No one wakes up one morning and decides to call their eyes “the edge.”
The retina is no peripheral sensor bolted onto the brain, it grows out of the developing forebrain and it computes before anything reaches the cortex: contrast, motion, the boundaries of things. The eye is the brain reaching out to meet the light. No one thinks of the eye as periphery because it is so plainly part of the self.
We call devices “edge” because we assumed a center. The word carries a whole worldview inside it. The real thinking happens somewhere else, in a building full of servers, while everything outside that building is a sensor with opinions. Edge is a perspective, not a reality. A neuromorphic system that accepts the perspective as reality has already stopped being neuromorphic, because the brain has no center. The human brain has unity and diversity, hubs and highways, local rules and shared signals, yet no sovereign seat where it all comes together.
The data center is not a destination
In August, BrainChip released the Symphony Community Akida Bundle, an open-source integration that lets Akida neuromorphic processors run under IBM Spectrum Symphony beside CPUs and GPUs. Inference jobs that would have gone to a GPU by default can now work on Akida instead.
The easy reading is territorial. According to this read, neuromorphic compute has left the edge and moved into the data center. The territorial reading misses what actually happened. Nothing moved. What changed is that Akida became a peer. Akida is now addressable and accountable inside a shared system of work, one participant among many, receiving the jobs suited to its nature and contributing its results back into the whole.
The territorial reading also takes the data center’s name at face value. Most data is born far from any data center, in cameras, microphones, vehicles, phones and factory floors. Much of it is used right where it is born. For anyone whose work happens mainly at the so-called edge, the data center is no center at all, just one more place in a wider field. “Data center” is as much a perspective as “edge.” Each name makes sense only in relation to the other; neither describes the reality.
The significance lies not in the location, it lies in the relationship.
Beyond thesis and antithesis
The territorial reading is tempting because edge and data center have long been framed as an antithesis. The data center plays the thesis: centralized, powerful, authoritative. The edge plays its negation: distributed, constrained, peripheral, defined entirely by what the center is not.
The pairing follows the triad popularly attributed to Hegel: thesis, antithesis, synthesis. The formula owes more to Fichte and to Hegel’s nineteenth-century popularizers than to Hegel himself, yet it has become one of the most common ways of seeing the world. Two opposites collide; a resolution splits the difference. The technology industry has followed the script faithfully, with the data center as thesis, the edge as antithesis and “hybrid” as the synthesis meant to reconcile them. Yet black mixed with white yields only gray. Never color.
Hegel himself saw further than his popularizers. His word for how thought moves beyond an opposition is Aufhebung, usually rendered “sublation,” a single German word meaning at once to cancel, to preserve and to lift up. For Hegel, the terms of an opposition were abstractions, one-sided moments of a richer whole neither could grasp alone. “The true is the whole,” he wrote.
Read that way, edge and data center do not blend. They dissolve. They were never two regions of reality awaiting reconciliation. They were two abstractions drawn from a single field of computation, fixed in place by a perspective that assumed a center. Remove the assumed center and the antithesis has nothing left to hold it together. What remains is not gray. What remains is color.
Goethe, whose theory of colors Hegel defended against its Newtonian critics, offers the image. Goethe observed that a prism held before a uniform field of light shows no color at all. Color appears only at boundaries, where light meets darkness. Color is born at the edge. Goethe’s edge, though, is no place at the rim of things. His edge is a meeting. Meetings happen everywhere difference touches difference.
The black-and-white view of computing, center against periphery, cloud against edge, edge against data center, can only ever see the binary. A heterogeneous view sees the spectrum: neuromorphic, GPU, CPU, quantum, mainframe, each a distinct color, each vivid precisely because it remains itself. The world is not black and white. It is bright and beautiful, rich with colors no binary can hold.
The grammar of participation
Participation has a grammar. To participate, something has to be addressable, so others can reach it. It has to be schedulable, so it can take its turn in shared work. It has to be trustworthy, so others can rely on its contributions. It has to receive as well as give, learning from what flows back to it. A device that can only be commanded is an instrument. A system fluent in the grammar of participation is a participant.
The brain works this way and so do we. No human being thinks alone, even in private. The language you think in was built by millions of people you will never meet. Every concept you reach for carries their accumulated labor. Each generation begins from ground the last one gained. People plant trees whose shade they will never enjoy and write for readers not yet born. Human cognition isn’t the product of one brain ever. Instead, real intelligence is a vast living web across the whole plane of human existence, past, present, and future.
Participation needs an ontology
Grammar presupposes a vocabulary. Before anything can participate, the system must know what each participant is, needing a way to say how a neuromorphic processor differs in kind from a GPU, a CPU, a quantum processor or a mainframe, what kinds of work suit each one, how they relate to one another and who governs the work flowing among them. Such a vocabulary is an ontology. Heterogeneous compute needs its own.
