Laguna Hills, California. A card the size of a business card slides into an ordinary workstation and starts classifying images. It draws less than a third of a watt. No fan. No cloud round-trip. No GPU.
The card is BrainChip's AKD1500 PCIe development board, put on sale September 17. It is the third shipping format of the same chip in under three months — after production silicon in June and a gum-stick M.2 module in July.
For a decade, neuromorphic computing — silicon that fires only when its input changes, the way neurons do — has been the technology that is always five years from mattering. In 2026 the open question is no longer whether the idea works. It is whether anyone buys it.
From sample to shelf
TIMELINE: BrainChip's AKD1500, lab to shelf
─────────────────────────────────────────────────────────────
2022 ──────── 2025 ──────── 2026 ──────── 2026 ──────── 2026
✅ 🔬 🧪 🏭 ◉ NOW
AKD1000 AKD1500 AkidaTag Production PCIe card
released unveiled + RF/ASIC shipments on sale
at EW NA partners (Jun 30) (Sep 17)
AKD1500 product announcements, 2022–2026.
The AKD1500 runs useful inference on under 300mW, a fraction of a GPU's idle draw.
The test ahead is commercial, not technical: whether designers pick a new architecture over a familiar one.
The production turn
BrainChip announced commercial availability and initial production shipments of the AKD1500 on June 30. The chip is built by GlobalFoundries on its 22-nanometer fully depleted silicon-on-insulator (FD-SOI) process — a mature, low-leakage node chosen for efficiency rather than peak clock speed.
The numbers are the pitch. In PCIe mode the part draws under 300 milliwatts; in serial mode, under 200. The company says that delivers near-tera-operations-per-second (TOPS) efficiency in a device that can sit on a battery.
The formats followed fast. An M.2 2230 module — 22 by 30 millimeters, the footprint normally reserved for storage or a modem — shipped with a peak rating of 800 giga-operations per second (GOPS) at 4-bit integer precision, 1 megabyte of on-chip memory, and a typical draw of 250 milliwatts. It runs fanless from 0 to 70 degrees Celsius.
The PCIe card, announced September 17, puts the same co-processor into any desktop, workstation, or single-board computer with a free slot. It is a developer tool, priced for evaluation rather than deployment, but it closes the loop: a designer can now buy the chip, the module, and the board.
The commercial availability of the AKD1500 production chip marks a profound milestone in our commercialization roadmap.— Sean Hehir, CEO, BrainChip
As we wrote in September, the AI compute economy has flipped from training to serving. The serving layer is where power budgets are won or lost — and it is exactly the layer it is now selling into.
Why sub-watt matters
Event-based inference is a different bargain from GPU inference. A conventional accelerator evaluates the whole network on every frame, whether or not anything changed. Akida's fabric computes only on events — a pixel that shifted, a vibration that crossed a threshold — so idle costs almost nothing.
Where event-based silicon wins — and where it doesn't
Loses: large-model training, dense floating-point workloads, and anything that already has a wall socket and a cooling fan. A GPU still owns the data center.
That trade-off defines the market. For a data center, 250 milliwatts is noise. For a drone, a hearing aid, or a sensor on a factory floor, it is the difference between shipping and not shipping.
Power is a systems problem before it is a silicon problem. Cut the draw and you cut the battery, the heat sink, and the size, weight, power and cost (SWaP-C) envelope that governs defense procurement. That is why BrainChip's earliest production customers are in defense and wearables, not cloud infrastructure.
From radar to orbit
In June, BrainChip released a communication reference platform for RF signal classification, aimed squarely at defense contractors and government agencies working under tight SWaP-C limits. A radar that can tell a drone from a bird, on-device and at under 2 watts, is a different product from one that streams raw data to a server.
The partnership list reads like a map of where the company expects demand. ForwardEdge ASIC, a subsidiary of Lockheed Martin, is embedding Akida into future custom silicon for aerospace and defense signal processing. Frontgrade Gaisler is exploring the integration of the Akida core into fault-tolerant, radiation-hardened microprocessors for space. Neuromorphyx built an embedded developer board around the part, and IBM folded it into its Spectrum Symphony workload manager.
None of these are large revenue events yet. They are options — and they are being written across three of Nexithon's coverage areas at once, which is the more interesting signal.
Turning points
The company's history explains the shape of the bet. It spent years selling Akida as licensable intellectual property, a model that depends on someone else deciding to tape out a chip. The AKD1500 reverses that: it builds the part, stocks the shelf, and takes the inventory risk itself.
Selling silicon is capital-intensive and slow. A $25 million raise in December 2025 funds the ramp, but the neuromorphic market is still an option on a future, not a present-day cash flow. Revenue follows design wins, and design wins follow developer adoption.
The $25 million capital raise, announced ahead of CES in December 2025, was framed around exactly that shift — funding chips and modules rather than pure research.
The investment question
Grand View Research projects the neuromorphic computing market will reach $20.27 billion by 2030, growing at roughly 19.9% a year from 2024. Treat the number with the caution every forward market estimate deserves. It describes a category, not a company.
AKD1500 draw, PCIe mode
Serial mode is under 200mW; the M.2 module runs at 250mW typical. · BrainChip, 2026
The competitive field is real. Innatera raised a Series A for neuromorphic microcontrollers aimed at sensors, and NVIDIA's edge modules set the default that most designers reach for. Its advantage is that it ships a commercial part today; its disadvantage is that a new architecture has to be learned before it can be used.
The PCIe card is a quiet move against that disadvantage. Put the silicon in a slot, hand developers open tools with no license fee, and let them discover whether their workload fits. That is how a niche architecture becomes a default — or fails to.
It has done the hard engineering part. The commercial part is now a question of adoption, and adoption is decided one design win at a time.