Another article for your consideration. Where does the knowledge go in America's reindustrialization? Every machine learns where it stands. Only the answer moves. https://x.com/i/article/2108011038398382081
Only the Answer Moves
Neuromorphic compute, the heterogeneous compute ontology, and industry rebuilds
Rye on the Monongahela
One of the first great American arguments about value leaving the place that made it happened in the hills of southwestern Pennsylvania.
In the 1790s, farmers in western Pennsylvania grew rye they could not sell. A wagon of grain could not pay its own way over the Alleghenies to the eastern markets, so they did the sensible thing and distilled it. A barrel of whiskey was worth carrying. Monongahela rye became the region's product, its savings account, and in many households its currency. Then a federal excise arrived, payable in cash the frontier rarely had and structured in a way that weighed lighter on large eastern distillers than on the small stills scattered up every hollow along the river. The country west of the mountains understood the arrangement perfectly well. Whatever they made, the value would be counted somewhere else.
We remember how that ended, Washington riding west at the head of an army. What we tend to forget is that the grievance underneath it never really went away. It just changed commodities. Coal left these valleys for two centuries. Steel left, and then the mills left. Gas leaves now, through pipelines and through a plastics plant in Beaver County that was supposed to change everything and mostly changed the skyline. The pattern is so familiar here that people describe it as weather rather than as a design decision. But it is a design decision. It always has been. Somebody decides where the value is counted.
America is now trying to rebuild its industrial base at a speed the country has not attempted since the Second World War and so are its allies. The question I want to put on the table is the same one the rye farmers were asking, applied to the newest commodity the new factories will produce in enormous quantities. That commodity is knowledge: what every machine, every hull, every crane, every sensor learns while it works. Where does it go? Who does it compound for? And can we build the nervous system of a reindustrialized nation in a way that, for once, lets the value stay where it was made?
I think we can. The hardware to do it is already shipping, the orchestration to do it descends from a lineage that has been in production since 1992, and I have spent the last ten months building the evidence.
The train has left the station
Anyone who doubts that reindustrialization is real has not been paying attention to Ohio. Anduril announced Arsenal-1 outside Columbus in January 2025, and the first Ohio-built Fury rolled off the line in July 2026, on a line that can produce up to 150 aircraft a year in its current configuration. The plant is designed to eventually produce tens of thousands of weapons each year and the company is working with Pickaway County high schools and Ohio universities to build the workforce it will need.
The shipyards are moving too. Hanwha has said its $5 billion expansion of the Philly Shipyard is aimed at taking annual output from fewer than two vessels to as many as twenty. Saronic raised $1.75 billion to scale autonomous shipbuilding and is expanding in Louisiana while planning Port Alpha in Texas. These are not press releases about intentions. They are steel, slips, and payroll.
And on the silicon side, the neuromorphic part of the story is crossing from samples to supply. BrainChip received its first production batch of 2,000 AKD1500 processors in Q2 2026 with the AKD2500 on track for tapeout in early December. Parsons is integrating Akida into mission-ready edge platforms. The timing is not a coincidence I would bet against.
Notice what every one of these programs produces at scale: autonomous machines and the factories that build them. A Fury, a Marauder, a crane on a dry dock, a press on a line at Arsenal-1, all of them sense, all of them decide, and all of them generate a continuous record of how the work actually goes. Rebuilding the arsenal of free nations means building millions of things that learn. The architecture that decides where that learning lives is being chosen right now, mostly by default, and the default is the one this region knows too well.
Two ways to build a nervous system
My doctoral work compares coordination infrastructures by asking what they assume about knowledge, not what they say about politics. The two systems I set side by side could hardly sound more opposed. Palantir speaks the language of national strength and operational advantage. The UN Sustainable Development Goals speak the language of global solidarity and shared progress. One is technocratic universalism, the other institutional universalism, and their admirers rarely attend the same conferences.
Yet underneath, they are built the same way. Both depend on centripetal data flows, where knowledge moves from the places that generate it toward a center that interprets it. Both impose schemas from outside, so the locality must describe itself in someone else's vocabulary before it counts. Both keep institutional memory centrally, so the compounding value of what was learned accrues where the data landed rather than where it was lived. And, both carry single points of failure, because a center that holds everything is a center that can lose everything, or withhold it.
