In Brief: The Ninety-Eight Percent Machine

Capitalism's winners keep about two cents of every dollar of value their innovations create. The other ninety-eight escape to society — as lower prices, better products, and goods that did not exist before. This report audits that payout across 106 primary documents, then asks what happens when agentic AI upgrades the machine that makes it.

In 2004, William Nordhaus went through the accounts of the American non-farm economy from 1948 to 2001 and asked a simple question: of all the value created by innovation, how much did the innovators keep? The answer was 2.2 percent — a central estimate in a range of 0.8 to 3.3 percent, with Schumpeterian profits amounting to just 3.8 percent of corporate profits. Everything else flowed to people who never signed a contract with the innovator.

That number inverts the standard story. The standard story says markets extract value from society and concentrate it at the top. The audited story says the market is a machine that forces its winners to give away roughly 45 times their profit in surplus (ENSI estimate: 1 ÷ 0.022). The mechanism is not benevolence but competition, imitation, and patent expiry: rivals copy, prices fall, and the surplus escapes to consumers.

The machine's total output is not small. The world produces roughly $150–180 trillion a year, in purchasing-power terms, above what a pre-industrial world of eight billion people would produce (ENSI estimate from the Maddison Project series) — frontier income per head is up about 30-fold since 1800. The natural experiments answer the attribution question: Korea split in 1948 and the market half now out-produces the planned half by an order of magnitude; the freest quartile of economies produces 7.6 times the GDP per capita of the least free ($52,877 vs $6,968, Fraser 2024). Assign even half the modern flow to the institutional package and ~$75–90 trillion a year exists because the rules allow prices, property, and entry.

And the receipts reached the bottom. Extreme poverty fell from 38 percent of humanity in 1990 to 8.4 percent in 2019 — more than a billion exits while population grew by 2.3 billion. China's market reforms account for roughly 800 million of them.

The best parts are counted at zero

The audit has a measurement problem that biases every number downward. GDP records transactions; the surplus escapes uncounted. Choice experiments put the median user's price for giving up search engines at $17,530 a year, email at $8,414, digital maps at $3,648 — services GDP values at roughly their advertising revenue (Brynjolfsson-Collis-Eggers, PNAS 2019). The uncounted US digital surplus plausibly runs $5–8 trillion a year (ENSI estimate; willingness-to-accept runs above willingness-to-pay, and even halved the figure stays in the trillions). Fold in mortality, leisure, and inequality, and welfare-relevant growth has run roughly 1.5 times measured consumption growth in catching-up economies (Jones-Klenow). Every headline number in this audit is therefore a floor.

The dominant term — and the honest debits

Rank the machine's seven mechanisms by value created and one towers. A dollar of R&D returns $13.30 of social benefit at baseline parameters — $5 under deliberately conservative assumptions, $20+ counting health spillovers — a social rate of return near 67 percent against private returns of 10–15 percent (Jones-Summers, NBER w27863). Because private actors capture so little, the world persistently buys less innovation than the social return justifies: capitalism's greatest asset class is permanently on sale.

The debits are entered at full strength, then netted. The global top 10 percent takes about 52 percent of income; the China Shock cost 2.0–2.4 million American jobs with decade-long local scarring; average US markups drifted from 18 percent over cost in 1980 to 67 percent by 2014; the high-growth startup premium collapsed from 16 to 4 percent after 2000. These are real, bounded, and — this report will argue — maintenance failures of competition, not refutations of it. The credits run one to two orders of magnitude larger.

The multiplier and the bet

The audit is not the news; the news is what happens next. Every mechanism by which markets create value — prices, competition, innovation, entrepreneurship, capital allocation, trade, institutions — is at bottom an information-processing or transaction-cost machine. Agentic AI is a general-purpose collapse in the cost of exactly those two things. The macro debate is one disagreement wearing two spreadsheets: Goldman Sachs projects a ~7 percent global GDP uplift over ten years; Acemoglu bounds the effect at 0.71 percent of total factor productivity. Acemoglu's arithmetic assumes AI stays a tool applied to fixed tasks; Goldman's assumes it reorganises tasks, firms, and markets. The tool scenario yields a rounding error. The actor scenario yields a second engine.

The evidence floor already tilts one way. Support agents with an AI assistant resolved 14 percent more issues per hour — 34 percent more for novices (5,179 agents, Brynjolfsson-Li-Raymond). Writing tasks ran 37 percent faster at measurably higher quality (Noy-Zhang). Eighty percent of US workers have at least a tenth of their tasks exposed; agent tooling lifts the share of quickly completable tasks from 15 percent to roughly half (Eloundou). Sum the mechanism-by-mechanism arithmetic and the conditional agentic dividend lands at $10–20 trillion a year within a decade (ENSI estimate).

Conditional — because the Nordhaus question returns. The 2.2 percent producer share was never a law of nature; it was an achievement of antitrust, patent expiry, open science, and low entry barriers. The agentic era re-runs the experiment: will the producers of intelligence keep 2 percent, or 40? The warning light is already on — markups rose from 18 to 67 percent over cost before a single agent transacted. The to-do list is therefore institutional, not technological: agent-readable markets, open-agent access, identity and liability law, micropayment rails, telemetry before strategy.

The Audit in Numbers

Twelve figures carry the whole argument. Three charts show the split, the poverty event, and the live bet. Every value traces to a library document; estimates are labelled.

Anthropic EI · ~36% of occupations use AI on ≥25% of tasks 57 / 43 · augmentation vs automation split 80% of US workers ≥10% task exposure · Eloundou

Contents

Method & Sources

Built on a downloaded, verified library: every number traces through one audited data pack to a primary document on disk.

The corpus spans the full perimeter: what capitalism is (Hayek, North), who captures the value (Nordhaus, Brynjolfsson), the growth and poverty record (Maddison, World Bank, UN), the seven mechanisms (NBER, Fraser, Kauffman, PIIE, CMA, CEA, OECD), the critiques (World Inequality Lab, Stiglitz, Ostrom), and the agentic frontier (Anthropic, DeepMind, BIS, x402, arXiv).

How the numbers were audited. Every figure lives in one data pack tagged to its source document; derived arithmetic sits in a calculations file with conservative and central variants. Verbatim quotations were admitted only after character-level verification against the library PDF; the unverifiable are paraphrased. Back-of-envelope aggregations are labelled "ENSI estimate"; source substitutions are disclosed on page 118.

The Donation Machine

Before the audit begins, one prior has to go: the belief that profit measures what capitalism takes from society. The audited accounts show the opposite — profit measures the roughly two percent of created value that the machine lets its winners keep.

