. They also have market caps where a neuromorphic win would be a rounding error for years. You buy these because you like the company; neuromorphic is the free option attached.

Intel NASDAQ: INTC

Intel has been the academic reference platform for neuromorphic since Loihi (2018) and Loihi 2 (2021). Its Hala Point system at Sandia National Labs packs 1.15 billion neurons and remains the largest neuromorphic machine built. Hundreds of research groups run on Loihi 2 through Intel's research community.

What's new: Loihi 3. Intel's neuromorphic lab head Mike Davies teased it in mid-2025, and it's intended to be the first Intel neuromorphic chip that gets commercialized rather than loaned to universities. Be careful here. A wave of articles this year described Loihi 3 as "launched" with detailed specs; we could not find an official Intel product release confirming any of it. Until Intel announces pricing and availability, treat it as a research chip with commercial ambitions. If it ships, Intel becomes the first tier-one vendor with a neuromorphic SKU you can order, and that changes the whole category's credibility overnight.

IBM NYSE: IBM

IBM built TrueNorth in 2014 and then, in 2023, published NorthPole in Science. NorthPole is a different animal from the spiking crowd. Instead of mimicking neurons firing, it removes off-chip memory entirely and spreads memory and compute across 256 cores on one die. Result: about 22x better energy efficiency than a leading GPU on ResNet-50 inference, at the time of publication.

IBM has not turned NorthPole into a product you can buy. What makes IBM interesting in 2026 is the adjacency: the IBM Symphony scheduler is now the front door through which BrainChip is trying to get neuromorphic chips into enterprise data centers. IBM isn't going to move on neuromorphic news. But if you want a blue-chip holding with a legitimate claim to the "eliminate the memory wall" thesis, this is the one.

Every neuromorphic chip, no matter whose architecture wins, needs two things: a memory that can act like a synapse, and a fab willing to run an odd, specialty process instead of chasing 2nm. These companies sell exactly that.

Weebit Nano ASX: WBT

Weebit Nano licenses Resistive RAM (ReRAM), a non-volatile memory that stores a value as the resistance of a tiny filament. Change the resistance a little and you've changed a "weight." That's, more or less, what a synapse does, which is why Weebit's CEO keeps describing the ReRAM bit as behaving like one.

The commercial story is about flash replacement first, neuromorphic second. In the past 18 months Weebit signed Tier-1 licenses with onsemi and Texas Instruments, got qualified at DB HiTek, has silicon at 130nm, 65nm and 22nm, and as of July 2026 has three customer chips taped out. It's targeting at least A$10 million in FY26 revenue and a third foundry deal this year. (One partner you can no longer buy on its own: SkyWater, the U.S. trusted foundry with Weebit's ReRAM qualified on its 130nm process, was acquired by quantum company IonQ on July 31, 2026 and now trades only as part of NYSE: IONQ.) Then in spring 2026 it raised A$87 million in a placement (plus a share purchase plan), taking pro-forma cash toward A$172 million, and explicitly earmarked A$25 million for AI architectures: in-memory compute and neuromorphic. That's the tell.

GlobalFoundries NASDAQ: GFS

GlobalFoundries is the fab that makes BrainChip's AKD1500 on its 22FDX process, and the poster child for the "More-than-Moore" node strategy neuromorphic chips want: fully-depleted silicon-on-insulator, low leakage, cheap enough for an always-on sensor chip that sells for dollars, not thousands. If neuromorphic edge silicon scales to millions of units, a disproportionate share of those wafers run through GF-style specialty fabs, not TSMC's leading edge. Low-drama way to own the volume.

Everspin Technologies NASDAQ: MRAM

Everspin is the world's largest commercial MRAM (magnetic RAM) supplier, and MRAM is ReRAM's main rival as a synaptic memory. Everspin reportedly won a roughly $10.5 million development contract in 2025 to build an MRAM-based neural accelerator, with a production-ready demonstration targeted for 2026. The thing to watch is a design win with a defense prime that uses MRAM as the compute layer rather than just storage. Until then it's a profitable, boring memory company with a neuromorphic lottery ticket.

Publicly tradable neuromorphic options are thin, and the most important 2026 news came from companies you can't buy. These four tell you where the theme is heading.

Unconventional AI (Private)

Naveen Rao already built and sold two AI chip companies (Nervana to Intel, MosaicML to Databricks for $1.3 billion). In September 2025 he founded Unconventional AI, and by December it had raised $475 million at a $4.5 billion valuation, one of the largest seed rounds ever, led by a16z and Lightspeed with Sequoia, Lux, DCVC, Playground and Jeff Bezos participating. Rao put in $10 million of his own money and has said the round could stretch to $1 billion.

