Seven Futures for the AI Economy

August 13, 2026

What follows is a scenario framework, built by granting the strongest version of the AI-abundance claim and then asking what else would have to be true. The short answer: intelligence is one of four locks on the door, and it is the only one the technologists hold a key to.

The Promise and the Catch

The boldest version of the AI promise runs like this: machine intelligence exceeds all of humanity’s combined by the early 2030s; humanoid robots follow; output becomes “quasi-infinite”; work turns optional; and by 2036, money itself stops mattering. Elon Musk says this outright, and millions quietly believe some version of it.

The catch is a gap in the evidence. Everything cited in support — benchmark scores, coding demos, robot prototypes — shows what AI can do in a test. The conclusions are about a world where no human is involved at all. Between those two points lie five stages. A task must first be exposed — something AI could touch at all. Then technically feasible, performed under benchmark conditions. Then economically viable, meaning cheaper once you count integration, oversight, errors, and lawsuits. Then adopted, actually run in production by real organizations. Then, finally, autonomous: no human signature left on it.

Call it the funnel. The evidence lives at stages one and two. The conclusions live at stage five. The economy — prices, wages, jobs — lives in the middle.

Chess is the tell. AI enthusiasts love saying machines are at “Stockfish level,” and in chess that meant instant, total human obsolescence. But chess is the one arena where the funnel has zero length by construction: no liability, no integration cost, no regulator, instant verification, total autonomy the moment feasibility arrives. The claim is exactly true in chess-shaped domains — and the real economy is made of sign-offs, permits, insurance policies, and building codes.

Four Kinds of Work, Four Speeds

Sort every economic task by the last thing blocking its automation, and the economy splits into four parts, each moving at its own speed:

Digital work — code, analysis, documents, support. Blocked only by AI capability. Moves at the speed of model releases: months.
Physical work — construction, logistics, hands-on care. Blocked by robots, power, and factories that must be physically built. Moves at buildout speed: years.
Permission work — medicine, law, auditing, engineering sign-off. Often technically automatable today, but the product being sold is a licensed human’s legally required signature. Moves at the speed of law: decades.
Scarcity goods — located land, status, attention, access. Valuable because they’re rare. Cannot be automated, ever, by definition.

These are constraints, not industries, which is why the split cuts across familiar sector lines. Two concrete examples show why that beats thinking in industries:

A house is a bundle: the structure (physical work), the land (pure scarcity), and the permission to build (a legal artifact). In expensive cities, the structure is only a third of the price. Even perfect robot builders would leave two-thirds of the cost untouched.

Healthcare is all four at once: diagnosis and paperwork (digital), nursing and procedures (physical), prescribing authority (permission), access to elite specialists (scarcity). So “AI will make healthcare cheap” is ill-posed — diagnosis could get 90% cheaper while your total bill rises, because the expensive parts aren’t the digital parts.

Three Rules That Follow

1. Cheap things get cheaper; dear things get dearer. When one part of the economy becomes hyper-productive, it shrinks as a share of everyone’s budget, and spending shifts to the parts that didn’t improve. Free intelligence therefore makes housing, healthcare, education, and credentials relatively more expensive — and those are precisely the unavoidable purchases. Somewhere between a twentieth and a tenth of the entire economy is currently work that machines could technically do, priced at licensed-professional rates — the low end of that range better evidenced than the high. That slice is what the coming political fight is about.

2. Insurance moves before law. The permission barrier doesn’t fall when a legislature votes. It falls the day an insurer prices coverage for AI mistakes below the malpractice premium for a human professional — at that moment the economics of the human signature collapse, and statutes catch up later. Whoever wants early warning should watch insurance pricing, not bills in committee.

3. Money concentrates rather than disappears. As everything makeable gets cheap, money doesn’t lose meaning — it becomes purely a claim on the two things that can’t be manufactured: place and permission. In every future, including full-blown abundance, wealth pools in land, position, and licenses. “Money won’t matter” gets it exactly backwards.

Seven Futures

The odds below are rough, honest judgments — tools for disciplined thinking, not precision forecasts. The structure matters more than the digits. Each future is a decade signature: what the 2030s look like from inside, not a description of some final state.

