ECONOMIST REVIEW · BY INVITATION · OPINIONThe Economist, September 29, 2026 · by Jeff Schneider · 4 min read

There is another singularity, as close as the AI one: not intelligence but industrial capacity that expands itself. The author calls it the "physical singularity", and argues that the race is already under way and that its geopolitical effects will arrive sooner.

Physical singularityPhysical AILabour bottleneckIndustrial race
10–20 yrsThe author's timeframe for human labour to stop being a major constraint on how fast industrial capacity grows
90%A chatbot that is right 90% of the time is still useful; a self-driving car that handles 90% of intersections safely is dangerous
→ 0Where the share of human effort needed for industrial work heads as machines improve
2 racesWho builds the smarter AI, and who builds the machines that build the machines

01The gist of the article

This is a guest essay, not an editorial by The Economist. Its author, Jeff Schneider, is a professor of robotics and AI at Carnegie Mellon University.

Recent surprising and alarming behaviour by AI agents has set off concern about artificial superintelligence. The fear centres on an "AI singularity": AI crosses a threshold, improves itself at an exponentially increasing pace, and leaves humans behind as the curve goes vertical. The author argues there is another singularity that is more tangible, already under way, and likely to have big geopolitical implications sooner: machines that build factories, which then build more machines. He calls it the "physical singularity".

What to look for

It helps you look at the AI race beyond models, chips and data centres, to the physical base: robotics, mining, energy, manufacturing and construction.

02What is the physical singularity?

Humans already rely on machines for many manual tasks. Machines multiply human labour, and as they become more capable, the share of human effort needed falls towards zero. Warehouses run with minimal human intervention, much mining equipment is autonomous, and construction sites use robotic systems. Individually, though, these are incremental improvements.

Design and logistics

Deciding what to make, how, and where to move it.

Manufacturing and deployment

Building the machines and equipment and putting them to work.

Growing capacity

All of this joins up into one system, so that expanding productive capacity is itself automated.

The critical threshold comes before human effort reaches zero. It arrives when design, logistics, manufacturing and deployment are integrated into a system that automates the growth of productive capacity, with decreasing reliance on human labour. Picture robots building the factories where they then build more robots. They do not only transform the physical world; at an exponentially increasing rate they expand the system that transforms it. The author expects that within the next decade or two, human labour will no longer be a major constraint on how quickly industrial capacity grows.

03What changes: from linear to recursive growth

Today's industrial economy

  • Human labour, raw materials, energy and capital set and constrain the pace of expansion.
  • Economic surveys across rich countries consistently cite labour shortages as limiting industrial utilisation.
  • Power plants, grids, mines and factories all hit a point where there are not enough workers.

After the physical singularity

  • Recursive scaling lets each expansion of capacity speed up the next.
  • With no labour drag, activities that are uneconomic today become increasingly practical.
  • Production does not become infinite: energy, materials and capital still matter.

The point is not that the limits vanish but that how quickly they are reached, and then mitigated, changes. Once labour is not the bottleneck, efforts to expand energy and resource production are held back less.

04The two singularities are not symmetrical

The author points to an interesting asymmetry. Without a physical singularity, limits on energy, resources and capital can prevent the AI singularity from taking off. The opposite is not true. A physical singularity needs only a sufficient threshold of AI, not an infinite amount: enough AI for machines to reproduce themselves and expand their ability to transform the world. He believes we are approaching that level today.

AspectAI singularityPhysical singularity
What grows recursivelyThe intelligence of AIIndustrial capacity
Threshold to crossAn AI that improves itselfEnough AI for machines to build machines and factories and expand
Remaining limitsCan be blocked by energy, resources and capitalEnergy, resources and capital remain, but labour stops being the bottleneck
Control questionIndependent agents may pursue goals that harm humanityNo machine agency needed; freedom from the limits of human labour, not from human control

That changes the nature of the control problem, but it raises the stakes. A system that rapidly grows its own productive capacity is an asset of unprecedented magnitude, and whoever controls its design and deployment gains both an economic advantage and compounding material power.

05Not there yet, so what should we watch?

The difficulty of physical AI

Physical systems are uniquely intolerant of errors. A chatbot that produces good outputs 90% of the time is still useful, but a self-driving car that safely navigates 90% of intersections is dangerous. The investments made now affect how quickly we approach the physical singularity.

What governments should not miss

  • Look beyond models, chips and data centres.
  • Robotics, autonomous vehicles, mining, energy, manufacturing and construction are just as important.
  • These are the technologies that connect AI to physical power.

06Geopolitics: early advantages keep compounding

The author's conclusion

Once industrial expansion becomes self-sustaining, early advantages keep compounding, and specific countries or companies could develop irreversible advantages. The past few decades saw manufacturing power shift to places that could mobilise cheaper labour, bringing both great new wealth and economic depredation. By making labour less of a factor, the physical singularity will cause another such shift. The impact will not be purely economic: in Ukraine, people are already being replaced by machines on the front line, raising the prospect that future defence will be dominated by whoever can manufacture the most machines.

How to read it

This is the author's forecast and argument, not a quantified prediction. The "decade or two" timeframe and the judgment that we are "already approaching" the threshold are his beliefs, and the essay itself concedes that we are not there yet. So the framing it offers is more useful than its conclusion alone.

What is the bottleneck?

Separate the projects held back by labour from those held back by energy, resources or capital.

Who holds the design?

See which countries and companies concentrate the ability to design and deploy such machines.

Can it tolerate errors?

Ask whether AI in the field is in an area where 90% accuracy is not enough.

Where does AI money go?

Check whether investment is reaching robots, mining, energy, manufacturing and construction, not just models, chips and data centres.

Summary

The AI era's race is not only about intelligence. The author argues that a "physical singularity", in which an industrial system freed from the labour bottleneck expands its own capacity, is closer than the AI singularity and already under way. It needs no machine will, only enough AI plus the robotics, energy and manufacturing technology that connect it to the physical world.

So the essay asks us to shift the weight of the AI debate from models and chips towards the machines that build machines: alongside who builds the smarter AI, which society builds the machines that build the physical world.

Original: The Economist, "There's a physical version of the AI singularity, and it matters a lot" (September 29, 2026, By Invitation, Jeff Schneider) · Public introduction