That convergence is the finding that matters. It means the decisive question about any coordination system is not whether it leans left or right, national or global, commercial or humanitarian. The decisive question is ontological: does the infrastructure extract knowledge from localities to centers, or does it let knowledge remain with the people and places that generate it?
A rich and free society should be uniquely bad at tolerating the first answer. Freedom that ends at the network boundary of a platform is tenancy as the frontier labs have shown us. Wealth whose compounding engine sits in somebody else's object model is rent. I made a version of this argument about sovereignty in Down in the River: an organization is sovereign only to the depth of the lowest layer it owns and can describe. The same is true of a shipyard, a county, and a nation.
So there are two ways to build the nervous system of a reindustrialized economy. The first is the familiar one. Instrument everything, ship the telemetry to a central platform or a hyperscale data center, interpret it there, and send instructions back. It works, in the narrow sense, and it reproduces the rye farmer's problem with better graphics: the place that does the work generates the knowledge and the place that holds the platform keeps it. The second way inverts the flow. Each machine learns its own normal where it stands. Each site holds its own memory. What crosses the network is the answer, not the record, and the centers that exist coordinate rather than accumulate. That second way used to be a philosophical position. It is now an engineering option and the rest of this piece is about why.
Monaca and Homer City
Western Pennsylvania has already run both of the recent experiments in what happens when the anchor is large and the value is counted elsewhere.
The first is the Shell plant in Monaca. Pennsylvania offered a $1.65 billion tax break to land it and the federal government once described it as the first of what could be several crackers across the Ohio Valley. It ended up costing $14 billion to build, it employs roughly 600 people, and it remains the only one that was built. In 2025, Shell's chief executive said the company wanted someone else to own the plant and this summer bidders including ExxonMobil, LyondellBasell, and Apollo submitted offers for Shell's U.S. chemicals portfolio, Monaca included. An Ohio River Valley Institute analysis found that since the project was announced in 2012, Beaver County's inflation-adjusted GDP has contracted 12 percent and its population has fallen 3 percent. People will argue about those numbers and they should. What nobody argues is that the cluster of local processors the plant was supposed to seed never really arrived. The ethane came out of the ground here, the polyethylene shipped out, and the know-how stayed inside one company's global portfolio.
The second experiment is underway in Indiana County, where the old Homer City coal plant is becoming a 4.5 gigawatt gas-fired campus for hyperscale data centers, a $10 billion bet on the largest gas plant in the country. It will bring real construction work and real permanent jobs, and no one can complain about that. But look at the shape of the flow. The gas comes from the region. Most of the power stays behind the fence. The computation inside serves whatever the tenants are training, for customers who are mostly not here, and the knowledge produced compounds in someone else's models. Meanwhile, Pennsylvania electricity rates rose nearly 14 percent between April 2025 and April 2026, almost double the national average, and the manufacturers we say we want back pay those rates.
When I ran a cognitive-warfare analysis of the data-center backlash earlier this year, the first finding was that the grievance is genuine: roughly 70 percent of Americans oppose a data center nearby. Organized campaigns have amplified it and foreign money has found it useful, but they did not invent it. People can feel the shape of an extractive flow even when nobody draws it for them.
Neither project is a villain. Both are the predictable result of reasoning about infrastructure from the top down and stopping at the first layer somebody can sell. The cracker treated the region as a feedstock. The data center treats it as a power supply. What neither treats it as is a place that learns.
Only the answer moves
The engineering reason the second way is now possible is that a kind of silicon exists that learns where it sits.
BrainChip's Akida is a neuromorphic processor built on Peter van der Made's insight that the brain's cortical column is a small repeatable unit of computation. It computes only when events arrive, draws power measured in milliwatts, and learns on the chip while it runs. That last property is the one that changes the politics of data. A processor that can learn in place does not need to send the world somewhere else to be understood.
