The Machine Eats the Ladder

Financialized AI and the End of Apprenticeship

The machine arrives with a bill before it arrives with a dividend. Not a metaphorical bill. A real one: power, memory, chips, land, water, cooling, engineers, permits, debt, security customers, grid priority, political fear, and a place in the queue ahead of someone else. AI is sold as intelligence in the cloud. It is being built as demand on the ground.

The first fight is over electricity. The next is over capital. The next is over which town becomes a datacentre annex. The International Energy Agency projects that global data-centre electricity consumption will double to about 945 TWh by 2030, still less than 3 per cent of global electricity demand. Fine. Global averages are how adults hide a local fire inside a spreadsheet. The real issue is speed and location: data-centre electricity demand is growing far faster than other electricity demand, and these facilities cluster in particular grids, water systems, planning regimes and political bargains. (IEA)

In the PJM region, the largest US power grid, the abstraction has already become an administrative problem. Data-centre demand is now part of the supply-demand balance for a grid serving 65 million people. Capacity prices have risen by more than 1,000 per cent since around 2024. The policy menu is boring only if you refuse to translate it: make data centres fund new supply, procure backstop power, or curtail them during stress periods. The translation is simple. The machine has entered the rationing queue. (Reuters)

Power is only the first invoice. Goldman Sachs estimates roughly $7.6 trillion of AI-related capital expenditure from 2026 to 2031 across compute, datacentres and power. The Bank for International Settlements says the five largest hyperscalers are set to spend more than $1 trillion on AI-related capex from 2025 through 2026, with commitments outrunning earnings and free cash flow for some firms. The naïve reading is: bubble. The better reading is: yes, and? A bubble is not always the opposite of infrastructure. Sometimes the bubble is how infrastructure gets built. (Goldman Sachs)

Railway manias left railways. Telecom manias left fibre. Dot-com wreckage left habits, talent, protocols, cloud precursors and platform ambition. The investor can be ruined while the empire keeps the asset. That is the American genius, and it is obscene. America can put a price on a future that has not happened, borrow against it, securitise it, wrap it in national security, sell it to institutions, use it to hire engineers, sign power deals, build campuses, and dare reality to catch up. Then the sceptic arrives and asks: where are the profits? Too late. The datacentre is already there.

This is not froth around the AI race. It is one of the race engines. America’s advantage is not realism. It is the ability to make fiction operational. That may be why America wins. Not Americans. America-the-stack: hyperscalers, frontier labs, Nvidia, private credit, security customers, founders, listed markets, infrastructure funds, and the state that calls all of this strategic. America-the-stack can win while Americans-the-people receive the invoice. A country can win a technology race the way private equity wins a hospital: by extracting the balance sheet, degrading the institution, and leaving with the asset.

That sentence sounds too violent until you ask what the AI buildout is already doing. It asks the grid to accommodate it. It asks capital markets to believe it. It asks the state to protect it. It asks firms to reorganise around it. Then it asks a quieter thing, more intimate than power or chips. It asks for the ladder.

The machine does not only eat electricity. It eats the tasks by which people learn to become useful: the first draft, the document review, the spreadsheet clean-up, the translation, the customer email, the junior policy memo, the litigation search, the bug fix, the ugly deck, the meeting note, the supervised mistake. These were not holy tasks. Most were boring. Some were stupid. Some were hazing with formatting. No one should romanticise clerical suffering. But they did one thing societies are about to miss. They let beginners become less useless under supervision.

The junior job was not waste. It was the nursery of judgment. The AI story says: remove the drudgery. Good. But the machine does not know the difference between drudgery and rehearsal. The senior person sees a bad first draft and asks why the firm is paying for this. The model can do it. The note can be summarised. The source sweep can be automated. The customer response can be drafted. The legal pass can be done by software. The first deck can be generated. The junior can be skipped. Correct. Now where does the junior learn?

Where does the analyst learn that sounding clever and being useful are different things? Where does the lawyer learn that cases have smells? Where does the programmer learn consequence? Where does the civil servant learn that the sentence buried in paragraph seven is the sentence the minister will die on? The labour debate asks whether AI creates jobs. Wrong question. The question is whether the new jobs form people. A task pays you. A ladder changes you.

AI will create tasks: model evaluation, data annotation, agent supervision, workflow design, red-teaming, compliance, AI operations, synthetic data, domain implementation. Fine. Some of these will become real professions. Some will be better than the work they replace. But a task is not a ladder. Data labelling can be work without being formation. Model evaluation can be work without being apprenticeship. Agent supervision can teach judgment, or it can make a human the liability wrapper around a machine they do not control. The test is brutal. Does the work carry a person upward? Does it teach judgment? Does it create trust? Does it attach a name to consequence? Does it lead somewhere?

