For the last two years I have written about AI from inside a factory — which photos it can retouch, which spreadsheet it can clean, which label it can print. That is the small view, and it is the one I trust most. But the small view keeps pointing at a larger question: if AI can do all of this, what happens to the work, and to the people who were doing it?

This piece is an attempt to follow that question all the way out. The figures below are the ones circulating in Chinese industry reporting, and they should be read as directional, not precise. The logic matters more than the decimals.

1. What AI is already replacing

The boundary of automation has moved. It used to stop at the wrist. Now it goes straight through the desk.

Three layers are going first:

• Standardised physical work — assembly lines, parcel sorting, port transport, mining. Robots do not get tired, do not take breaks, and do not make the same mistake twice.
• Standardised cognitive work — basic copywriting, junior coding, data entry, tier-one customer service, routine translation, standard financial analysis.
• Process execution — administrative coordination, meeting scheduling, checklist-style recruiting screens.

The important word is task, not job. AI rarely removes an entire profession at once. It removes the repetitive, mechanical, time-consuming part of a profession and leaves the rest. A Java engineer is a good example: AI can take the boilerplate coding and the documentation, but system architecture and the diagnosis of a live outage still need a person who understands the business.

The speed is what has changed. In 2025, autonomous AI alone was reported to have eliminated more than 50,000 white-collar positions, with another estimate putting the total closer to 170,000. Several of the large international consultancies have paused campus recruiting for analyst roles — which is a quiet way of saying they expect the work those analysts did to be done differently.

2. What AI cannot replace

There are categories where a human is not a temporary substitute for a better machine:

• Genuine emotional presence — therapy, high-end care, education that involves actually raising a child.
• Deep cross-domain creation — strategy, art, scientific breakthroughs.
• Legal and moral responsibility — medical decisions, judicial discretion, the direction of a company. Someone has to be accountable, and accountability cannot be delegated to a model.
• Work in messy physical environments — plumbing, construction, farming.
• Improvisation — reading a room, handling a household crisis, the thousand small judgments a waiter makes in an evening.

The economist Lu Ming makes the structural version of this argument. Services have a marginal-cost structure that protects them: hardware wears out and must be depreciated, someone has to travel to the customer's home, and very low-frequency tasks never justify the engineering cost of automating them. Those cost floors are where a lot of human employment survives.

That is a more useful way to think about safety than "creative vs routine". The question is not whether a machine could in principle do the task. It is whether the economics of doing it with a machine ever close.

3. The growth paradox: where the gains actually go

Here the story stops being about technology and starts being about distribution.

The Chinese economy grew 4.6% year-on-year in the first three quarters of 2026, while real per-capita consumption grew 2.7%. Labour income was 43.6% of GDP in 2025, against a developed-country average of about 55%. More than half of surveyed companies say AI has improved their margins; roughly one in ten has raised pay at the same time.

A projection from CICC takes the trend to its limit: over the long run, compute's share of total output approaches one and labour's share approaches zero. Whether or not you accept the endpoint, the direction is hard to argue with.

The mechanism is simple, and it is why this round of automation feels different from earlier ones. Capital no longer needs labour as an intermediary in order to reproduce and improve itself. A robot does not draw a wage, does not pay social insurance, does not take leave, and does not negotiate. The cost saved flows to whoever owns the robot, and the worker who used to capture part of that cost through wages is simply no longer in the chain.

Productivity rising is not the same thing as prosperity spreading. Those were two separate questions all along; AI just makes the gap impossible to ignore.

4. The ceiling of the Earth market

Capitalism has always handled distribution by growing. New markets absorb the output, new jobs absorb the workers, and everyone can avoid the fight over shares. That escape route is narrowing.

• At home: overcapacity, strong supply and weak demand. Retail sales grew only 1.1% year-on-year in January–September 2026.
• Abroad: deglobalisation, rising trade barriers, and supply chains being deliberately shortened.
• The next markets: Africa and Southeast Asia are real, but their purchasing power is rising far more slowly than production capacity is.

When growth hits a ceiling, a society has to move from dividing a growing pie to dividing a fixed one by rules. That is a political problem, not a technical one, and no model release solves it.

5. Energy: the key that opens the rest

The chain I keep coming back to runs like this:

• Energy becomes effectively free → the marginal cost of production approaches zero.
• Marginal cost approaches zero → goods approach being free.
• Goods are nearly free → "equivalent exchange" stops describing anything real.
• Equivalent exchange fails → money loses the basis it has had for several thousand years.

Elon Musk has said something close to this: once AI and robots can manufacture chips, build solar panels, extract resources and update their own systems, the traditional monetary system loses its foundation. You can treat that as a prediction rather than a fact, but the logic inside it is not exotic — it is just the labour theory of value taken to its conclusion.

What does not become free is equally important. Location and land. Attention and time. Creativity. Power and status. Energy abundance does not dissolve those; it moves competition onto them. If material goods stop being the scarce thing, people will compete on relative position — and relative position is, by definition, a zero-sum game. That is the part of the "everything free" future that gets discussed least.

6. Four scenarios for the near future

The World Economic Forum framed four employment scenarios in early 2026. They are a reasonable map, even if the probabilities are guesses.

Scenario	What it looks like	Judgement
Rapid progress	AI develops quickly; people become "orchestrators of agents"; new professions appear at scale, but social protection cannot keep up.	Ideal, but hard
Substitution era	AI displaces workers faster than they can adapt; unemployment spikes; society fragments.	Largest risk
Copilot economy	AI develops gradually; human–machine collaboration becomes normal; most people learn to drive the tools.	Most likely
Stagnation	AI progress slows; legacy processes harden; the gap between those who use the technology and those who do not widens inequality.	Most pessimistic

The copilot economy is the one to plan for, precisely because it is the least dramatic. It does not require a breakthrough or a collapse. It requires a few hundred million people to learn a new interface, unevenly, over about a decade — which is roughly what happened with smartphones, except that this time the interface can do part of the job.

7. What this means for an ordinary person

The most exposed position is not "low-skilled". It is fixed: a fixed salary, repetitive duties, no customers of your own, and no judgement of your own. That combination is exactly what a model plus a workflow can absorb.

The safer positions are three:

• People who can operate AI. Not people who can chat with it — people who can point it at a real process and get a result. One person with that skill covers the output of a small team.
• People with their own business. Your own customers, your own judgment, your own lane. A business is a bundle of relationships and decisions that does not fit inside a prompt.
• People doing what AI cannot. Emotional care, complex decisions under responsibility, genuinely new creation, and work that happens with your hands in an unpredictable physical world.

The sentence I would put on a wall: the thing that replaces you is not AI. It is a competitor who has learned to use AI and is now twice as fast as you at the same job.

8. The endgame

Put the pieces in order and the chain is complete:

AI replaces a large share of labour → productivity rises sharply → the gains concentrate in the owners of capital → growth runs into the ceiling of a saturated planet → energy becomes the binding constraint → solve energy and marginal cost collapses toward zero → the distribution problem is exposed with nowhere left to hide → the social form has to change → and the direction of that change is "from each according to ability, to each according to need".

That is not a slogan I invented. It is a route map drawn more than a century ago, and AI plus cheap energy are, step by step, making it practical. The transition will be painful, because it redistributes something that people currently hold through wages, titles and ownership. But the direction is no longer ambiguous.

Which brings it back to the factory. I am not waiting for the endgame. I am using the tools that exist now, keeping the business that pays today, and watching which of the four scenarios actually arrives. That is the only honest position available to someone who has to make payroll this month.