The conventional view is coherent and wrong. It says: firms profit, therefore society pays; the richer the winners, the larger the extraction; GDP measures the benefit, so whatever GDP misses does not exist; and AI is simply the next rent-seeking industry arriving on schedule. Each link in that chain fails against a primary source. Profits are not the value — Nordhaus's accounting shows they are about 2.2 percent of the value, with the rest escaping to consumers through falling prices and better goods. Growth is not zero-sum — poverty collapsed fastest precisely where markets integrated, and more than a billion people exited extreme poverty while population grew. GDP is not the benefit — the median user values free search at $17,530 a year and GDP records it near zero. Innovation does not mainly reward innovators — a dollar of R&D returns $5–13.30 to society.

The reframe, then: capitalism is best understood not as an extraction machine with side benefits but as a donation machine with a retention fee. The donation is not chosen; it is forced — by imitation, entry, and patent expiry — which is exactly why it is reliable. Charity fluctuates with sentiment. The competitive giveaway has run for two centuries because no single winner can stop it.

Anatomy of the 2.2 percent

Nordhaus's method deserves a page because the entire report leans on it. He defined Schumpeterian profits — profits arising from the appropriation of returns to innovation, as distinct from returns to routine capital — and estimated them for the US non-farm business economy across five decades of national accounts. Result: Schumpeterian profits averaged 3.8 percent of corporate profits, and producers captured 2.2 percent (range 0.8–3.3) of the total social surplus their innovations generated. The finding is conservative by construction: it uses measured productivity gains, and the new-goods literature (Hausman) shows measured gains undercount welfare from genuinely new products.

How does the surplus escape? Through four repeating steps, each documented later in the mechanism catalog:

1. Rivals imitate. Knowledge leaks; spillovers roughly double private returns to R&D (Bloom-Schankerman-Van Reenen), and the imitator's customers pocket the difference.
2. Entrants price-compete. New firms enter above-profit markets and bid margins down — the surplus migrates from the income statement to the price tag.
3. Patents expire. Temporary monopoly is the system's deliberate, time-boxed exception; at expiry the remaining surplus transfers to users permanently.
4. New goods diffuse. Products spread from early adopters to everyone; the welfare gain of novelty itself is the part GDP counts worst (Hausman).

The 45x giveaway — and why it is an achievement, not a law

To feel the scale, take one firm. Apple's cumulative net income is on the order of $1 trillion. If smartphone-era profits sit anywhere near the Nordhaus split, the consumer surplus created by the smartphone category is measured in the tens of trillions of dollars — an ENSI back-of-envelope figure, but one independently consistent with the willingness-to-accept evidence on page 6, where a single free service is worth five figures a year to its median user. The most profitable company of the era is, on the audited arithmetic, one of history's largest net donors — involuntarily.

Involuntarily is the operative word, and it cuts both ways. Competition forces the donation: no firm can hold surplus that imitation, entry, and patent expiry are structurally designed to pry loose. But that means the 2.2 percent split is not physics. It is the output of maintained institutions — antitrust enforcement, patent terms that actually expire, open science, low entry barriers. Where the maintenance lapses, the split drifts: average US markups rose from 18 percent over cost in 1980 to 67 percent by 2014 (De Loecker-Eeckhout) — the machine still creating value, the giveaway share quietly eroding. The reframe therefore ends in an obligation, not a celebration: the giveaway is the thing the rest of this report exists to defend — into an era in which the producers of intelligence itself will test it again.

NBER w10433 · Nordhaus 2004 NBER w23687 · De Loecker-Eeckhout PNAS 2019 · Brynjolfsson et al. ENSI Calculations §2 · giveaway multiplier

01 The Counterfactual

The modern market economy produces on the order of $150–180 trillion a year of output that a pre-industrial world would not — the largest positive number in the human record, and one almost nobody has ever seen written down.

Every audit begins with a baseline, and the baseline is the part everyone forgets. For nearly the whole span of human existence the material life of the median person did not improve. A peasant of 1750 lived at roughly the standard of a peasant of 1200, or of Roman Egypt: a few dollars a day, half of all children dead before adulthood, life expectancy stuck in the thirties. The Maddison Project's reconstruction — GDP per head for more than 160 countries, annual back to 1820 — draws the same flat line through every pre-modern society. Stagnation was not a failure state. It was the normal state.

Then, around 1800, in a corner of northwestern Europe, the line bent and kept bending. Real income per head in the frontier economies rose roughly 30-fold — McCloskey's "3,000 percent" — a break with all prior history that compounding at a mere 2 percent a year, finally sustained instead of lost to war, plague or expropriation, was enough to manufacture. The human mind is not built to feel a thirty-fold change; we round it away and treat the modern standard of living as a fact of nature rather than the most improbable event in the economic record.

Maddison Project · 2020 update McCloskey · The Great Enrichment ENSI · CALCULATIONS §1

The experiments history ran

Cross-country correlations invite every objection in the book — rich countries can afford good institutions, geography confounds everything. The decisive evidence is different in kind: cases where the same people, on the same land, at the same moment were split between institutional regimes, and one clean instrument that isolates institutions from everything they travel with. Together they function as the control group the world never volunteered to be.

Read together, the cases control for culture, continent, resources and starting income — and the variable that tracks the outcome every time is the institutional package. This is as close to a randomised trial as macro-history allows, and it points one way: the flow computed on the previous page is not a coincidence of geography or a dividend of culture. It is produced by rules, and the rules can be identified.

AJR · MIT AER 2001 NBER w9124 · Botswana Maddison [03] · divided-country series ENSI library · angle 15, case studies

The freedom gradient, and the honest half

The natural experiments establish causality at the extremes; the cross-country gradient shows the dose-response in between. Sort the world's economies into quartiles by economic freedom — property rights, sound money, open trade, light regulation — and income per head climbs the ladder step by step, ending 7.6 times higher at the top than at the bottom.

The objection writes itself: gradients can be bought at the bottom's expense. The same dataset answers it. The poorest tenth of the population in the freest quartile earns $7,610 a year — more than the average citizen of the least-free quartile ($6,968), and seven and a half times the $1,010 the poorest tenth receives where freedom is lowest. The gradient does not trade the poor's welfare for the average; it lifts both.

Now the attribution, stated with deliberate conservatism. The experiments imply same-people gaps of 10:1 and up; the instrument puts a factor of seven on institutions alone; the gradient runs at 7.6x across 165 economies. Any of these read literally would attribute the great majority of the modern flow to the institutional package. This audit does not need the majority. Assign it half.

A payout of that size forces the next question, the one the whole politics of capitalism turns on: who keeps it? The answer — measured, not asserted — is Analysis 02.

02 The Split

For every dollar of value post-war innovation created, the innovating firms kept about 2.2 cents. The other ninety-eight went to everyone else. The fortunes we argue about are the visible sliver of a dividend that is mostly invisible.

The intuitive model of capitalism — the one that powers most political argument — is that firms create value and capture it, leaving the public whatever they are compelled to leave. The data invert this. Working through US non-farm business data across the entire 1948–2001 period, William Nordhaus asked what fraction of the total social returns to innovation the innovators themselves actually kept as profit, rather than losing to lower prices and better products for consumers. His answer: 2.2 percent, ranging from about 0.8 to 3.3 percent across specifications. The public is not the residual claimant on the value markets create. The public is the primary claimant, and the firm is the residual.