Platform capitalism already showed the trap: work without careers. AI can repeat the trick at a higher cognitive layer. The early data is not proof. It is worse than proof. It is plausible. PwC’s 2026 AI Jobs Barometer says AI-exposed “seniorised” entry-level roles have grown 35 per cent since 2019, while other entry-level roles declined by 10 per cent. PwC’s own framing is managerial and optimistic: organisations must rethink how they mentor and train juniors so they step into complex decision-making earlier. Exactly. The ladder is being compressed. (PwC)

Switzerland gives the same smell. A jobs.ch study covering more than 7.3 million job advertisements found entry-level postings in 2025 were 32 per cent below the 2019–2022 average; senior roles in AI-affected fields rose by 26 per cent, while junior roles in those sectors fell by 16 per cent. Do not overclaim. Hiring cycles matter. Interest rates matter. Sector weakness matters. Post-pandemic distortion matters. Still. The pattern has a name: AI seniorises the junior role. It asks beginners to arrive already formed. (Reuters)

This is financialized AI’s revealed preference. It does not want the beginner. It wants the already-trained human with agents attached: senior engineer plus agents, senior lawyer plus agents, senior analyst plus agents, senior operator plus agents, senior founder plus agents. Terrifying productivity. The senior has taste, context, mandate, memory, scars, clients, enemies, timing, caution. The machine multiplies that. Wonderful. The beginner has needs. The beginner needs time. The beginner needs correction. The beginner asks basic questions. The beginner slows the meeting. The beginner writes badly before writing well. The beginner must be exposed to consequence in small doses. The beginner requires someone expensive to pay attention. The beginner is bad margin.

So the machine economy does what financialized systems do. It buys the finished asset and stops funding the nursery. That is the centre of the argument. Not “AI replaces jobs.” Too crude. Not “AI raises productivity.” Too innocent. Financialized AI wants judgment without the institutional cost of producing judgment. It wants the fruit without the orchard.

China has its own ghosts. This is not just a Western professional-class story. Reuters reports that Chinese firms are using AI to quietly reduce graduate hiring and shrink functions in tech, entertainment and advertising, even while Beijing pushes AI adoption through its “AI Plus” agenda. Reuters also notes analyst warnings that AI-driven job creation is lagging displacement, with early-career workers especially exposed. (Reuters)

China also has a youth labour problem large enough to make every official sentence sound brittle. The 16–24 unemployment rate, excluding students, was 15.6 per cent in May 2026. That is the revised, less alarming statistical series, after students were removed from the denominator. So call them what they are: highly educated ghosts. Not unemployed in the old factory-gate sense. Not peasants waiting outside a mill. Not surplus hands for assembly lines. Credentialed, exam-trained, family-funded, platform-fluent, politically nervous young people standing before ladders that are narrowing before they can step onto them. (Reuters)

The Chinese ladder is not being eaten by American venture capital. It is being eaten by automation, industrial policy, graduate overproduction, platform pressure, local fiscal stress, and a state that wants AI to be disruption and cure at the same time. Different hand. Same bite.

The one-person unicorn is therefore not a miracle. It is a threat model. One founder. Ten agents. No juniors. No middle. No apprenticeship. No institution except the founder’s taste and the machine logs. Research, code, sales, support, investor updates, design, recruiting, legal triage, customer service: all pushed through agents and contractors. Output rises. Headcount stays beautiful. The cap table smiles. But what has been built? Not an institution. A private amplifier. The founder compounds. The agent memory compounds. The platform compounds. Everyone else gets gigs.

This is the class form hiding inside the productivity fantasy. The AI-native firm can become a machine for scaling judgment without reproducing the conditions that formed judgment. The phrase “AI-native organisation” should not be treated as a compliment. It should be treated as an indictment until proven otherwise. Can it train people? Can it pass judgment downward? Can it turn error into trust? Can it promote anyone who was not already formed before the machine arrived? If not, it is not a new organisation. It is a private learning monopoly.

Do not freeze the old hierarchy. Most old hierarchies deserved to be burned. Much junior work was not apprenticeship. It was sadism with stationery. But the answer is not to remove the junior. The answer is to make the junior fight the machine. The junior does not write the first draft in the old way. The model can do that. The junior judges the drafts: which source did it miss, which assumption is lazy, which answer sounds right because the prose is smooth, which answer is correct but unusable, which answer would fail in the room, which answer would embarrass the client, which answer has no owner. That is work. Better: that is formation.

