I was reading a casual description of Von Neumann probes recently, the kind of thing an AI assistant produces when you ask an idle question. One probe builds two. Two build four. The galaxy gets colonised in a few million years. The Fermi paradox is puzzling because the math is so clean.
Earlier the same month, I'd been working through a Mars colonisation requirements document. Its synthesis chapter contained a flat, unhappy sentence: the semiconductor supply chain has ten or more tiers of depth, requires nation-state-scale industry, and any Mars colony will depend on Earth for advanced electronics for decades to centuries, regardless of progress in any other domain.
I noticed the two claims couldn't both be true. If a probe can build its own chips from asteroid dust, the Mars document is wrong about semiconductors. If the Mars document is right, Von Neumann probes can't exist in the form the math assumes. I didn't know which one was wrong. So I worked it out.
What Von Neumann actually proposed
In the late 1940s, John von Neumann designed a theoretical universal constructor: a machine that could build a perfect copy of itself from raw materials. The design had four parts. A blueprint with the instructions. A factory to gather and assemble. A copier to duplicate the blueprint. A controller to run the whole thing.
The power of the design is that, in principle, all four parts can be built from the same substrate the constructor mines. The exponential math follows from that. If the probe can produce every component of itself, replication time stays constant across generations, and the doubling never slows.
The word doing all the work in that paragraph is controller. In Von Neumann's 1948 thought experiment, the controller was a simple cybernetic mechanism, the kind of thing you could imagine being fabricated with hand tools and foresight. For a real probe doing real mining, real refining, real assembly, and real interstellar flight, the controller is a modern processor stack. And a modern processor stack, as the Mars document notes with the quiet precision of an engineer who has actually done the sums, is not something a single machine can produce.
How deep the stack actually goes
If you want a flight computer that works, you need chips. Chips come from fabs. Fabs need lithography machines. The ones at the frontier are built by one company in the Netherlands.
The mirrors in those machines are polished to tolerances measured in picometres across half a metre of surface, among the most precise objects ever manufactured. The light source fires droplets of molten tin forty thousand times per second and hits each one with a pulsed laser. The resist chemistry depends on specialised fluoropolymers. The metrology depends on other chips made on earlier-generation fabs. The cleanrooms depend on industrial gas supply. The gas supply depends on rare earth separations. The rare earth separations depend on mines in specific countries. Every tier has a tier beneath it. The bottom of the stack sits in the ground, in supply chains, in geopolitics.
No probe can contain this. A probe that tried would have to be the size of a small country, at which point it isn't a probe, it's a civilisation that moved. Civilisations have moved before. It takes a while.
This is the replication constraint. A self-replicating system is capped in replication rate by the most complex thing in its stack, and the most complex thing is always several tiers deeper than the replicator itself. Exponential math that ignores this is describing a system that doesn't exist.
Biology is the only exception, and it's instructive
The objection that usually arrives here is that biology is an actual self-replicating system at small scale. Cells copy themselves. They don't ship electronics from Earth. So the constraint can't be absolute.
This is almost right, and the reason it's almost right is the most useful thing in the whole argument.
Biology works because the four Von Neumann roles aren't separate components. DNA is both blueprint and part of the copier. Ribosomes are factories that are themselves produced by ribosomes. Enzymes catalyse the reactions that produce more enzymes. The entire stack is collapsed to a single molecular tier. Nothing in the cell depends on an industrial process the cell can't do itself.
Engineered replicators have never managed this. The RepRap project started in 2005 with the goal of making a 3D printer that could print its own parts. Twenty years later, every RepRap still needs bought-in screws, motors, heating elements, belts, bearings, and the chips on the control board. The fraction of a printer that prints itself has crept up slowly. It has never approached the threshold where one printer can, unaided, bootstrap another one from raw materials. Every attempt to close the loop reveals a new tier underneath.
The strongest version of the objection is that molecular nanotechnology, in the Drexler sense, would someday collapse all the tiers into a single substrate and reach biology's trick through engineering rather than evolution. I take this seriously, but not all the way. Either that nanotechnology would need its own controller stack, which pushes the constraint down one level rather than eliminating it, or it wouldn't need one, in which case it would functionally be biology, and biology has had four billion years to build kinematic replicators at engineering scale and hasn't. The absence is evidence.
Why this matters for aerospace
Aerospace is where the constraint bites hardest and most visibly, because aerospace systems are the ones that most obviously want to replicate, and are the furthest from being able to.
A modern launch vehicle sits on top of a long stack. Specialist alloys from a handful of foundries. Rocket-grade composites from a smaller number of fabricators. Flight computers on process nodes from fabs worth billions. Inertial measurement units calibrated at sub-arcsecond precision by vendors you can count on one hand. Software stacks whose verification involves institutional processes no single engineering team can replicate. Every Starship sits on this. No matter how vertically integrated SpaceX becomes, the stack runs deeper than SpaceX.