Nothing in the chart is a moral claim; it is an accounting identity about where the surplus landed. The interesting question is why a system built on the pursuit of profit pays out ninety-eight percent of the winnings to people who never signed the cheque — and whether that ratio is a law of nature or an achievement that can be lost. That is where this analysis goes.

Pricing the giveaway

Ratios anaesthetise; dollars wake the reader up. Invert the Nordhaus share and the split becomes a multiplier: if producers keep 2.2 percent, then every dollar of genuine innovation profit you can see implies roughly forty-five dollars of surplus you cannot — dispersed into lower prices, better products, saved time and new capabilities, collected by consumers who never sent the innovator a cheque.

Apply it to the most scrutinised fortune of the era. Apple's smartphone-decade profits are the standard exhibit for the prosecution — the margin on the device in a billion pockets. The Nordhaus lens flips the exhibit: if profits of that scale sit anywhere near the historical split, they are the small number in the transaction.

This is the arithmetic that the political argument about "big tech profits" never runs. It does not acquit any firm of any particular abuse — market power, privacy, lock-in are audited elsewhere in this report at full strength. It establishes something prior: the direction of net flow. The transaction between the innovating firm and society is, by two orders of magnitude, a transfer to society.

NBER w10433 · Nordhaus 2004 NBER · Hausman, Valuation of New Goods ENSI · CALCULATIONS §2

03 The Invisible Ledger

The median American would need to be paid $17,530 a year to give up search engines — value the national accounts record at roughly zero. The largest line items in capitalism's payout are the ones GDP cannot see.

GDP counts what is bought. It is close to silent on what is gained. When a free search engine replaces a paid encyclopaedia, GDP falls; when a mapping app worth thousands a year to a household costs it nothing, GDP records nothing. The error runs in one direction — against the system being audited — so an honest audit must re-instrument. The cleanest instrument is the one Brynjolfsson, Collis and Eggers built in massive incentive-compatible online choice experiments: stop asking what people pay for a good, and ask what they would have to be paid to give it up.

Two readings of the chart, both correct. The comfortable one: free digital goods are a gigantic unrecorded gift. The uncomfortable one: our headline measure of prosperity has drifted so far from experienced welfare that its single most-used category — the search bar — is valued at seventeen thousand dollars a year by its median user and at approximately nothing by the statistical system that judges the economy delivering it.

PNAS 2019 · Brynjolfsson-Collis-Eggers NBER w25695 · GDP-B Stanford/AER · Goolsbee-Klenow

Booking the entry: what the uncounted surplus sums to

Aggregate the instrument readings for the United States alone, and label the arithmetic for what it is. Search's median WTA of $17,530, carried across roughly 210 million adult users, implies on the order of $3.7 trillion a year of surplus from one product category. Extend across the full digital bundle — email, maps, streaming, social, the rest of the free stack — and the plausible total lands at $5–8 trillion a year of consumer surplus that the national accounts record at approximately zero (ENSI estimate, CALCULATIONS §3). Set that against the measured digital-sector contribution to GDP, near $2 trillion, and the invisible entry is two-and-a-half to four times the visible one.

Nothing here is new behaviour by the economy — it is old behaviour by the instrument. New goods have always entered the accounts at zero surplus; the free digital stack merely industrialised the blind spot. Which is why the correct reading of this entry is not "add $5–8 trillion to the score" but something stronger and more disciplined: every measured number in this audit is biased downward.

The floor principle

Three analyses in, a pattern has emerged that deserves to be named, because it governs how every remaining number in this report should be read. The counterfactual flow of Analysis 01 is computed from measured output. The 97.8 percent society-share of Analysis 02 is computed against measured surplus. And Analysis 03 has now shown that measurement itself is structurally blind to whole categories of value — free goods, new goods, quality — and blind in one direction only.

The rule has teeth. It means the $150–180 trillion counterfactual understates the machine's true annual payout; on the corrections surveyed here, the unmeasured surplus plausibly adds another 50–100 percent on top of the measured flow (ENSI estimate, CALCULATIONS §10). It means the producer share of innovation value is at most 2.2 percent, and the giveaway multiplier at least 45. And it means the burden of proof in the political argument sits where few place it: the case against the machine must be made net of trillions in payouts its critics' own statistics never counted.

• Free goods — priced at zero, valued at thousands per user per year; $5–8T/yr uncounted in the US alone (ENSI estimate; Brynjolfsson-Collis-Eggers, PNAS 2019).
• New goods — the fall from "unavailable" to "purchasable" is the largest welfare event in a product's life, and price indices never record it (Hausman, NBER).
• Health & longevity — the doubling of life expectancy appears in no output statistic; Analysis 06 prices it, and it may be the largest entry of all (Jones-Klenow, AER 2016).

The floor principle also disciplines this report's own side of the argument. Where ENSI aggregates run above what conservative methods support — WTA above WTP, stylised counterfactuals — they are labelled, haircut, and quoted as ranges. An audit that inflates its credits forfeits its debits; ours must survive the halving of every soft number, and does.

What the floor principle cannot tell us is why the machine keeps generating value to undercount — why the surplus is created at all, year after year, rather than exhausted. The answer is the machine's core reactor: ideas, and the extraordinary arithmetic of what a dollar of research returns to society. That is Analysis 04 — The Compounding Engine.

04 The Compounding Engine

Every dollar the market system invests in research returns roughly $13.30 to society — a 67 percent social rate of return no private portfolio has ever matched. The investor keeps a sliver; everyone else gets the compounding — and the engine still runs chronically under-fuelled.

The first three analyses counted value already delivered. This analysis audits the machine that manufactures the next round. Its headline is the most consequential ratio in the modern growth literature: one dollar invested in research and development returns, on the best economy-wide estimate, $13.30 of social benefit in present value — about $5 under deliberately conservative assumptions, more than $20 once health gains and international spillovers are counted.

Expressed as an interest rate, the baseline is a 67 percent social internal rate of return, against the 10–15 percent private investors earn on research-intensive assets. That gap is the defining fact about how a capitalist economy treats its most productive activity — and it is the phenomenon itself, not a flaw in the study design: firm-level work by Bloom, Schankerman and Van Reenen finds social returns to corporate R&D run roughly twice private returns from technology spillovers alone. Knowledge leaks — by imitation, employee mobility, patent expiry — and almost none of the return can be invoiced by the party who paid for the research.

NBER w27863 Jones & Summers · Social Returns to Innovation NBER Bloom-Schankerman-Van Reenen · Technology Spillovers Stanford Bloom-Jones-Van Reenen-Webb · Ideas Harder to Find

Five, thirteen, twenty: a return no market can arbitrage

The Jones–Summers figure is not a survey average of project-level studies; it is an economy-wide accounting exercise. If long-run growth in income per head comes from innovation, then the increment of GDP attributable to that growth is the return on the economy's total innovative effort — an accounting that nets out every spillover margin at once: imitation, business stealing, intertemporal displacement. The denominator is everything the United States spends on R&D across business, government, universities and non-profits: about 2.7 percent of GDP. The entire innovative effort of the largest economy on earth is a rounding error against its own output, earning a return no pension fund will ever see.