The bet is analog, not digital spiking: run neural networks directly on the physics of silicon (oscillators, dynamical systems) instead of simulating them with numbers, targeting roughly 1,000x better efficiency for inference. The company has released a first image-generation model, Un-0, running on a simulator of the architecture, and has said it's working with TSMC on what would be one of the largest analog chips ever built, with first silicon expected to tape out in 2026. Crucially, this is aimed at cloud providers and supercomputers, not wearables. No IPO chatter; expect more private rounds first. If Unconventional's first chip works, it re-rates every name on this list.

Innatera (Private)

Delft-based Innatera launched Pulsar at Computex in May 2025 and called it the world's first mass-market neuromorphic microcontroller. It pairs a spiking-neural-network engine with a RISC-V CPU and a small CNN accelerator, runs in the sub-milliwatt range, and its entire pitch is: sit next to the sensor, filter the noise in analog time, and only wake the main processor when something matters.

By CES 2026 it was showing Pulsar inside real hardware: 42 Technology's motor-vibration sensors, Aaroh Labs' smoke detectors with presence detection, and Joya's wearables. In March 2026 it signed with Synopsys to scale the next generation. Watch for a Series C or a strategic acquirer (a sensor-hub incumbent would be the obvious buyer).

Prophesee (Private)

Paris-based Prophesee makes event-based vision sensors: cameras with no shutter and no frame rate, where each pixel fires independently only when the light on it changes. Microsecond response, 120dB dynamic range, a fraction of the data. It co-developed the IMX636 sensor with Sony, launched the tiny GenX320, and has a multi-year supply deal with Qualcomm to get event sensors into smartphones.

Prophesee went through a court-supervised restructuring in 2024, re-emerged under new leadership, and has spent 2026 rebuilding commercially: a letter of intent with IDS Imaging at Embedded World, and a real-time event-processing software platform called Mantara in June. It's the "eyes" of neuromorphic. Roughly €126 million raised to date. An IPO is unlikely near-term; a strategic exit to a sensor giant is the more probable path.

SynSense (Private)

Founded out of ETH Zurich and now headquartered in Chengdu, SynSense sells the Speck (vision) and Xylo (audio and motion) chip families, and has a collaboration with BMW on neuromorphic driver-monitoring and gesture control for the intelligent cockpit. It's raised about $77 million total, including a $27.7 million round in mid-2025, and has been reported to be closing a Series B. Watch for a production-vehicle design win; that would be the first time a neuromorphic chip ships in a car at scale. Also worth knowing in the same neighborhood: Dresden's SpiNNcloud, which deployed a 650-million-neuron neuromorphic supercomputer at Leipzig University in July 2025.

Failure mode #1: the software never catches up. The hardware works. The problem is that spiking networks are harder to train than ordinary ones, every vendor has its own toolchain, and porting a mainstream model means rewriting it. Intel's own lab head said years ago that neuromorphic hardware was ahead of neuromorphic software, and that's still true. What to watch: whether standard frameworks (PyTorch-style tooling, ONNX export, scheduler integration like the IBM Symphony bundle) make neuromorphic chips a drop-in target. If a developer still has to "think in spikes" in 2028, the market stays niche.

Failure mode #2: "good enough" conventional silicon eats the edge. Qualcomm, Arm, and every microcontroller vendor keep shipping tiny NPUs that get more efficient every year. They're not 500x better, but they run existing models with existing tools. Neuromorphic has to win on power by a margin big enough to justify the pain of switching. What to watch: design wins in consumer products from brands you've heard of, not reference platforms and proofs-of-concept.

One more thing, specific to 2026. There's now a flood of AI-generated "neuromorphic goes mainstream" articles claiming products have launched when they haven't. Verify shipments against the company's own IR page. If a chip is "commercially available," there's a press release, a part number, and a distributor. If there isn't, it isn't.

The next 12-18 months are the first stretch where the theme gets graded on shipments instead of slide decks. Catalysts, roughly in order:

• Q4 2026: Intel Loihi 3 availability. An official product announcement with pricing, or another year of silence. Nothing on this list matters more.
• Late 2026: BrainChip AKD2500 and the GenAI demo. Also, watch how BrainChip funds itself now that the LDA facility is done.
• 2026: Unconventional AI's first tape-out and the remainder of its $1 billion round. First real data on whether analog brain-scale compute works outside a simulator.
• 2026: GSI's Plato tape-out and Hsinchu Phase II. Follow-on phases would be the first recurring software revenue from the APU. Management has told you to expect the ramp in 2027; hold them to it.
• 2026: Weebit's third foundry deal and first named AI customer. Both were promised for this year.
• CES 2027: "Powered by" moments. Innatera or Prophesee inside a product from a brand your parents recognize. That's the signal the edge market has arrived.

Our stance: this is an edge story first. Own the picks-and-shovels for the volume, keep a small position in one pure-play for the leverage, and hold the giants for a data-center future that's still being built in someone's fab.