The agent produces surface area. The human learns discrimination. The supervisor’s job changes too. The supervisor does not merely mark up the memo. The supervisor marks the judgment. What did the agent propose? What did the junior accept? What did the junior miss? When did the junior override the machine? Was the override right? Which mistakes keep returning? Which forms of judgment are improving? That is a ladder. Not the old ladder. A real one.

The questions are simple. Who owns the memory? Who learns from the error? Who gets trusted next time? Who is promoted? Who is merely kept in the loop so the machine has someone to blame?

This is Agentworld made practical. Benjamin Bratton’s frame names a world where human and non-human agents act, remember, negotiate and communicate alongside one another. The brief calls this a “preemptive anthropology” of open-world centaur societies, with agent institutions, agent-to-agent interaction, open-world ecologies and hybrid forms of agency. Fine. Here is the darker version: agents will not wait for law, unions, management theory, labour-market redesign, public consensus, universities to decide what to teach, or firms to decide whether juniors still matter. They will enter first. They will take routine work first. They will acquire memory first. Then institutions will retrofit the settlement after the damage has already begun. (*)

This is why the cheerful Agentworld story is incomplete. It asks what kind of society emerges when humans and agents interact. Wrong question. The first question is who gets displaced from the training ground while that society is being assembled. Are humans being trained above the machine, displaced below it, or trapped beside it as supervisors of outputs they cannot judge?

Every organisation has a learning ledger. In the old firm, it was crude but visible. The junior wrote a bad memo. The senior marked it up. The next memo was less bad. The worker learned. The supervisor learned what the worker could handle. The institution converted error into trust. AI breaks the ledger. The task still gets done. The email is answered. The document is summarised. The bug is patched. The deck is improved. But who learned? The junior, the supervisor, the firm, the agent memory, the platform, the lab?

This is the class question inside AI-native work. Not who touched the task. Who booked the experience? If the learning stays with the worker, AI can rebuild the ladder. If it stays with the firm, AI can strengthen institutions. If it flows to the platform or the lab, the economy becomes their training environment. Then the ladder is eaten twice. First the machine takes the junior task. Then it takes the learning from the task.

The China question is not chips. The hawks were fixated on the wrong constraint. The question is not whether China can get Nvidia-equivalent chips. The question is whether China can deploy enough compute, memory, interconnect, power, cooling, software workarounds and datacentre capacity fast enough to matter. Export controls may slow China. They may also train China. A blocked system does not simply stop. It substitutes. It duplicates. It wastes. It learns.

Meituan says LongCat-2.0 was trained and run entirely on a 50,000-chip domestic Chinese cluster. Reuters notes the company did not name the chipmaker and that most Chinese models have historically been trained on US chips. Good. Be cautious. Do not declare parity. But do not miss the point either. This is not a press release about victory. It is evidence of adaptation under restriction. (Reuters)

CXMT points to the same stack logic. Reuters reports that the Chinese memory maker signed a long-term DRAM supply agreement with Tencent worth more than 20 billion yuan ahead of a planned 29.5 billion yuan STAR Market listing. That is not pure Wall Street hallucination. It is not old state-bank lending either. It is offtake plus domestic substitution plus capital-market story. The STAR Market reforms matter for the same reason. China is opening the listing path for strategically important but unprofitable AI large-model companies. Translation: Beijing is trying to make the right kind of speculation politically tolerable. (Reuters)

Huawei’s roadmap makes the systems question explicit. Huawei says the Atlas 960 SuperPoD will assemble more than 15,000 Ascend chips into one deployable system by Q4 2027. Treat the numbers as company claims. The direction is still clear. China is not just asking: can we make the blocked chip? It is asking: can we turn domestic chips, memory, interconnect, software and power into usable compute at scale? That is the actual contest. Not chip. System. Not import. Deployment. Not purity. Survival. (Huawei)

The usual story says America has finance and China has the real economy. It is usually meant as a criticism of America and a compliment to China. In AI, the compliment may be wrong. America’s financial system is not froth around the machine. It is part of the machine. It turns a claim about future dominance into money today. It lets a founder sell inevitability, a market price the fantasy, a fund buy the debt, a utility sign the power agreement, a governor bless the campus, a security agency become the anchor customer. This may be grotesque. It may also work.

China’s danger is not that it cannot build. That would be stupid. China can build. It can build fabs, chips, batteries, ports, datacentres, power lines, robots, platforms, cities, supply chains and industrial parks. The danger is subtler. China may build too literally. The Self-Strengthening mistake was not that China failed to buy enough machines. It was mistaking the machine for the system that made the machine world-historical. Ships mattered. Guns mattered. Arsenals mattered. But so did finance, law, universities, insurance, accounting, corporate form, logistics, capital markets, and the permission to speculate.