Future	Odds	In one line
The Grind	28%	Abundance in the app store, inflation in real life
The Rentier Decade	19%	Superintelligence behind locked gates
The Robot Decade	13%	A compressed industrial revolution
Full Abundance	9%	Every gate opens, nothing throttled
The Plateau	8%	Progress stalls; the spending boom busts
The Credit Bust	8%	The AI works; the debt financing it doesn’t
The Split World	8%	Different countries land in different futures
None of the above	7%	Decades that refuse to be named

The last row is not an eighth future but an honesty budget: mostly decades that straddle two signatures — a gate that opens in 2035 looks like one world in the statistics and another in the streets. And two things changed from this essay’s first printing, both from running the framework’s own update machinery rather than from editorial second thoughts: the Rentier Decade overtook the Robot Decade for second place, and the Split World was promoted from the leftovers to a branch with its own mechanics. The reasons are in the full document’s changelog; the short version is at the end of this page.

The Grind — 28%, still the single most likely decade. AI keeps improving; robots stay scarce; the permission barrier holds. Growth runs a strong 3–5% early, then slows by design in the mid-2030s — not because AI failed, but because the cheap digital slice keeps shrinking as a share of spending. (Expect that slowdown to be misread as “the AI story is dead.”) There is no unemployment spike, and that’s the strangest part: displaced office workers are absorbed into protected sectors — compliance, administration, coordination, care — real jobs, but slower, lower-paid, gate-kept ones. The inefficiency of the protected sectors quietly functions as the welfare state, funded through the prices everyone pays rather than through taxes. Daily life: unlimited free intelligence in every pocket; rent, premiums, and tuition eating the paycheck. The clearest losers are young, credentialed renters — and the telltale statistic is professional-school applications per seat surging while incumbent professionals’ incomes also rise.

The Rentier Decade — 19%, now the second most likely. AI becomes self-improving — and deployment stays bottlenecked anyway, because robots can’t be built fast enough, permission is withheld, or governments deliberately throttle. Astonishing capability overhead; wealth pooling to whoever owns energy, compute, land, and licenses; wages sagging without a mass-unemployment moment to organize around. The uncomfortable kicker: governments’ default posture — subsidize AI for national security, restrict deployment to protect professions and ratepayers, too fiscally strapped to redistribute — is a machine for producing exactly this world. And the machine has a second gear the first printing missed: the state is also AI’s most eager customer. Permission has a direction. Governments move slowly when AI would replace doctors and auditors, and quickly when it strengthens their own hand — continuous tax compliance, claim scoring, fraud flagging, predictive caseloads — because there the approver and the buyer are the same office. So the informational state can become superhuman long before the physical economy does: cheap cognition for everyone, extraordinary cognition for institutions, formal rights intact while the operative decisions migrate into models and thresholds nobody votes on. Call it computational rentierism — arrived at incrementally, legitimately, one efficiency improvement at a time. Capability optimism mostly raises these odds, not utopia’s.

The Robot Decade — 13%. Machines reach mass production before ~2035 and automation goes physical. Prices fall broadly, growth hits high single digits — and two dangers arrive with it. Workers’ slice of national income collapses mid-transition, before any new social contract exists. And broadly falling prices make existing debts heavier in real terms: governments owe fixed sums against a shrinking price level, with public debt already enormous — the 1930s configuration. This is the one future where “deflation will be the issue” is genuinely right, and it’s a warning, not a comfort. One variant reaches this decade without the robots: if the permission barrier collapses at scale instead — insurers repricing the human signature out of existence — the growth surge and the labor-share crunch arrive through falling professional prices rather than machines. Either way, the transition outruns the social contract.

The Plateau — 8%. Progress stalls near current levels. The trillion-dollar datacenter boom unwinds around 2027–29 as an ordinary investment bust. But already-built AI keeps diffusing for years afterward — so pressure on entry-level office jobs continues through the bust.

The Credit Bust — 8%. The AI is fine; the financing isn’t. The buildout is increasingly funded with debt secured against chips whose working life is itself disputed: the operators depreciate them over five to six years, while sceptics put the economic life closer to two or three. The gap between those two numbers is the whole question. Around 2028–29 comes the test: does AI revenue cover debt payments plus perpetual hardware replacement? If not, the losses land on the lightly-regulated funds and insurers holding the paper. Crucially, this is no reprieve for workers — cost-cutting accelerates automation, and nobody uninstalls systems that work. With tech hiring and the construction boom both frozen, this is the worst decade of all to be entering the workforce.

