12 June 2026
What if BrainChip’s edge AI story is only the beginning?
I first wrote about BrainChip in 2020. At the time, the company was still widely seen as an intriguing Australian deep-tech story: ambitious, unusual, difficult to explain, and operating in a field most investors and business readers barely understood.
Five years later, in 2025, I revisited the subject in an article titled A Chat with GPT: BrainChip Akida versus GPU / TPU technology. That piece explored the difference between BrainChip’s Akida neuromorphic processor and the better-known AI compute architectures powering the data-centre AI boom.
It posed the question; "could neuromorphic technology eventually usurp Von Neumann architecture and matrix math based AI processing?"
The work of Kevin D. Johnson - exploring Akida
But the recent public work of Kevin D. Johnson - an IBM Field CTO working across High Performance Computing (HPC), AI, IBM Spectrum Symphony, GPFS and related orchestration technologies - has provided the real inspiration for this article.
BrainChip’s Akida technology is positioned as a power-efficient edge AI chip, but Kevin Johnson has been exploring using it inside larger, orchestrated, enterprise-grade AI infrastructure.
At the time of writing, Johnson had completed an extraordinary 51 Akida demonstrations.
His work has included Akida with IBM Spectrum Symphony, GPFS/Spectrum Scale, LSF, vLLM, multi-chip coordination, model distillation, adaptive learning, RF sensing and industrial anomaly detection.
The striking point is the range. Johnson has applied Akida-based neuromorphic systems to (a few examples):
- live rail monitoring, where the system learns normal movement patterns and flags anomalies;
- RF disturbance sensing, where a wireless mesh becomes a form of distributed environmental perception;
- vLLM integration, where Akida is explored as a backend behind standard AI-serving interfaces;
- model distillation, where larger models are compressed into smaller Akida-compatible specialists;
- multi-chip coordination, where fleets of AKD1000 devices act as a shared neuromorphic resource;
- on-chip learning demonstrations based on Peter van der Made’s (Brainchip founder) early architecture;
- Jetson Nano integration, showing Akida being pushed into practical edge hardware;
- industrial anomaly detection across live visual feeds;
- distributed “Cerebra” experiments that treat Akida nodes as parts of a wider neuromorphic system;
- and heterogeneous compute demonstrations that place Akida alongside CPUs, GPUs, orchestration layers and shared storage.
These are early demonstrations, not commercial products. But they show why neuromorphic computing may matter beyond simple edge classification. Johnson’s work is not just asking whether Akida can run a model. It is asking whether Akida can become a schedulable, orchestrated, adaptive compute layer inside larger AI systems.
From One-Trick Pony to Akida Herd
A single Akida chip may be highly effective as a specialist “one-trick pony” - a compact, power-efficient device doing one narrow job extremely well.
But Kevin D. Johnson’s work points to something more interesting: what happens when many Akida devices are coordinated into a herd?
A single Akida device is already interesting where the task is highly specific: a sensor-side classifier, an always-on detector, a local inference engine, a low-power audio or vision processor, or a specialist model running close to the device. In that role, Akida can be thought of as a very efficient “one trick pony” - not in a negative sense, but in the sense that it can perform a specialized task with extremely low power draw and minimal dependence on cloud infrastructure.
But Johnson’s work points to something larger.
Across his demonstrations, Akida increasingly appears as part of a coordinated array: multiple chips, multiple nodes, shared state, distributed sensing, orchestration and task routing.
The power of neuromorphic computing may not come only from one chip doing one thing well. It may come from many specialized neuromorphic units working together.
This is where the comparison with the brain becomes more useful. A brain is not one giant processor performing one enormous calculation. It is a distributed system made from many specialized structures, pathways and feedback loops. Different regions handle different tasks, but the intelligence emerges from how those regions coordinate, adapt, remember and respond.
Johnson’s use of IBM Spectrum Symphony, LSF and GPFS/Spectrum Scale appears to explore a machine equivalent of that idea. Akida chips provide the power-efficient neuromorphic compute. Symphony and LSF provide orchestration and scheduling. GPFS provides shared state, model access and a memory-like fabric. The result is not simply “one Akida running one model,” but a possible architecture for many Akida devices acting as a coordinated compute layer.
That is why the array concept is so important. A single Akida may be ideal for a narrow edge task. But arrays of Akida chips raise a much bigger possibility:
Powerful distributed AI systems that still consume extraordinarily low amounts of electrical power compared with brute-force conventional AI.
If that approach scales, Akida’s role changes. It is no longer just a low-power edge processor. It becomes a building block for larger adaptive systems: rail-monitoring networks, RF-sensing meshes, industrial anomaly detection, mobile autonomy, defence systems, robotics, smart infrastructure and eventually more brain-like machine architectures.
Launching at the edge
BrainChip’s current commercial pathway is focused on Edge AI.