Kubernetes cannot provide it. Kubernetes is an extraordinary system for placing containers onto nodes. Nothing here diminishes what it does well. K8s model of the world, though, is a flat pool of capacity. Even with device plugins and newer dynamic resource allocation APIs, an accelerator enters Kubernetes as an attribute of a node, something to be counted, matched and allocated. Kubernetes answers the question “where does this pod fit?” It has no native way to ask what kind of work a job is, what kind of resource suits it, or how that decision should be governed across competing consumers. Kubernetes places. A compute ontology understands.
“Hybrid compute” does not reach it either. Hybrid names a blend of two: on-premises and cloud, classical and quantum. Hybrid is an accurate word for what it describes, but it describes the wrong thing. The word tells you where workloads live or how two things are paired, not what the resources are or how they relate. A hybrid is also, by definition, neither of its parents. The distinctness of each is what gets lost. Hybrid is the popular dialectic’s synthesis, the gray that comes from mixing black with white. Heterogeneous computing is no blend. Its whole value depends on difference being preserved and put to work, each kind of compute doing the work it does best in relationship with every other kind. Hybrid describes a mixture. A heterogeneous compute ontology describes a community of kinds, a spectrum rather than a shade of gray.
The brain returns here as the model. Unity and diversity: many kinds of tissue, each doing what only it can do, coordinated without a coordinator. The retina is not a hybrid of eye and brain, it is a distinct kind of neural tissue taking its place in a whole. A heterogeneous compute ontology gives silicon the same possibility. Workload management built on such an ontology types each resource by what it is, governs the consumers who compete for it and routes each job to the kind of compute that suits it. When Akida joined Symphony, it joined an ontology, not a mere cluster.
Computing is a living web in the same way human life is. Workloads written decades ago still run beside models trained last month. Systems built by people who have long since moved on carry obligations into tomorrow. Neuromorphic compute is not merely analogous to this reality, it can share in it. Event-driven processing responds to things that actually happen, when they happen. A spike is not a simulation of an event. It is the system’s encounter with one. Timing is meaning and experience something neuromorphic compute lives.
The categories we inherited, edge versus cloud, accelerator versus host, on-premises versus hybrid, describe positions in a hierarchy. A participant is defined by its contribution and its belonging. The shift is paradigmatic rather than incremental for exactly this reason.
For the market realists
None of the argument so far is abstraction for its own sake. Markets price categories. The category you accept determines the market you are allowed to address.
“Edge AI chip” is a niche with a ceiling: wearables, sensors, per-unit cost pressure, long design cycles. “Accelerator in a hybrid environment” is hardly better, because it keeps neuromorphic compute subordinate to someone else’s stack. “Peer in a heterogeneous compute ontology” is a different thing entirely, opening the whole economy of workloads. Enterprises have been burning GPU cycles on small inference jobs, not because those jobs need GPUs, but because there was nowhere else to send them. Every job that doesn’t need the heaviest resource in the building is a cost that can be reclaimed: in power, in cooling, in capital locked up in hardware doing work beneath its capability.
Only an ontology captures that value. A flat container pool cannot tell workload that needs a GPU from one that merely defaults to one. Placement fills whatever fits. Intelligent routing sends workload where it belongs. When the addressable market is defined by kind of work instead of kind of place, geography stops bounding it. Work of that kind exists wherever compute exists. The opportunity is far larger than any market-labeled territory.
The vocabulary will not change overnight. Analysts, customers, and companies will keep speaking of edge and data center long after the underlying reality has moved beyond both because paradigm shifts take time to implement. Old categories linger in product lines, contracts, org charts, and roadmaps long after the thinking behind them has changed. Companies will sometimes need to speak the market’s current language simply to be understood. They cannot let that language set the limits of what they build.
Market forces reward whatever language already sells. Left alone, they will keep neuromorphic compute in the box it already occupies: edge chip, inference accelerator, niche. The fuller neuromorphic vision therefore has to be pressed constantly by the companies and individuals leading the charge, in what they build, in how they describe it, and in where they run it. Paying lip service to the vision while bowing to self-interested market forces will leave the old paradigm standing.
A garden, not a machine shop or factory
The environmental story may be the most important one.
According to the International Energy Agency, data centers used roughly 415 terawatt-hours of electricity in 2024, about 1.5 percent of the world’s total. The agency projects around 945 terawatt-hours by 2030, slightly more than all of Japan consumes today. The machine-shop answer to that curve is familiar: bigger machines, more power, more cooling, more of everything applied uniformly to every task regardless of what the task requires.
The industry’s most powerful company has made the metaphor explicit. NVIDIA calls its systems AI factories. In Jensen Huang’s framing, the data center is no longer a data center at all. It is a factory that manufactures intelligence, measured by the tokens it produces. The name is honest. But, “AI factory” also reveals a worldview.
A factory makes one kind of product at scale. The logic is throughput: raw material in, uniform output out, every machine judged by how much it produces per watt. Huang himself has acknowledged how much power AI factories squander. NVIDIA’s answer is a better factory: more tokens per watt from the same kind of machine. Better factories are real progress, yet they leave the deeper question unasked. A factory never needs to ask what kind of work a job is because a factory makes only one kind of thing. The mistake lies in taking the factory as the model for all of computing.