In August, I put ten of them inside ten storage nodes of an IBM Storage Scale file system and let each one learn its own node's normal behavior on the chip itself, watching 35 of its own signals through a small ReRAM array. The learned model of normal for each node was 4,157 bytes. The whole fleet came to 58 kilobytes. When I drove heavy I/O one node at a time, that node departed from its learned normal in 29 of 30 trials while its neighbors held still, an 11.4 sigma separation. The raw telemetry was discarded on the node that produced it. Only the answer crossed the network. I titled that demonstration Only the Answer Moves and I have come to think the phrase is the whole design philosophy of a non-extractive industrial base in four words.
The system generalizes. Ten chips watching live rail cameras classified railcars at 91 percent accuracy with a model that fits in a megabyte of on-chip memory, producing freight intelligence comparable to satellite services that cost tens of thousands of dollars a year, all from public cameras. Redundant chip pairs across a ten-node cluster traced flaring and venting events to their subsurface cause by fusing well telemetry, satellite methane, thermal, and seismicity on-chip, with automatic failover when a chip went dark. A single AKD1000 beside a ReRAM lens caught a seamless GPS time spoof as an independent witness the attacker could not reach. None of these needed a data center. All of them kept the record where it was generated.
But a chip is a component and components alone do not make a nervous system. This is where the heterogeneous compute ontology does its work. I have argued at length in Old Black Water that this layer should not be called an operating system because Linux already owns the machine and a coordination layer that pretends otherwise is building a moat, not a system. Instead, the layer is actually a fabric. IBM Spectrum Symphony places services on the right kind of silicon, keeps models resident, fails over automatically, and lets many tenants share one pool on their own terms. IBM Storage Scale moves and protects state across edge, core, and cloud by policy. The compute ontology names what each resource is: CPU for control, GPU for dense training and large reasoning, quantum for the bounded problems where its physics wins, storage for memory and lifecycle, and Akida for the always-on, event-driven sensing and adaptation that a rack of conventional compute can't stand watch over.
The numbers say this fabric is not a sketch. In my NAECON 2026 work with Steven Harbour, a 46-node Akida fleet under Symphony held 8,832 concurrent inference contexts at 99.4 percent of ideal scaling, dispatched in 5 milliseconds, hot-swapped a hundred resident models at 11 milliseconds each, and cut GPU cost by 30 to 50 percent by routing work to the tier that actually needed it. Heterogeneity is not a complication to be abstracted away, it is the precondition for owning anything.
A fleet that teaches itself
There is one dependency left in the picture and it is the deepest one. Even a system that infers locally usually trains centrally. Models are born in GPU farms, shipped outward, and periodically called home to be retrained on data that has to travel back with them. As long as that is true, the center still holds the compounding loop and every site is a customer of its own experience.
So the next stage of this program is a neuromorphic fleet that trains without outside help. Not by asking Akida to imitate a GPU but by using the methods that suit a fleet of small learners. Fixed rich reservoirs supply features that never need training and chips learn the readout on the spot. Local plasticity rules, the kind brains use instead of backpropagation, shape feature layers on FPGA fabric beside the chip. Gradient-free search treats the fleet itself as the optimizer with every chip evaluating a different candidate and only seeds and scores crossing the network. Agreement among many chips becomes the labels that teach any one of them, so the neuromorphic hive trains the individual.
The capacity to do that already exists. Because BrainChip's SDK is bit-for-bit identical to the silicon, the same 46 nodes that held 8,832 concurrent contexts become roughly nine thousand evaluators for a search, more than six times the 1,440 CPU cores landmark evolution-strategies work used. In September, Symphony HostFactory brought up 131 Akida nodes in 37 cities across 20 countries and six continents, from nothing to a fleet answering in 14 minutes and 50 seconds, for about a dollar forty, on commodity clouds rather than the hyperscalers. Every node from Johannesburg to Santiago returned the same classification, bit for bit.
Put those together and the shape is clear. Train broadly on a bit-exact fleet that any nation or company can own outright. Execute on silicon beside the work. Keep learning on the chip in the field. Promote what proves itself, roll back what does not, and never require the knowledge to leave home in order to grow. A shipyard that can train its own models from its own experience is not a tenant of anyone's intelligence. It is a place that learns.
Sovereignty at every scale
The same principle holds at every size which is how you know it is a principle and not a product feature.
The machine. A crane on a dry dock, a press at a munitions plant, a transformer at a substation each learns its own normal with a chip whose card sells for not much more than a night out on the town . What it learned is a few kilobytes, belongs to the machine's owner, and it never has to leave the floor to be useful.