Stress-test it and the number bends but does not break. Impose a 20-year diffusion delay and the benefit is still $4.90 per dollar — the internal rate of return falls to 11 percent, still above any hurdle rate a market applies to anything. Count health gains, inflation mismeasurement and spillovers to the rest of the world, and the authors' own arithmetic runs to $10, even $20 per dollar — for the US economy alone.

The firm-level evidence agrees from below. Across US panel data for 1980–2001, Bloom, Schankerman and Van Reenen used firms' distinct positions in technology space and product space to separate knowledge spillovers from business stealing: both operate, the spillovers dominate, and the economy under-invests in R&D as a result. Nor is the return confined to private laboratories — the Dallas Fed finds high long-run productivity returns to government R&D as well (Fieldhouse & Mertens, 2023).

The under-fuelled engine

The tension: ideas are getting harder to find

The case for the engine would be simple if research productivity were constant. It is not — it has been falling for as long as we can measure it. Bloom, Jones, Van Reenen and Webb, working across every domain where inputs and outputs can both be counted, find the same pattern: sustaining a constant rate of progress requires an ever-larger army of researchers, with total growth held steady only because research effort scaled massively to compensate (Stanford, 2020).

Read the two findings of this analysis together and the tension resolves into a single sentence: the binding constraint on capitalism's dominant value mechanism is the volume of effective research effort — an input getting more expensive precisely as its measured social return sits at 30–70 percent. Movement III returns to this machine and asks what happens when AI agents cut the cost of the one input the engine is short of.

1. Innovation is the dominant term in the audit. $5–13.30 of social benefit per $1 of R&D at floor and baseline, $20+ at the ceiling — no other documented expenditure category returns in this range (Jones & Summers).
2. The giveaway is structural, not incidental. Spillovers dominate business-stealing at the firm level; producers keep ~2 percent at the economy level. The mechanism is competition, imitation and patent expiry — not benevolence (Bloom-Schankerman-Van Reenen; Nordhaus).
3. Ideas are getting more expensive. 18x the researchers for the same cadence; research productivity down ~41-fold since the 1930s. The constraint is effort volume, and it tightens annually.
4. The policy arbitrage is standing. Public and private R&D both earn measured 30–70 percent social returns; funding the gap is the closest thing macroeconomics has to a free lunch (Movement III).

05 The Poverty Event

In 1990, nearly four humans in ten lived in extreme poverty. By 2019, fewer than one in eleven did — while the planet added 2.3 billion people. It is the fastest mass improvement in human welfare ever measured, and this analysis records both its size and the honest allocation of the credit.

Begin with the baseline, because the baseline is what makes the event legible. In 1990, close to four in ten human beings lived below the extreme-poverty line — $2.15 a day at 2017 prices, the threshold of bare subsistence. That figure was already an improvement on every earlier century; for most of recorded history, most of humanity lived at or near that line as a permanent condition. Poverty was not a policy failure. It was the default state of the species.

Then, in a single generation, the default broke. By 2019 the extreme-poverty rate had fallen to 8.4 percent. More than a billion people crossed the line upward — and they did it while world population grew by roughly 2.3 billion. Hold that denominator in view, because it is what separates this event from every partial improvement before it: the global economy did not lift a fixed pool of the poor over a threshold. It absorbed two billion additional human beings and still cut the absolute count of the destitute by over a billion. Halving a rate while the denominator swells is a categorically harder feat than halving a count in a static population, and it is the feat the data record.

The accounting here is unusually clean, which is why this entry carries the audit's highest confidence grade. The World Bank has counted the same thing, the same way, at a consistent line, for a generation; the United Nations ran an independent flagship series over the same period on different survey infrastructure. The two agree on direction, magnitude and timing. Whatever else in this report rests on estimation, the poverty event does not.

What remains is the part an honest audit cannot skip: who and what produced it. Three-quarters of the decline happened in one country, and that country's own auditors credit markets and state capacity, not markets alone. Trade was a central channel. And a welfare-weighted reading — what a dollar is actually worth to someone living on two of them a day — makes this the most valuable entry in the entire audit per dollar moved. Those are the next three pages.

Three decades, one line, one direction

Two independent flagship series carry the same story. The Millennium Development Goal, set in 2000, aimed to halve the 1990 extreme-poverty rate by 2015; the UN's final accounting confirms the target was met five years early, with the rate more than halved and over a billion people out (UN, MDG Report 2015). The World Bank's series, built on different survey infrastructure, concurs on every order of magnitude. This matters for the audit because it retires the reflex that the number is an artefact of one institution's method: two instruments, one event.

The China honesty — and the trade channel

What a dollar is worth at two dollars a day

GDP arithmetic treats every dollar the same. Welfare arithmetic does not — and the difference is the reason the poverty event outranks its modest weight in world output. Under the standard log-utility assumption that anchors the Jones–Klenow welfare framework, the value of a marginal dollar falls in proportion to the income of the person receiving it. A dollar reaching someone at $2 a day is therefore worth on the order of 10–30 times the same dollar at rich-world incomes (Jones–Klenow method; ENSI application). The event this analysis audits moved income to exactly the people for whom income is most valuable.

Priced as a flow: lifting a billion people from roughly $1.90 to $5.50 a day equals about $1.3 trillion a year of income now reaching the world's poorest (ENSI estimate). Against a $110 trillion world economy that is barely one percent — and that is precisely the point. Weighted by what the dollars are worth to the people who received them, the smallest major entry in the GDP ledger becomes one of the largest entries in the welfare ledger. Any audit of what capitalism has paid society that stopped at unweighted dollars would miss it.

Log-utility weighting: the marginal value of a dollar scales with 1/consumption, shown relative to a $60/day reference consumer. ENSI illustration of the Jones–Klenow method — the source of the 10–30x range used in the text.

1. The event is real and twice-verified. Two independent flagship series — World Bank and UN — concur on a fall from ~38% to 8.4% of humanity, 1990–2019 (very high confidence).
2. The denominator makes it historic. The world absorbed 2.3 billion additional people while cutting the count of the destitute by more than a billion; the MDG halving target arrived five years early.
3. The credit is shared, honestly. China accounts for ~75% of the decline, and its own auditors attribute the result to market reform plus state capacity — a finding to be quoted whole, never in half.
4. Trade was a central channel. WTO–World Bank analysis ties the billion-person exit to openness; the value-chain evidence shows the wage mechanism at work.
5. Welfare-weighted, this is the audit's most valuable entry per dollar. At 10–30x weighting, ~$1.3T/yr flowing to the poorest (ENSI estimate) outweighs far larger flows higher up the distribution.