A flat resource model is machine-shop thinking in software: every job a part, every node a bench, the only question which bench is free.
A garden works differently. A gardener puts the right plant in the right place. Nothing is wasted because every part of the system feeds another. Diversity is no complication to be standardized away. Diversity is what makes the whole resilient and beautiful. The gardener works with time instead of against it, tending what is already growing and planting for generations still to come.
An ontology is how the gardener knows the plants. You cannot put the right plant in the right place without knowing what each plant is.
The brain runs on roughly twenty watts, not because it is small, but because it spends energy only when something happens and only where energy is needed. Event-driven neuromorphic compute carries that same disposition into our systems. Routed by an ontology to the work it suits, it takes what it needs and no more. I call that stewardship in silicon, not a sustainability feature bolted onto a product but a way of participating shaped from the start by care for the whole.
We are standing before an extraordinary garden of technological innovation. We can treat it as a machine shop, extracting output until neither the grid nor the climate can bear the load. Or, we can tend it and build systems that belong to the world where they compute.
Beyond training and inference
We have also narrowed our picture of intelligence to two verbs: training and inference. A model is trained once at great expense, then frozen and deployed to answer questions. The current state of the art has achieved remarkable things within that frame. The frame itself, though, is not brain-inspired. The human mind does not work in two phases.
No one is trained for twenty years and then switched to inference for the rest of their life. For us, learning and acting are one continuous process. Every perception adjusts the perceiver. Synapses change as we use them. Even remembering is a kind of learning. Neuroscience has found that recalling a memory can make it briefly malleable again so each act of recall restores it slightly changed. The brain does keep rhythms, sleep among them, when the hippocampus replays recent experience and the cortex slowly folds it into what we already know. Those rhythms are timescales within a single ongoing life of learning, not a factory’s separation of production from use.
Intelligence also includes far more than learning and answering. We attend, letting go of what doesn’t matter so that what matters can emerge. We imagine, dream, grow curious, decide what is worth learning, notice what surprises us, and learn from one another. None of those capacities fits neatly into training or inference.
A frozen model cannot discover the limits of its own knowledge, it can only answer from within them. A living mind learns most readily at the moment of surprise, where the unexpected marks the limit of understanding. The difference between a frozen model and a living mind is why I am fully engaged in “building systems that discover what they don’t know.”
The split carries a familiar hierarchy as well. Training happens at the center, in the factory, on the largest machines. Inference happens at the edge, where frozen knowledge is consumed. The training-inference split is the edge-data center antithesis again, drawn across time instead of space: knowledge made in one place and merely used in another.
Neuromorphic chips are too often perceived through that frame, as low-power inference engines for models trained somewhere else. Their event-driven plastic nature points toward something richer: continual learning, adaptation in the moment, knowledge that keeps living after deployment. On-chip learning today is an early step yet it points in the right direction.
A heterogeneous compute ontology changes the question. Instead of training at the center and inference at the edge, different kinds of compute can take on different rhythms of one continuous learning life: fast, local adaptation on neuromorphic silicon, slower and deeper consolidation on GPUs, shared knowledge moving among participants the way understanding moves among people. The brain itself divides learning this way, with fast learning in the hippocampus and slow integration in the cortex, neither one the center. Learning becomes something the whole system does, together, all the time.
Thinking together
Neuromorphic compute took its name from the brain. The deeper promise it signals reaches beyond any single brain. No mind has ever thought alone. Human intelligence is collective: many distinct minds, each local, each partial, coordinated without a coordinator across families, communities, disciplines and generations. We learn from what happens to us, when it happens. We spend attention where something matters. We receive what others have learned, change it by living with it, and hand it forward to those who come next. No one sits at the center of human thought. Human thought has no edge either.
Neuromorphic compute should represent that collective intelligence, the way we actually think, and the way we actually exist. Brain-inspired computing isn’t brain in a box imitating one skull’s worth of neurons. Rather, neuromorphic compute is a participant in the shared work of thinking, shaped like the life it serves. A spike that fires only when something happens is closer to lived experience than any clock cycle. A network that learns locally while listening for signals from the whole resembles a community more than a machine. A heterogeneous compute ontology, where each kind of compute contributes what only it can contribute, resembles human society at its best: unity and diversity, difference held in relationship, no one reduced to gray. A technicolor dreamland.
The move toward the data center was never about moving toward the center, abandoning the edge, or cramming more chips wherever they can be sold. The effort here is about neuromorphic compute taking its place as a distinct kind within a heterogeneous compute ontology, sharing in the living, unfinished work of human thought and prosperity augmented by machine intelligence that carries forward the thinking of people past, present and future, together.
Human flourishing by any other name. That is the new paradigm to embrace and why I say, “Run Akida anywhere. Run Akida everywhere.”