The site. A shipyard going from two hulls a year to twenty cannot afford to ship every weld signature and vibration trace to a central platform and wait. With a site fleet under a heterogeneous compute ontology, every station learns locally, the yard trains and promotes its own models, and leadership, whether in Philadelphia or in Seoul, sees answers across the yard without the yard's production record ever leaving the fence line.
The town and the county. This is the scale extractive architectures forget entirely. A county that hosts the river terminal, the rail junction, and the substation can own the fleet that watches them. The freight intelligence, the emissions record, the grid health, all of it can live in a public institution or a local firm rather than in a vendor's catalog, and it can be the asset a development authority shows the next manufacturer that comes looking.
The vessel and the satellite. An autonomous boat or a satellite cannot depend on a link home. It receives models as signed, encrypted bundles, learns on board, and sends down verdicts rather than video. Capture it and you get a hardened host and ciphertext. The physics of the chip and the politics of the data point the same direction.
The nation. A free country should hold its own keys. A sovereign fleet, trained on its own bit-exact infrastructure and operated by its own people, is a purchase of capability rather than a subscription to it. Encrypted computation, an egress gate that controls exactly what leaves, and consensus across chips so that no single compromised node changes the answer: I have built and demonstrated each of these pieces, and together they mean that the operator can run the fleet without being able to read it.
The alliance. Free nations do not need one platform at the center of their defense industrial base. They need federation. BrainChip is listed in Sydney, the Philly yard is Korean-owned, and the allies rebuilding their own industry stretch from Tokyo to Warsaw. Each partner keeps its own region and its own keys, and they exchange answers and threat signatures under policy, not raw data under a vendor's terms. This avoids a weaker alliance and operates without a single point of failure, which is the only kind worth having in a contested century.
Many centers, each genuinely different, held together as one working whole without being collapsed into sameness and without being split apart. I have written elsewhere about the older grammar underneath that sentence. Here it is enough to say that it is also a network topology and it is the one a rich and free world ought to choose.
The edge nobody has priced
All this is yet another reason I keep saying the edge market is seriously underestimated. In The Neuromorphic Market Analysts Have Yet to See, I laid out how the published forecasts box neuromorphic computing into milliwatt sensing for drones, wearables, and IoT, then size the category from that box inward. The estimates for the same market in the same year, 2026, run from roughly $125 million to more than $8 billion. A spread of more than sixty times is not a disagreement about data. It is a disagreement about where the category ends, and every one of those boundaries is drawn too close.
Reindustrialization moves the boundary again, in a direction the reports have not modeled at all. The edge they count is consumer and gadget-shaped. The edge being built in Ohio, Philadelphia, Louisiana, and Texas is the arsenal itself. Every airframe on a line designed for tens of thousands of weapons a year, every autonomous hull from a yard aiming at twenty ships a year, every crane, press, weld cell, substation, pipeline segment, rail junction, and barge terminal that keeps those lines running is an edge node that has to sense, decide, and learn under a power budget and without a reliable link home. None of that appears in a forecast built around smart watches.
The sovereignty argument multiplies the effect. If the knowledge a shipyard or a munitions plant generates must stay on site, and in a contested century it must, then the work that the current model routes to a distant data center has to happen at the edge instead. That is not only inference. Once a fleet can train itself from its own experience, a share of the training budget moves out to the edge with it. The edge stops being the cheap periphery of the AI market and becomes the place where a growing fraction of the market's work is actually done. Really, this is all about perspective and our consideration about what constitutes the edge needs to think more deeply about where and what it really is.
The power arithmetic closes the case. Pennsylvania's industrial customers are already paying for the gigawatts being built to serve data centers. Every recognition task handled on a chip drawing a fraction of a watt beside the machine is a task that never has to be shipped, stored, and processed on the grid's most contested capacity. My paper sized the opportunity from the workload outward: an inference envelope in the hundreds of billions a year, an embodied-intelligence layer of comparable scale, and an efficiency dividend on compute already committed. The reindustrialized edge belongs in every one of those lines, and analyst reports have not yet counted it in any of them.