The new version is AI with Chinese characteristics: build the chips, build the fabs, build the datacentres, build the models, build the power stations, but underbuild the financial machinery that turns possibility into command. China does not need “more finance” in the abstract. It has plenty of bad finance: property finance, local-government debt, collateral worship, administrative safety, political credit. The problem is not too little finance. The problem is that the wrong finance has been allowed to live too long, and the right finance remains politically dangerous.

Chartbook 455: The AI boom, China's walled-in-wealth and the financial barriers that separate the two. - World Economy July 2026.

China does not lack wealth. It has walled-in wealth. Tens of trillions in yuan deposits sit behind capital controls, watching the US AI boom inflate the assets they are not allowed to buy. Beijing cannot simply “financialize” in the American way, because opening the gate may not fund Chinese AI. It may send Chinese wealth rushing toward the American tech party, dragging the RMB and domestic assets with it. So China must attempt something harder: a controlled AI bubble, a patriotic casino, a STAR Market future in which trapped wealth is offered domestic chips, memory, models and industrial platforms instead of Nvidia, SpaceX and Wall Street. America turns global belief into infrastructure. China must turn captive savings into conviction without letting capital become politically sovereign.

America tries to turn belief into infrastructure. China tries to turn restriction into learning. The American sin is hallucination. The Chinese sin may be literalness. One moves too fast because money believes too easily. The other may move too slowly because the state wants finance to behave.

The outcome depends on what AI becomes.

If AI remains a scarce frontier capability — best model wins, capability gaps persist, access is tiered, agents become proprietary infrastructure, security customers anchor demand — America’s financial engine has the advantage. It was built to monetise scarcity before scarcity has fully arrived.

If AI becomes an industrial utility — cheap, diffused, embedded in manufacturing, logistics, robotics, services, schools, clinics, procurement and local government — China’s production density becomes terrifying.

So the question is not “who is winning?” Wrong question. The question is: what kind of thing is AI becoming first? Frontier rent machine or industrial utility? The American system is built for the first. The Chinese system may be better at the second.

Both can overbuild. Both can leave useful infrastructure. Both can damage the human ladder while doing it. American financialized AI eats the ladder by buying seniors, scaling founders, automating juniors and privatising the learning ledger. Chinese industrial AI eats the ladder by automating factories, services and office work while producing graduates faster than institutions can absorb them, then selling AI adoption as both disruption and cure. One hand is public-market delirium. The other is state-directed overbuild. Both tell society the same thing: feed the machine first, settlement later.

Southeast Asia should watch this with hunger and fear. Middle regions can catch what empires discard: lawful second-tier compute, open models, unwanted talent, regional service work, AI-native public goods, small experiments too boring for frontier myth but useful to everyone else. But they can also sell water, land, power and young people too cheaply. Hosting the machine is not the same as owning the capability. The scarce asset may not be the largest model. It may be the ability to combine decent models, local context, institutional trust and human formation.

The uncomfortable conclusion is simple. AI is being sold as abundance. It is being built as priority access: priority access to power, capital, chips, policy attention, security customers, and the tasks that once trained people. The machine comes first. The settlement comes later. This is not an argument against AI. Spare me. The machine will be built. It will be useful. It will do real things. It will create new work. It will improve some institutions and ruin others. It will make some people vastly more capable. That is not the question.

The question is who pays for the buildout, who waits in the grid queue, who gets the rents, who learns from the work, who owns the memory, who gets trusted next time, and who never gets the first chance. If AI firms rebuild entry pathways, the ladder thesis weakens. If new AI jobs become real professions, it weakens. If productivity appears broadly outside the frontier stack, it weakens. If AI tools distribute capability downward, it weakens. But if AI capex absorbs power and capital faster than broad productivity appears; if grid costs are socialised while compute rents remain private; if junior hiring keeps compressing; if entry-level jobs require experience that beginners cannot acquire; if AI work remains gig work with nicer nouns; if AI-native firms scale output without training people; if states treat AI as strategic infrastructure while neglecting the social infrastructure that makes citizens capable — then stop calling this productivity.

Call it extraction.

America may win the AI race the way private equity wins a hospital. China may build the machine and discover, again, that the machine was not the system. Everyone else will be told to adapt. But adaptation is not magic. Someone has to build the next ladder. Someone has to decide where learning accumulates. Someone has to say that productivity without formation is extraction. Until then, the machine eats first, and the young are asked why they never learned to climb.