Scaling this to interstellar operations doesn't relax the constraint. It tightens it. An autonomous probe meant to run without resupply for decades, mine and refine at its destination, manufacture replacement components, and build a successor, needs its controller tier to be either carried in full, which makes the probe civilisation-scale, or manufactured locally from raw materials, which puts you back at the Mars document's unsolvable problem. There is no third option in current physics.
The interesting engineering question, then, isn't how to make things replicate. It's how close you can bring the controller tier to the replicator scale before the physics, chemistry, and tolerances stop you. That question doesn't have a known answer. I think it's one of the more important open problems in manufacturing science, and almost nobody is framing it that way.
What this does to the Fermi paradox
The standard Fermi argument says a single spacefaring civilisation could colonise the galaxy in a few million years because probes replicate exponentially. The replication constraint breaks that argument, not by removing the paradox, but by dulling its sharpest edge.
If replication time grows with each generation, because every new probe has to either import its controller tier from an increasingly distant supply chain or rebuild it locally at civilisation scale, then galactic dispersal slows by several orders of magnitude. Millions of years becomes hundreds of millions, maybe billions. A civilisation expanding now wouldn't be visible yet. The silence gets easier to explain without invoking any filter at all.
This doesn't solve the paradox. The great filter, rare Earth, and dark forest explanations are still in play. But one of the paradox's most load-bearing numbers, the short exponential dispersal time, turns out to rest on a physical assumption that doesn't hold. That's worth noticing.
Where I sit
I'm nowhere near aerospace R&D right now. What I'm doing today is software and industrial operations, several tiers of the Von Neumann diagram away from the stack this piece is about. Pretending otherwise would be silly.
The reason I wanted to work this out is because a later-stage piece of what I'm building will have to address the aerospace-adjacent manufacturing questions the replication constraint describes. I'd rather commit publicly to the constraint now than quietly have the option to ignore it later. If the constraint is real, any serious work in this space has to be framed around stack-collapse, not around replication rate. Framing it around replication rate is the mistake this piece exists to identify.
So the honest version of my position is small. I'm an operator running early-stage software and industrial companies, with a long intention to work in aerospace R&D once the operating base supports it, and I'm writing this down now because the thesis will outlast the interval before the work becomes possible. Published in 2026, it's harder to quietly revise in 2032.
The asymmetry
Planners who take the replication constraint seriously will build slower-looking systems that actually work. Planners who ignore it will build faster-looking systems that stall at the first tier of the stack they can't themselves produce.
This isn't abstract. The industries that currently supply the aerospace controller tier (lithography, advanced composites, specialist alloys) are geopolitically concentrated in a way most planners have noticed but few have absorbed. A future in which those supplies aren't reliably available isn't a fringe scenario. Several serious governments are planning around it already. The programmes that correctly identify which tiers are replicable and which aren't will survive the next twenty years. The ones still doing Von Neumann math will read well in press releases and not fly.
The practical payoff is that the replication constraint is a diagnostic. Apply it to any system claiming exponential growth. Find the tier beneath the tier being counted. Ask whether it scales at the same rate. If it doesn't, the exponential claim is wrong in a way that matters. If it does, you've found something rare, and it's worth paying close attention to.
Summary
- Von Neumann's universal constructor assumes the blueprint, factory, copier, and controller can all be built at the same scale from the same raw materials. Modern controller stacks, especially semiconductors, need ten or more tiers of infrastructure no single replicator can contain. This is the replication constraint: a self-replicating system is capped by its most complex component, and the most complex component is always several tiers deeper than the replicator itself.
- Biology is the only existence proof of self-replication at small scale, and it works because the Von Neumann roles aren't separate components. DNA is both blueprint and copier. Ribosomes make ribosomes. The stack is collapsed to one tier. No engineered system has ever done this, which is why seventy-five years of serious attempts have produced nothing close.
- The strongest counter-argument is molecular nanotechnology. Either it would need its own controller stack, which pushes the constraint down rather than eliminating it, or it wouldn't, in which case it would functionally be biology, which has had four billion years to build engineering-scale replicators and hasn't.
- The Fermi paradox's short exponential-dispersal timescale relies on the Von Neumann assumption. With the replication constraint accounted for, galactic expansion slows by several orders of magnitude, which makes current silence much less surprising without resolving the paradox.
- A lot of current forecasting (AI foom, crypto adoption curves, autonomous fleet projections) inherits the Von Neumann assumption without examining it. The replication constraint is a diagnostic. Find the tier beneath the tier being counted. If it doesn't scale at the same rate, the exponential claim is wrong.
- The real engineering question isn't how to replicate. It's how close you can bring the controller tier to the replicator scale before the physics stops you. That's where serious aerospace R&D in the 2030s will have to live.
- Bora