Starting where the rivers meet
If this is going to be built, it should start in the place that has paid most for the other way of building things. That's likely your town and mine. Western Pennsylvania sits between Arsenal-1 in Ohio and the Philly yard in the east. It has three navigable rivers, two Class I railroads, a century of metalworking and electrical manufacturing in its bones, and a robotics culture in Pittsburgh that the rest of the country borrows from. It also already has a neuromorphic fleet: ten or twenty chips near Pittsburgh that served one vLLM endpoint alongside nodes in Dallas and Washington, placed live, in July.
The practical start is not a megaproject. It is four modest things done well.
1. Instrument the corridor and let it keep its own knowledge. A river terminal, a rail-served brownfield, and a gas gathering area, each watched by a local fleet owned by a county authority or a regional firm. The freight picture, the emissions record, and the power-quality data become public assets that make the next site easier to sell, not telemetry that makes someone else's platform more valuable.
1. Build the trusted module and provisioning center here. Akida modules for American and allied programs need to be assembled, ruggedized, loaded with keys, attested, and tested by people with clearances in a place with the right logistics. That is light electronics manufacturing a rural site with a trained workforce can win and it puts the region at the front of the supply chain for the autonomous systems everyone else is building.
1. Make the neuromorphic lab a standard fixture of community colleges and tech schools. A ten-node cluster of small hosts and Akida chips costs less than a single high-end GPU server. Every welding and electrical program within fifty miles of the rivers should have one so the people who maintain the new industrial base can also maintain the intelligence inside it.
1. Give the old product a new nervous system. The rye is coming back to the Monongahela. A craft distillery tracking fermentation and barrel aging on a chip in its own rickhouse is a small customer, but it is a symbolic one, and a region that once went to war over who counts the value of its whiskey would enjoy the symmetry.
From there, the corridor becomes a reference design: the place allied delegations visit to see a non-extractive industrial base actually running, before they build their own.
Rich and free, and staying that way
I don't really like the word "developed" for the countries now trying to rebuild their industry. Being a developed country implies a finished state, puts it conveniently over so-called developing countries, and it carries a whole theory of history in which development is something one place does to another. What these nations actually are is rich and free and the uncomfortable truth is that they are not guaranteed to stay either. Wealth that compounds somewhere else stops being yours. Freedom that runs through infrastructure you cannot carry out the door is nothing more than permission.
The reindustrialization now underway will decide a great deal about both, it will put millions of learning machines into factories, shipyards, vessels, substations, and orbits. If we wire them the familiar way, the learning will flow to whoever holds the center, and we will have rebuilt the arsenal while quietly handing away the knowledge of how to run it. If we wire them the other way, every machine, yard, county, and nation becomes a place where knowledge accumulates for the people who did the work.
The hardware to choose the second way is shipping. The fabric to coordinate it comes from a high performance computing lineage that has run demanding workloads for decades. The evidence is in the record, eighty-two demonstrations deep and counting. What remains is the decision and it is the same one the rye farmers on the Monongahela understood two centuries ago. Count the value where it is made.
Only the answer has to move.
Sources
- Anduril and Arsenal-1: JobsOhio, Orange County Business Journal, Axios Columbus
- Shipbuilding: Industrial Info Resources, Pulse 2.0 on Saronic
- BrainChip: June 2026 quarterly update, Parsons agreement
- Monaca: Spotlight PA, Ohio River Valley Institute
- Homer City and power prices: Utility Dive, Sierra
- Demonstrations: #27 SymRail, #43 SymGrayZone, #57 SymBasin, #70 Akida Cascade, #71 GPS time spoof, #72 Only the Answer Moves, #82 Run Akida Anywhere
- Papers: Neuromorphic Orchestration for Efficient LLM Infrastructure, IEEE NAECON 2026
- Prior essays: Down in the River, Old Black Water
- Shell sale process: Hydrocarbon Engineering
- Whiskey Rebellion: HISTORY; Monongahela rye revival: WHYY
- Akida M.2 card specs and price: BrainChip shop
- Evolution strategies at 1,440 cores: Salimans et al., 2017
- LSF and Platform Computing since 1992: The Register
- Market analysis: The Neuromorphic Market Analysts Have Yet to See