06 The Welfare Multiplier

GDP counts what people buy. It does not count how long they live to enjoy it. Price the mortality, the leisure and the equality alongside the consumption, and the welfare delivered by modern growth has been running roughly 1.5 times the measured number — the audit's largest correction, and it points upward.

Income is a means. This audit would close hollow if the thirty-fold enrichment of Analysis 1 and the poverty exit of Analysis 5 had not converted into the things people actually want — years of life, healthy children, literate minds, hours not worked, the disappearance of routine catastrophe. It did convert, and this analysis prices the conversion. The headline finding: measured growth has systematically understated the welfare delivered, and by a wide margin.

The most direct statement of the multiplier comes from Jones and Klenow's welfare accounting, the most rigorous attempt yet to put consumption, leisure, inequality and mortality into a single consumption-equivalent number. Across the countries where the calculation can be run over time, welfare growth averaged 3.1 percent a year against income growth of 2.1 percent between the 1980s and the mid-2000s — welfare compounding half again as fast as GDP, chiefly because people were living longer and the extra years are worth a great deal (Jones & Klenow, AER 2016).

The physical fact underneath the accounting is the largest welfare gain any statistic captures: over the market era, life expectancy in the frontier economies roughly doubled — from the thirties and forties of 1900 to the high seventies and eighties today — and the poorest countries have been converging on the frontier faster in health than in income. None of that doubling is ever purchased in a shop, so almost none of it appears in GDP. A statistic built to track production will faithfully record the pharmaceutical's price and entirely miss the decades of life it added.

The pages that follow do the pricing in three steps. First, the master relationship — the Preston curve — and its most optimistic property: it keeps shifting up. Second, the mechanics: what actually drove mortality down, and how the Jones–Klenow method converts longevity, leisure and equality into consumption-equivalents, with France as the worked example. Third, the multiplication: what the welfare wedge does to the audit's headline counterfactual flow.

The Preston curve — and why it keeps shifting up

Samuel Preston's 1975 finding is the master relationship of this analysis: plot national income per head against life expectancy across countries, and the points trace a strong positive curve that rises steeply among the poorest countries and flattens among the richest. Money buys life most efficiently exactly where there is least of both — one more reason the poverty event of Analysis 5 is the audit's most welfare-dense entry.

But the deeper finding — the one that carries this analysis — is what happened to the curve over time. It did not sit still. The entire curve shifted upward across the twentieth century: at any given level of income, people lived substantially longer in later decades than earlier ones. Two forces, then, not one. Income growth moved societies along the curve; the diffusion of knowledge — germ theory, sanitation, antibiotics, vaccines, public health — moved the whole curve up. And both forces were downstream of the same market-driven enrichment: the growth paid for the science, built the water systems and manufactured the vaccines at prices poor countries could eventually afford. The doubling of life expectancy is the compound interest on both accounts at once.

Preston 1975 WHO Bulletin reprint NBER w11963 Cutler-Deaton-Lleras-Muney OECD How Was Life? 1820–present UNDP Human Development Reports 1990 · 2023-24

Pricing a longer life: the Jones–Klenow correction

Two hundred years of welfare, priced

Now run the multiplier through the audit's headline. Analysis 1 established a counterfactual flow of roughly $150–180 trillion a year of output above the pre-industrial baseline. That is a GDP-metric number — it counts the consumption and misses the lives. Apply the Jones–Klenow wedge, under which welfare-relevant growth has historically run about 50 percent above measured growth where mortality was falling, and the welfare value of modern growth lands at the equivalent of $225–270 trillion a year of consumption — an order-of-magnitude estimate, labelled as such, but built from the most careful welfare accounting the literature has produced.

The two-century record supports the direction of the correction everywhere we can measure it. The OECD's How Was Life? project, assembling global well-being data since 1820, finds the market era is also the era in which real wages, height — the biologist's proxy for childhood nutrition — literacy and schooling went from elite privileges to near-universal norms. And the UNDP's Human Development Reports bracket the modern arc: the first report, in 1990, was written precisely as a corrective to income-centric accounting — people as the real wealth of nations — and the 2023–24 edition's long-run series records a world higher on every component of the index, income, schooling and longevity alike, than the world the first edition measured, recent setbacks included.

1. Life expectancy roughly doubled over the market era — the single largest welfare gain any statistic captures, and one GDP records only incidentally (Preston; OECD).
2. The Preston curve shifted up, not just along. Income growth and knowledge diffusion compounded — and both run downstream of the same enrichment that funded the science and the sanitation.
3. Welfare growth ≈ 1.5x measured growth — 3.1% vs 2.1% a year in the Jones–Klenow accounting — because added years of life are worth so much once priced.
4. The correction is honest in both directions. France rises from 60% to 92% of US living standards; high-mortality, high-inequality countries fall below their measured income. This is a yardstick, not an advocacy tool.
5. Priced properly, modern growth is worth $225–270T/yr consumption-equivalent (ENSI estimate) — and the audit's headline numbers remain floors, not ceilings.

07 The Honest Debits

An audit that reports only the credits is not an audit but a brochure. Before this report nets anything, it books the four largest entries in the red column — at full strength, in their authors' own numbers, with nothing softened.

The preceding six analyses assembled the credit side of the ledger: a $150–180 trillion annual counterfactual flow, a ~98 percent surplus giveaway, a billion-person poverty exit, trillions of uncounted digital surplus, a doubled life expectancy. A reader is entitled to suspect an audit that finds only good news. So this analysis inverts the burden. The four debits below are the strongest documented entries against the machine, drawn from the same library and stated the way their authors state them — not the way the machine's defenders would prefer them stated.

They are: the concentration of income at the top of the global and national distributions; the concentrated, slow-healing job losses that trade inflicted on identifiable American communities; the four-decade drift of market power visible in markups and returns on capital; and the thinning of the high-growth entrepreneurial tail that does the machine's creative destruction. Each is real. Each is measured by economists with no brief for capitalism's defence — the World Inequality Lab, the labour economists who named the China Shock, the industrial-organisation team that built the markup series.

Two of the four, we will argue, are questions about who holds the surplus; two are maintenance faults that degrade tomorrow's surplus. But that classification comes later. Pages 39 and 40 present the red ink uninterpreted. Page 41 does the netting — and shows the arithmetic that keeps the audit, decisively, in the black.

Real, bounded, concentrated — and one to two orders of magnitude smaller than the credits

Now the netting, with the debits conceded in full. Three things are true of all four entries. They are real — measured, replicated, undisputed in direction. They are bounded — none reverses or approaches the scale of the credit entries. And they are concentrated — which is precisely why they dominate the politics: a $2.1-trillion national dividend is diffuse and invisible, while a closed factory has an address.

