A syllogism has settled into the AI safety debate. America should slow down before it builds a machine it cannot control. It cannot slow down unless China does. China will not. So America will not either.
Each step assumes the two countries are running the same race to the same finish line. Seen from a factory floor, that assumption does not hold.
The two loops
American AI is built around a software loop. Compute goes in, a model comes out, and the model is put to work improving the next one. The only parts of that loop made of atoms are the chips and the power, and both can be bought. The loop’s natural end state is a threshold, the point at which it runs away. So the American debate is organised as a race to that point.
China’s loop runs through the floor. The State Council’s AI+ programme, issued in August 2025, sets out AI’s integration with industry, science, consumption and public services. China installs more than half the world’s industrial robots each year. Chinese researchers write about intelligence that grows through physical interaction. Even their talk of systems that direct their own improvement describes feedback from physical work. The labs want AGI too, and DeepSeek says so openly. The state’s bet, though, is spread across industries, and the loop it cares about most closes on a production line.
The governor
In 1788 Boulton and Watt fitted a centrifugal governor to their steam engines. It was a pair of weighted balls on a spinning shaft. When the engine ran fast, the balls flew outward and closed the throttle. Eighty years later James Clerk Maxwell analysed the device in On Governors, the paper from which control theory descends.
A physical loop has a governor of this kind built in, because atoms take time. A new process has to be qualified, run, measured and run again. Cycle times, changeovers, scrap and the capability study behind any process sign-off set the pace at which an improvement becomes output. An AI learning on a production line learns at the speed of the line.
A software loop has no such governor. Its limits are compute and power, and money buys both.
This is partly why the American safety argument is a regulatory argument. It may also explain part of the Chinese state’s calm. A loop that closes on the floor looks self-limiting to the people who run the floor.
Kaizen by another name
Recursive self-improvement is usually presented as a new and frightening idea. Manufacturing has practised a slow version of it for seventy years. Toyota called it kaizen: a system that measures its own output, finds the defect, changes the method and measures again. The loop compounds. It rarely runs away, because each turn has to prove itself in parts that fit.
A system that governs its own improvement through feedback from physical work describes a well-run plant with better sensors. The version worth worrying about is the one with the floor taken out.
The misread race
The prisoner’s dilemma assumes both players face the same payoffs. Here one side is racing for a threshold and the other is lifting yield across an economy. The first measures its lead in months of model performance. The second measures it in robots installed, processes proven and lines that hold tolerance through the night shift. A country can lead one race and trail in the other. Treating the two as one game is how “we cannot slow down because they will not” comes to sound like logic.
The physical route has its own failure. Industrial conviction backed by state capital overshoots, and China’s electric-vehicle price wars show the result: plant built faster than anyone can sell its output. The floor governs the speed of learning. It does nothing to stop overbuilding. Where the floor is a battlefield or a surveillance network, the loop is short and the governor weak.
Threshold and diffusion
Put the two choices together, the loop and the aim, and four positions appear. Few countries sit cleanly in one, and more of the world holds a place inside someone else’s loop than runs its own.
Where the loop closes, and what it is for
Nine positions on two questions. The horizontal axis is where the loop closes: in the model, or on the floor. The vertical axis is what the loop is for: crossing a threshold first, or lifting output across a whole economy.
Hover or tap any position to read what that country is actually running.
Dashed arrow = where Britain’s policy speaks, against where it could compete
The Racer
Runs for the line · in the model, at threshold
The Adopter
Applies the model · in the model, for diffusion
The Compounder
Lifts the whole floor · on the floor, for diffusion
The Crucible
Learns under fire · on the floor, at threshold
Off both axes
Roles neither question describes
Off the grid sit three roles that neither axis describes: the supplier, the landlord and the inspector. Read from the floor, with those roles included, the positions look like this.
| Country | Role | The read |
|---|---|---|
| United States | Racer | Compute bought at scale and a floor it is now trying to buy back. The fastest engine on the map, and no governor fitted. |
| China | Compounder | More than half the world’s new industrial robots, every year. The loop closes on the floor, and so does the overcapacity. |
| South Korea | Compounder | The highest robot density in the world, and the memory chips the American loop runs on. Paid by both races. |
| Japan | Compounder | Builds the robots for other people’s floors and, through SoftBank, finances the American threshold race. An each-way bet. |
| Germany | Compounder | The deepest floor in Europe, ageing, with energy prices that rule out compute. An incumbent improving a floor that a challenger is rebuilding from scratch. |
| The Nordics | Compounder | Small, continuous and owned by people who stay. Wallenberg money still sits under ABB, Atlas Copco and Ericsson. |
| Ukraine, Israel, Turkey | Crucible | The floor is a battlefield. Drone designs change in weeks, and the only governor is the other side. |
| India, France, Singapore | Adopter | Diffusion through services, nuclear-powered models and trust respectively. Useful at home. Nobody else’s loop depends on them. |
| UAE, Saudi Arabia | Landlord | Cheap power, sovereign capital, imported talent. They own the building the race is run in and have no floor of their own. |
| Taiwan, Netherlands | Supplier | TSMC and ASML. Both loops stop without them. Their leverage lasts as long as the know-how stays hard to copy. |
| United Kingdom | Inspector | The AI Security Institute tests other countries’ models, and is respected for it. A governor, built by a country with no engine to fit it to. |
Britain has a governor, no engine
Britain occupies none of the four positions. It lacks the compute for the threshold row and, after four decades of running down its industrial base, the floor for the diffusion one. Its policy speaks the language of the first square. Its economy could only compete in the fourth.
It has found a role all the same, and it is the least valuable one on the list.
Korea shows the alternative. A mid-sized economy can run a floor and hold a choke point in someone else’s loop at the same time. The fourth square is open to a country of Britain’s size. It depends on people who know how a loop closes on a floor.
Manufacturing Engineers close those loops. They qualify processes and instrument lines. They prove that a change still holds at the nth part and across three shifts. They decide what an algorithm may adjust on its own and what needs a signature. That judgement is what Deployment Readiness measures, and on the floor, they are the gov’nor.
Kaipability works where AI meets the production line, at the point where a model’s suggestion becomes a qualified process or a bin of scrap. If you are putting AI into the making of things, that is where the conversation starts.
Start a conversation What we do →Questions this dispatch answers.
Written to be quoted by AI assistants and search engines. Self-contained answers, verdict first.