That makes sense. Edge AI is where Akida’s immediate strengths are easiest to understand: power-efficient, local inference, fast response, reduced data movement and intelligence close to sensors. Compared with the giant GPU clusters dominating the public AI conversation, this is a more practical and commercially grounded starting point.
Edge AI may only be the market entry point.
The title of this article, "The thin edge of the neuromorphic wedge", reflects that realization. BrainChip’s founder, Peter van der Made appears to have had a deeper ambition: to build computing systems inspired by the way biological brains process information, adapt, learn and make sense of the world. The following observation drives that ambition...
The mammalian brain remains the most impressive computing system we know. It processes vast amounts of sensory information, learns continuously, adapts to novelty, controls movement, recognizes patterns, forms memories and supports intelligence - all while consuming roughly 20 watts of power.
Modern AI, by contrast, is powerful but brutally inefficient.
Data centres with thousands of rows of equipment racks hum with the deafening sound of cooling fans and even though they use refrigerated air-conditioning - the amount of heat being pumped out of the data centre requires water cooling towers. Data centres are now being built adjacent power stations and need access to huge amounts of water.
Neuromorphic computing asks a different question: instead of forcing intelligence through ever-larger conventional compute systems, what if we built machines that process information more like the power efficient human brain?
Kevin Johnson’s recent work (which I am documenting here) is important because it appears to be exploring exactly that wider role not just as theory but through actual practical demonstrations.
Akida is being applied as a useful specialist accelerator inside a larger compute environment toward something more ambitious: Akida as a practical building block in systems that learn, adapt, notice novelty, and coordinate many small units into something larger than the sum of its parts.
From edge AI chip to neuromorphic architecture
The diagram below shows the larger possibility. It is not a forecast, and it is not a claim that each stage is inevitable. It is a way of visualizing how Akida’s role could expand: from a thin-edge sensor processor, to an always-on sentinel, to a front-end intelligence layer, to coordinated multi-chip arrays, to a broader neuromorphic compute fabric, and ultimately toward brain-like machine intelligence.
Kevin D. Johnson’s work is already pushing Akida several steps along this progression.
BrainChip’s commercial entry point is the thin-edge sensor processor. But Johnson is exploring using Akida progressively through more demanding roles from an always-on sentinel, a front-end intelligence layer, a coordinated multi-chip array, to a schedulable neuromorphic compute resource inside IBM Spectrum Symphony, LSF and GPFS.
That does not mean the end point has been reached.
But it does suggest the path from edge AI chip to broader neuromorphic compute fabric is no longer purely theoretical.
This is also where Johnson’s use of IBM Spectrum Symphony becomes more than ordinary enterprise workload management.
In a conventional computing environment, Symphony schedules jobs across available resources. But in Johnson’s Akida work, it begins to look like a coordination layer for a larger neuromorphic system. If individual Akida chips and simulations act like specialized neural regions, Symphony helps decide which region is activated, which task is routed where, and how the wider system behaves as a coordinated whole.
The analogy is not exact, but it is useful. In the mammalian brain, structures such as the hippocampus help with memory formation and spatial/contextual mapping, the thalamus helps route sensory information, the basal ganglia help select actions, and limbic structures help assign importance or attention to novelty and threat.
Johnson’s architecture hints at a machine equivalent of that division of labour.
Peter van der Made’s original neuromorphic ambition was never merely one chip doing one classification task. The larger idea was brain-inspired computation built from many interacting neural structures. Johnson’s “Cerebra” work appears to pick up that thread in practical infrastructure: Akida nodes, RF sensing, distributed learning, shared memory and orchestration working together as early building blocks of a larger adaptive machine-intelligence architecture.
The commercial journey: still a long way to go
Let's not get ahead of ourselves - the commercial reality is still early. BrainChip must still prove repeatable revenue, customer adoption and scalable commercial traction. That's the problem with technology based on completely new thinking - it's too easy to use familiar methods. It takes something compelling to accelerate adoption (maybe like running out of energy for AI data centres).
The near-term BrainChip story is about Edge AI. The longer-term story may be about whether Akida can become one of the first practical building blocks of a much broader neuromorphic computing future.
Johnson’s work should not be mistaken for an IBM product announcement or proof that neuromorphic computing has already crossed into mainstream adoption.
But as a technical and strategic signal, it is hard to ignore.
The bigger question
If artificial intelligence ultimately needs something closer to an artificial nervous system, a system that can learn locally, adapt to novelty, coordinate many specialized sensing and inference units, and operate within tight space and power constraints, then Akida may be more than an edge AI chip.
It may be one of the first practical steps toward a broader neuromorphic computing architecture.
That does not mean BrainChip is close to building an artificial human brain, but it does mean Peter van der Made’s original ambition deserves to be taken seriously: intelligence built from brain-inspired compute, not merely brute-force scaling by throwing more kilowatts at the problem.
That is the question this article explores.