Classify them and the ledger clarifies. Debit 1 is a question about who holds the surplus, not about whether it exists: the 52 percent the global top decile takes is a share of an income pool the machine multiplied thirty-fold — a distribution problem inside a creation success, treated in full in Analysis 09. Debit 2 is a gross loss already netted inside a credit entry that remains roughly ten-to-one positive: against the $2.1 trillion annual US trade dividend, the China Shock's adjustment costs run on the order of $0.2 trillion a year (ENSI estimate, from the ~10:1 net in this report's calculations) — real, unequally borne, and an indictment of the missing compensation layer rather than of exchange itself. Debits 3 and 4 are not destroyed value at all but maintenance faults: warning-gauge readings that tomorrow's giveaway will shrink if the competitive machinery is left unserviced. Nowhere in the 106-document library is there evidence of a failure in the machine's core function — that innovation stopped returning multiples, prices stopped aggregating information, or entry stopped creating jobs. The faults sit in the enforcement and access layers: the wear surfaces.

The netting statement, then: the debits are real and the credits run one to two orders of magnitude larger — and the two debit entries that most threaten the future (markups, dynamism) are exactly the ones with known, tested fixes. Which raises the question the next analysis answers: if the faults are fixable, what is the repair manual?

WIL World Inequality Report 2022 Berkeley Piketty & Saez, US top shares 1913–1998 ADH The China Shock, Annu. Rev. Econ. NBER De Loecker & Eeckhout w23687 CEA Benefits of Competition, 2016 NBER Decker et al. w21776

08 The Maintenance Thesis

Every debit on the preceding spread is an engineering fault with a documented fix. Read properly, the market-failure literature is not the machine's obituary. It is the machine's specification sheet — and its repair manual has been on the shelf for decades.

Start with the strongest theoretical objection, because the thesis stands or falls on absorbing it. Joseph Stiglitz's work on the invisible hand demonstrates that under imperfect and asymmetric information, with incomplete markets, competitive outcomes are generically not constrained-efficient — the assumptions of the First Welfare Theorem fail pervasively, so market results can in principle be improved upon (Stiglitz, NBER w3641). This is decisive against laissez-faire as a theorem. It is not decisive against the machine, for a reason engineers will recognise: a specification that states the operating conditions under which a system performs — and the conditions under which it degrades — is the most useful document a system can have. Stiglitz tells us where the faults will occur: wherever information is asymmetric and costs fall on parties outside the transaction. That is a map of the maintenance schedule, not a demolition order — especially since the alternative coordination technology, a planner with better information than the price system, is precisely what the calculation debate showed cannot exist.

Next, the repair manual. Elinor Ostrom's Nobel lecture documents a large empirical class of commons and externality problems — the canonical market failures — solved in practice by polycentric, self-governing institutions: neither market nor state, but nested layers of local monitoring, graduated sanctions and cheap conflict resolution (Ostrom, Nobel Lecture, 2009). The correction layer for price failures, her fieldwork shows, is itself usually decentralised. And at the level of the firm, Hart and Zingales argue that the objective function is adjustable from inside: companies should maximise shareholder welfare, not market value — because shareholders are people with ethical and social preferences, the corporate machine can internalise harms its owners actually care about without any change to its ownership structure (Hart & Zingales).

The honest limit of the thesis is political, and the library states it without flinching:

The gauges: five instruments on the machine's dashboard

If capitalism is infrastructure, it should be monitored like infrastructure — continuously, on named gauges, with thresholds that trigger service. The library supplies exactly five. Each has a published instrument, a current reading, and a toolkit that has worked before. Two are in the red; two are amber; one is the meta-gauge that decides whether the other four ever get serviced.

Sources: De Loecker & Eeckhout (NBER 2017); CEA (2016); Decker et al. (NBER 2015); Hsieh & Klenow (QJE 2009); World Bank WGI; Zingales (NBER 2017).

Read the dashboard as a whole and one pattern jumps out: the red gauges are all in the enforcement and access layers — competition, entry, dynamism — while the machine's core functions (innovation returns, price signalling, job creation by entrants) show no documented degradation anywhere in the library. That is the profile of a maintenance backlog, not a design failure. It is also the profile of a system whose faults are coupled: the markup drift feeds the entry decline, the entry decline thins the skewness tail, and unchecked concentration eventually reaches the governance gauge — the one fault that, if it trips, disables the repair crew itself. Hence the maintenance ordering: service gauge 1 first, and instrument gauge 5 permanently.

The steelman, answered line by line

The maintenance thesis earns its name only if it survives the strongest versions of the objections — stated the way their proponents state them, then answered or conceded. Six rows; two concessions inside the answers; no objection paraphrased down.

What the maintenance framing buys is not comfort but jurisdiction. "Is capitalism good?" is a referendum — unanswerable, unfalsifiable, permanent. "Which gauge is in the red and what is the service procedure?" is a work order. Every objection above converts into a line on that work order, which is the entire point: the critics' evidence, taken at full strength, specifies the repairs rather than the demolition. One question survives the conversion, though — the one no gauge settles: even a well-maintained machine distributes its output somehow. Who, over two centuries, actually got the Great Enrichment?

09 The Distribution Question

Who got the Great Enrichment? Two hundred years of data return a double answer: the gap between countries rose for 160 years and then began to close, while the gap within countries compressed for 70 years and then reopened. Most political argument quotes one curve and hides the other.

The authoritative source is Chancel and Piketty's reconstruction of the world income distribution from 1820 to 2020, built on the World Inequality Database. Its headline is sobering: global inequality "has always been very large," with the top decile's share oscillating in a 50–60 percent corridor across the entire period. But the level hides the mechanics, and the mechanics are the finding. Global inequality has two components — the gaps between countries and the gaps within them — and for the last century they have moved in opposite directions:

Hold both clauses at once, because each political camp owns exactly one. The globalising decades after 1980 produced the poverty collapse and the first sustained convergence between nations in modern history — and a widening of the gap inside most rich nations. The critic who cites only the second and the champion who cites only the first are each telling half the story. This analysis tells both halves, then asks what follows.

Two curves, one question

The schematic makes the audit's distributional finding legible in one look. For 160 years, the dominant engine of global inequality was geography — which side of the industrialisation frontier your country sat on, an accident of birth no individual effort could overcome. The mid-century compression inside rich countries (1910–1980) coincided with the era's wars, taxation and welfare-state construction. Then 1980: the curves cross trajectories. As markets integrated, the poor giants grew faster than the rich world for the first time in the modern record — pulling the navy curve down — while the rules-mix inside rich countries let their internal gaps reopen, pushing the gold curve up. The result, Chancel and Piketty note, is that early 21st-century capitalism carries a level of global inequality comparable to that of 1910 — but produced by a completely different mechanism, under a different set of rules. And rules, unlike geography, can be rewritten.

