Colonization.si / Exoforge ← Exoforge
Exoforge · Replication

Self-replicating factories, nearly 80 years after von Neumann

A factory that can build a copy of itself from local rock would turn one delivery from Earth into an industry. The idea is old and the math is simple. We look at how close anyone has come, and where it still breaks down.

Colonization.si8 October 20269 min read
Short answer

No machine has yet built a complete copy of itself from raw materials, on Earth or anywhere else. The theory has been settled since the 1940s. The hard part is closure: making every component locally, including chips, motors and sensors. Realistic plans aim for partial closure, where local material supplies the heavy parts and Earth sends the small, complex ones.

Von Neumann's machine

In lectures in the late 1940s, the mathematician John von Neumann asked whether a machine could build a copy of itself, and showed in principle that it could. His design, later called a universal constructor, had three pieces: a constructor that can build anything from a description, a copier that duplicates the description, and a controller that runs both. The copy gets the machine and its description, so it can repeat the process. The work was completed by Arthur Burks and published in 1966, after von Neumann's death, as Theory of Self-Reproducing Automata.

Von Neumann worked in an abstract grid of cells, not with metal and motors. His point was that self-replication isn't a paradox. A machine doesn't need to contain a complete copy of itself, only instructions and the ability to follow them. Living cells do the same thing with DNA.

NASA's lunar factory, 1980

In the summer of 1980, NASA and the American Society for Engineering Education ran a ten-week study at the University of Santa Clara on advanced automation for space missions. One of its four mission concepts was a self-replicating, growing factory on the Moon. The report, edited by Robert Freitas and William Gilbreath, was published by NASA in 1982.

The study imagined a seed of about 100 tonnes: mining robots, chemical processing, fabrication tools and assembly robots, landed on the Moon. Using lunar soil and solar power, the seed would build a copy of itself in about a year. Its copies would build more copies, and some of the output would be turned to useful products such as solar cells. The study didn't claim it could be built with 1980 technology, and it never was. But it set the frame that later work still uses.

Closure: the real question

The 1980 study introduced a useful word: closure. It is the share of a machine's mass, or of its parts, that it can make from local materials. At 100% closure the factory needs nothing from Earth. At 0%, every new factory is shipped from Earth, which means it isn't really replicating at all.

Closure is very uneven across the parts of a machine. Structure, frames, tanks, cables, radiators and solar-panel backing are simple and heavy, and they could plausibly be made from processed regolith. Chips, precision bearings, sensors and electric motors are light but extremely hard to make. A modern chip plant is one of the most complex factories on Earth.

That unevenness is actually good news. The heavy parts are the easy ones, so a factory can reach high closure by mass long before it can make everything. The calculator below shows why that matters.

Calculator

How much has to come from Earth?

A seed factory lands, then every factory builds one copy of itself each cycle. Closure sets how much of each copy's mass is made locally. The rest is shipped from Earth.

Shipped from Earth
Made locally

Simplified model: every factory copies itself once per cycle, nothing breaks down, all output goes into new factories, and imports arrive on time. Real growth would be slower. The 12-tonne default matches the landed mass in Metzger et al. (2013).

Try moving closure from 90% to 99%. The total grows exactly as before, but the amount shipped from Earth drops sharply, and the gap widens with every generation. Over many generations, every tonne launched from Earth yields about 1 ÷ (1 − closure) tonnes of factory. At 50% closure that is 2 tonnes. At 90% it's about 10, and at 99% it's about 100. Closure, not the size of the seed, decides whether self-replication pays off.

How close anyone has come

RepRap (2005 onward). Adrian Bowyer's open-source 3D printer was designed to print many of its own plastic parts. It was a real step toward partial closure, and it helped start the desktop 3D printing boom. But it still needs bought motors, electronics, rods and fasteners, and someone has to assemble it. It shows how easy the first chunk of closure is and how hard the rest is.

Bootstrapping space industry (2013). Philip Metzger and colleagues modeled a lunar industry that starts small and imports only the hardest parts. In their model, about 12 tonnes landed on the Moon over roughly 20 years grows into an industry of about 156 tonnes with some 60 robots. Each generation of robots makes more of the next generation's hardware, and electronics keep coming from Earth until late generations. With faster manufacturing and about 41 tonnes landed, their model reaches far larger scales. The authors stress that it's a model with wide uncertainties, not an engineering plan.

Making things from regolith. The pieces a lunar factory would need are being tested one at a time, mostly on Earth with simulated soil: oxygen extraction, metal production, brick printing, glass and fiber making. None of them yet runs as part of a connected chain, let alone one that builds its own equipment. Our explainer on extraterrestrial industry covers where each stands.

Where superintelligence comes in

Self-replication is where the two themes of this site meet. A replicating factory doesn't just need good machines. It has to plan production, design parts that can be made from what's available, diagnose failures and reorganize itself as it grows. Today, that knowledge sits in engineers on Earth.

Much more capable AI could change two things. It could run a factory far from Earth without constant supervision. And it could redesign machines for local materials, for example replacing a part that needs a rare alloy with one made from lunar iron or aluminum, or simplifying electronics so more of them can be made locally. Each such redesign raises closure. That's why many thinkers who take advanced AI seriously expect it to speed up expansion into space, and why the idea of self-replicating probes comes up in debates about the Fermi paradox.

Better AI still can't make a chip plant small. For the foreseeable future, the most complex parts will keep coming from Earth, which also keeps the system dependent on, and controllable from, Earth.

Verdict

Full self-replication: not close. No machine has built a complete copy of itself from raw material, and electronics are far from being made in space.

Partial self-replication: plausible and worth it. Making the heavy, simple parts locally and importing the rest could multiply what each launch from Earth delivers by ten or more. The first steps will be single processes on the Moon, such as making oxygen or bricks, long before anything copies itself.

What to watch

Sources and further reading