The Split World — 8%. No single global outcome. China can grant AI legal permission by government decree — no insurance market required — and may open its gates first, faster but with nobody independently pricing the risk of correlated failure. Europe keeps its gates shut without building an AI frontier of its own: importing others’ abundance, exporting its savings to fund their buildout. India faces the harshest cell: its youth bulge peaks exactly as the outsourcing ladder melts, with almost no protected sector to absorb the shock. The thing to watch is whether the first decree-driven opening stays local: a permission barrier falling in one bloc while holding in the others is this future arriving.

Full Abundance — 9%. Every gate opens, nothing is throttled. Ordinary economic reasoning expires here; the only decisions that still matter are the ones made in advance — how widely ownership was spread, who governs compute, what guardrails were locked in.

The punchline. Grant the optimists total certainty about machine intelligence, and full abundance still only comes up about one time in four. The other three-quarters flow overwhelmingly to the rentier decade, with slivers to the robot decade and to a world split along its borders — because the remaining barriers are robots, permission, and politics, and no amount of intelligence flips those on its own. If the capability prediction is fully right, the most probable result is the world least like the one predicted.

What Happens in Nearly Every Future

Whatever the branch, expect:

Entry-level white-collar disruption — in booms and busts alike.
Two price worlds at once: digital collapse alongside rising housing, healthcare, and credential costs — with headline inflation looking deceptively calm, because averaging a plunge and a surge produces a placid number. Young renters and older homeowners will live in visibly different economies, and both will be reading their own reality correctly.
Land and position claiming a growing share of national income.
A trades boom — electricians, builders — that is real but self-liquidating: the buildout bids up wages, then trains its own competition, and in the Robot Decade the trades build the very machines that replace them.
The energy buildout continuing regardless.

Six Traps Under the Surface

The training-rung trap. Junior white-collar work was always two products in one: cheap output plus an apprenticeship that manufactured future senior professionals. AI destroys the output’s value; no one ever separately paid for the apprenticeship; so the senior class of 2035 quietly doesn’t get made. The early evidence fits: firms are adjusting by not hiring the young, not by firing the experienced. The cheapest fix in this entire landscape is explicitly funded apprenticeships — completing a broken market, not writing welfare checks.
The one-lever trap. Central banks hold a single interest rate against four diverging price trends. Raising rates cannot lower prices set by legal permission or land scarcity — but it can strangle homebuilding and investment, worsening the very scarcities voters are angry about. Whichever number the central bank targets, it amplifies one distortion or another.
The scissors. Government revenue leans on wages precisely as wages erode, while needs peak — displaced workers, transition costs, interest bills already exceeding defense budgets. This is why restriction beats redistribution by default: restriction is off-budget. Licensing rules become redistribution laundered through consumer prices.
Finance loses in both directions. If AI disappoints, the datacenter debt implodes. If AI triumphs, offices, legacy software firms, and older chips get stranded. The financial system is only safe in the orderly middle — and the two tails are held by different institutions, each blind to the other, neither hedged.
The safest asset is the future tax target. Since value flows toward land and position in every future, that is exactly where taxation must eventually aim. The endgame political fight has a pre-printed address, and the sophisticated money and the future taxman are converging on the same coordinates.
The one-way ratchet. Nearly every feedback loop in this story — professional lobbies, ratepayer anger, insurer caution, budget stress — pushes toward less deployment, never less capability. Pressure builds behind the gates rather than dissipating. Two forces run the other way. One comes from abroad: when a country deploys by decree, the failure data it generates lets insurers everywhere finally price the risk — permission barriers can be eroded from outside a country’s own politics. The other comes from inside the gatekeepers themselves: institutions adopt AI eagerly when it strengthens their hand — audits, eligibility checks, fraud scoring — even as they throttle the versions that would replace them. The ratchet, it turns out, throttles deployment for you, not on you.

Politics Arrives in Two Waves

Wave one (2026–28): power bills. Households pay for the datacenter buildout in their electricity rates before any visible benefit arrives. Expect rate revolts, datacenter tariffs, and siting moratoria.