The compensation layer: the fix is transfers and mobility, not autarky

The China Shock has two candidate remedies, and they are not symmetric. The first — close the border, repatriate the supply chains — cancels the roughly $2.1 trillion annual dividend to protect against a ≈$0.2 trillion adjustment cost (ENSI estimate, Analysis 07): a ten-for-one destruction of national income that would also, applied globally, throw the between-country convergence machine into reverse against the very populations it just lifted. The second remedy accepts what Autor, Dorn and Hanson actually demonstrated. Their finding was never that trade fails to pay; it was that the adjustment the textbook assumed does not happen on its own — wages and participation stay depressed for a decade in exposed places. That is a finding about a missing policy layer, not about exchange. The fix it implies is a compensation layer funded out of the dividend it protects.

Even the authors of the inequality series point down this road. Chancel and Piketty's own forward-looking discussion concerns alternative rules — they model how fiscal revenue-sharing arrangements could significantly reduce global inequality — a redesign of the distribution machinery, not an abandonment of the production machinery. The two-century lesson of their data is that the top-decile share is set by rule-sets: it compressed under one (1910–1980) and reopened under another (1980–2020), around a production engine that ran throughout. Distribution is a policy output. Treat it as one.

1. Keep the aggregate gain. Openness stays; the dividend is the fund from which every repair is paid. A remedy that shrinks the surplus to rearrange it fails both curves at once.
2. Pre-commit the compensation. Adjustment assistance decided after the losses concentrate arrives a decade late — the decade ADH measured. Trigger it automatically on exposure, not on politics.
3. Fund people and places. Decade-long local scarring means mobility support alone is not enough; the evidence indicts both the worker's stranded skills and the county's stranded tax base.
4. Instrument the within-country gauge. Top-share telemetry (Piketty–Saez, WID) belongs on the maintenance dashboard of Analysis 08, watched continuously like the markup series.
5. Fix distribution with distribution tools. Taxes, transfers, predistribution and revenue-sharing act on the gap directly; tariffs act on it by burning the surplus. Match the instrument to the fault.

The audit's distributional verdict, then: the Great Enrichment was real and increasingly widely shared between nations; its sharing within nations is a live, worsening, and entirely policy-tractable fault. What the verdict still lacks is a control group — what the alternatives to this machine actually paid the societies that ran them. That is Analysis 10.

PSE/JEEA Chancel & Piketty 1820–2020 WIL World Inequality Report 2022 ADH The China Shock PIIE Hufbauer & Lu, PB 17-16 World Bank PSP 2022 · Four Decades UNU-WIDER Vietnam 2018

10 · The Systems Counterfactual

Analyses 1–9 priced what the market system delivered. This analysis prices what the alternatives delivered when actually tried — and the gap is the cleanest single piece of evidence in the file.

The standard critique compares capitalism with perfection and finds it wanting. An audit cannot work that way; the only defensible comparison is with the systems that actually competed with it. The twentieth century, at appalling human cost, twice split a single people on a single geography into two operating systems — Korea in 1948, Germany in 1945–49 — and within two generations the income gaps ran at ten-to-one to twenty-to-one. No other variable in social science moves an outcome by an order of magnitude while holding everything else constant.

The evidence arrives in four registers. The theoretical root: Hayek's demonstration of why a planning apparatus cannot, even in principle, do what a price system does. The micro ledger: the state-versus-private ownership literature. The marginal measurement: Hsieh and Klenow's misallocation accounting — socialism in miniature, priced per dose. And the before-and-after panels: Poland, Vietnam, Estonia and the Korean split.

These are histories, not randomised trials — but read together they control for continent, culture, colonial past, resources and starting income, and the gradient always runs the same way. The Fraser index states the cross-section bluntly: the freest quartile of economies produces 7.6 times the GDP per head of the least-free quartile.

The root and the receipts: from the calculation debate to the ownership ledger

AER 1945 Hayek · The Use of Knowledge in Society [01] Hayek Competition as a Discovery Procedure [01] NBER 1998 Shleifer · State versus Private Ownership [01] NBER 2003 Djankov, Glaeser, La Porta, López-de-Silanes & Shleifer [01]

Socialism in miniature: the misallocation meter

Every market economy administers some of its allocation — subsidised credit, licensing, connected firms, state banks — and Hsieh and Klenow built the meter that reads the resulting damage. Their method compares the marginal products of capital and labour across manufacturing plants within narrow industries. In a frictionless market these converge: inputs move until the gaps close. Where administration overrides prices, the gaps persist — and their dispersion measures how far allocation has drifted from what the price system would have chosen.

In China and India, the 90th-to-10th percentile ratio of plant-level revenue productivity runs at a factor of 6–7, against under 2 in the United States. Hypothetically reallocating inputs to US-level efficiency — no new technology or capital, just prices allowed to do their job — would raise manufacturing TFP by 30–50 per cent in China and 40–60 per cent in India, with output gains approaching a factor of four once capital accumulation responds. Misallocation is one of the largest sources of forgone income on the planet.

Read the right way round, the meter is the systems argument in marginal form: every administrative override of the price signal is a small dose of the same disease that killed the planned economies; the full command economy is simply the 100 per cent dose. The dose-response curve is continuous — China's partial reforms bought partial recovery exactly over the years its markets were liberalising. The counterfactual is not a Cold-War memory; it is a live gradient every economy sits on today.

Before and after: four societies that switched the machine on

The strongest evidence macro-history allows is the recorded before-and-after. Four panels, four decades, four continents' worth of excuses controlled away.

Read together, the panels control for continent, culture, colonial history, natural resources and starting income — communist and post-colonial, Asian and European, resource-poor and war-ruined — and the one variable that tracks the outcome every time is the adoption of market mechanisms under decent institutions. That is as close to a controlled experiment as macro-history allows, and it points one way. The alternatives were not cheaper, fairer versions of the machine; they were the same populations earning a fraction of the income, and in every recorded case the switch — not the endowment — moved the number.

11 · The Agentic Bet

Goldman Sachs says AI adds ~7 per cent to global GDP. Acemoglu's bound says at most 0.71 per cent of TFP in ten years. Both use the same task data. The ten-fold gap is not a disagreement about coefficients — it is a bet on whether AI stays a tool or becomes an actor.

The loudest macroeconomic argument of the decade looks like a duel of forecasts. Goldman Sachs estimates that generative AI could raise global GDP by roughly 7 per cent — about $7 trillion — and lift US productivity growth by almost 1.5 percentage points a year over the decade after widespread adoption, with the equivalent of 300 million full-time jobs exposed globally. Acemoglu, applying Hulten's theorem to the same exposure data, bounds total-factor-productivity gains at no more than 0.71 per cent over ten years — and nearer 0.55 per cent once hard-to-learn tasks are discounted. Ten years of argument compressed into one spread: these are not competing estimates of the same object.