Wave two (2028–32): the displaced young. Educated, urban, renting, queue-stuck — and demanding access before handouts: entry to licensed professions and housing they can afford. Each profession then splits internally: established insiders want doors closed; the queue outside wants seats opened. The early test cases are already legible — accounting (severe shortage, weakening guild) should open its doors; law (oversupplied at the top, strong guild) should slam them shut. If the opposite happens, this whole picture needs rethinking.

A second current rides the same wave: backlash against the government’s own algorithms. It has already toppled a government — the Netherlands’ childcare-benefits scandal — had a welfare-fraud scoring system struck down in court, and ended Australia’s automated debt scheme in a royal commission. As agencies automate themselves, episodic scandal is how the public fights its half of the permission battle: the citizen-facing gate closes loudly, after the fact, through courts and resignations rather than statutes.

How to Tell Which Future Is Arriving

The price of AI liability insurance versus human malpractice premiums — the earliest and best signal of the permission barrier cracking.
Hiring versus firing rates — the signature isn’t layoffs but a frozen market: nobody fired, nobody hired.
Employment of 22–25-year-olds in exposed occupations versus their older colleagues.
Robots actually shipped — production numbers, never demos — and robot cost-per-hour versus roughly twice the median wage.
AI revenue versus debt-plus-replacement bills, coming to a head around 2028–29.
Land’s share of national income and of house prices — the scarcity thermometer.
The gap between wholesale AI costs and consumer prices — measuring whether savings are passed through or pocketed.
Government AI procurement and algorithm registries — whether the state is automating itself faster than it permits the economy to be automated.

And three signs this whole picture is wrong: spending on digital work rising for years even as its price collapses (bottlenecks don’t rule after all); licensed services getting cheaper with no change in law or insurance (the permission barrier was imaginary); or sweeping redistribution passing before 2030 (politics is faster than assumed — the one welcome way to be wrong).

The Scorecard on the Original Prediction

Right: electricity and chips are the binding physical constraints, and China’s power advantage is real.
Right in shape, wrong in kind: prices will fall — but relative prices, not everything at once; the dangerous everything-falls version appears only in the Robot Decade, where it’s a debt crisis, not a bonus.
Wrong: “no way to compete” is true only in chess-shaped domains with no liability and no gatekeepers — which is to say, almost nowhere that pays a salary.
Inverted: money doesn’t stop mattering by 2036. It ends up mattering more, and more unequally, concentrated on the two things no machine can manufacture: place and permission.

The general pattern behind these errors: predictions about hardware run merely late; predictions about institutions run wrong. The error grows with the number of regulators, insurers, and licensing boards standing between a demo and a deployment — and the future of the economy is, more than anything else, the story of that crowded middle.

Where this comes from. The above is a distillation. The full working document — Musk’s Economic Model: Reconstruction, Critique, and a Scenario Framework — is roughly fifteen times longer, and carries the parts that had to be cut: the growth accounting behind the four classes, the scenario tree the seven futures are composed from and its sensitivity analysis, the debt arithmetic under the credit bust, the jurisdictional breakdowns, the dashboard of indicators, and the falsifiers.

Its caveats apply here too. Its probabilities are stated subjective priors rather than measurements, and it keeps an explicit register of what it gets wrong. Its source-verification checklist has since been executed against five independent research passes; the corrections that touched this page have been applied, and the full ledger of what survived, what was stale, and what was simply wrong is in the appendices. Anyone inclined to lean on a number above should go read how it was built.

Since this essay’s first printing, the framework has also run its own update machinery once (v3, August 2026), and the odds above were recomputed mechanically rather than re-judged. Three things moved. The capability evidence came in faster than the draft assumed, which raised the self-improvement branch and with it the Rentier Decade. Cross-country divergence was promoted from the unnamed leftovers to a branch with its own conditions, which is why a “none of the above” row now appears. And Full Abundance rose from 6% to 9% — bookkeeping, not optimism: the update raised the odds of self-improving AI without also strengthening the political-throttle assumption, a deliberate restraint against counting one intuition twice; the framework expects that number to come back down if the evidence for institutional self-automation firms up. The changelog, and the governance mechanism this revision added, are in the full document.