Acemoglu assumes — explicitly, and to his credit — that AI remains a tool: a cost saving applied to a fixed set of existing tasks, with no new tasks, no reorganisation of markets, no change in the process of science. Goldman's number quietly assumes the opposite: economy-wide reallocation, new occupations, compounding adoption. If agents merely assist workers at today's tasks, the sceptic's bound is roughly right and the whole episode is a rounding error on trend growth. If agents become participants — arbitraging prices, shopping for consumers, running research pipelines, underwriting loans, transacting with each other at machine speed — the gain arrives through channels Hulten's theorem was never asked to price, because the theorem holds task shares fixed while agents change what the tasks are.

The sceptic's arithmetic, shown in full

An audit steelmans its strongest witness, and Acemoglu's calculation deserves to be displayed rather than summarised. By Hulten's theorem, aggregate TFP gains equal the GDP share of the tasks affected times the average cost saving on those tasks. Every input is drawn from the empirical literature this report has already cited — including the very experiments the optimists lean on:

Within its assumptions, the arithmetic is unimpeachable. The audit question is what sits outside the assumptions — because each exclusion is not an oversight but a deliberate scope line, and each one is a channel this report has separately priced:

• New tasks. The task set is held fixed at today's O*NET catalogue; occupations and products that do not yet exist contribute nothing — yet new tasks are where past general-purpose technologies earned most of their return.
• Reorganisation. No reallocation across firms and markets, no new market structures, no Coasean redrawing of firm boundaries — the channel Analyses 2–4 priced in the trillions.
• Science. Acemoglu explicitly sets aside AI "changing the process of science" — the one channel with the Jones–Summers 5–13x multiplier standing behind it (Analysis 4).
• Agents as actors. No transacting, negotiating, arbitraging software participants at all — the entire subject of Movement III.

If the assumptions hold through 2034, the bound is right and modest AI policy is correct policy. The empirical tripwires that would break them are specified on page 56 — several are already under measurement.

The floor under both camps: what field evidence already shows

Beneath the forecast war sits a base that is no longer speculative: randomised and staggered deployments in real firms, an exposure census of the US task economy, and — since 2025 — live usage telemetry. Every number on this wall is measured, not modelled; and every deployed system studied so far was a suggestion tool, the weakest possible form of agency, which makes each figure a floor.

The last tile is the quiet hinge of the whole macro debate: the leap from 15 per cent of tasks to roughly half is precisely the leap from model to agent. And note the direction of the distributional finding — in both flagship studies AI assistance compressed skill gaps inside firms, the opposite of what three decades of skill-biased technical change taught economists to expect.

What would settle the bet

A disagreement this large, resting on named assumptions, is empirically decidable — and unusually for a macro dispute, the deciding instruments already exist. Acemoglu's bound fails only if its scope lines break: if tasks appear that O*NET never catalogued, if automation displaces augmentation in the telemetry, if the cost of profitable automation keeps falling, or if agents begin transacting rather than suggesting. Each of those is observable quarter by quarter. The audit therefore closes this analysis not with a verdict but with an instrument panel — six indicators, each tied to a specific assumption, each with a named data source:

1. The automation share. Anthropic's telemetry currently reads 57/43 augmentation-to-automation. A sustained climb in the 43 — the model fulfilling requests, not assisting them — is the actor row arriving in the data. (Anthropic Economic Index [13])
2. The profitable-automation fraction. Svanberg's 23% is a 2024 cost estimate, and it falls as inference costs fall. Each revision mechanically loosens Acemoglu's bound — using his own arithmetic. (MIT 2024 [12])
3. Tasks outside the catalogue. The first appearance, at scale, of measured task portfolios containing work that did not exist in O*NET breaks the fixed-task assumption directly. (EI telemetry vs O*NET [13])
4. The research-productivity number. Nobody has yet measured agentic R&D's effect on research productivity at economy level. Because the Jones–Summers 5–13x multiplier stands behind this channel, that paper will be the most important growth-economics result of the decade. (Analysis 4; [04][12])
5. Sectoral price dispersion. If consumer and firm agents act as market participants, dispersion collapses Kerala-style, sector by sector; where it persists, the residual maps protected rents. (Jensen QJE 2007 template [15])
6. The markup series. De Loecker–Eeckhout's 18%→67% curve is the capture gauge: agentic commerce bending it down signals open competition; accelerating it signals the toll booth. This one decides the split, not the size. (NBER [06])

What the bet is not about deserves one line: nobody serious disputes the field-evidence floor. Double-digit gains in real deployments, skill compression inside firms, half the task economy in scope once tooling counts — that is common ground. The dispute is confined to whether the machine's mechanisms themselves get upgraded. Which is why the audit's final analysis stops forecasting a number and starts mapping the futures the bet can land in.

12 · Scenarios 2035

Eleven analyses reduce to two live uncertainties: does AI stay a tool or become an actor, and does the model layer stay open or concentrate? Crossed, they yield four futures for the value engine — and only one of the two axes is decided by technology.

Everything else in this audit is settled enough to carry forward. The mechanisms and their coefficients (Analyses 1–6), the debits and their repairs (7–8), the distribution record (9), the systems counterfactual (10) — these are the fixed physics of the scenario space. What Analysis 11 could not settle is the capability question: whether agents remain suggestion tools, in which case Acemoglu's bound holds and the decade adds a rounding error, or become transacting participants, in which case the field evidence and the amplification arithmetic point to an addition measured in trillions per year. That is the row of the matrix, and it will be decided mostly by laboratories and adoption curves.

The second uncertainty is older than AI. Nordhaus's 2.2 per cent — the audit's founding fact — was never a law of nature; it was produced by competition, imitation and patent expiry relentlessly passing surplus to consumers. Whether the agentic economy repeats that split depends on the market structure of its choke point: the model and agent layer. Open — interoperable, contestable, agent-access guaranteed — and the historical split holds. Concentrated — a handful of gatekeepers owning the agents that shop, negotiate and underwrite — and the 98/2 division can flip for the first time in the modern record. That is the column of the matrix, and it is decided by policy: protocol rules, access rights, competition enforcement. The one lever states actually hold.

The four cells that follow are not predictions. They are the audited coefficients of this report run forward under each pair of assumptions — Acemoglu's bound for the tool row, the amplification arithmetic of CALCULATIONS §8 for the actor row, the Nordhaus split and the De Loecker–Eeckhout markup drift for the two columns.

The scenario space

One picture carries the whole analysis. The vertical axis is the Analysis-11 bet; the horizontal axis is the Nordhaus question. Today's telemetry places the economy just below the centre line and drifting upward — while the model layer, supplied by a handful of firms, starts the decade on the right-hand side.

The colouring is deliberate: both right-hand cells are reachable from today's starting point without any technological surprise at all — concentration is the default commercial instinct of every platform. The left-hand column has to be built.

The four futures, card by card

Each card carries the scenario's arithmetic, its surplus split, and the leading indicators that would signal its arrival. All magnitudes are ENSI estimates (BOE), assembled per the p57 labelling rule.

Note what the matrix does not contain: a cell where the technology fails to matter at all — the field-evidence floor of Analysis 11 already rules that out — and a cell where concentration produces the big dividend and shares it. Concentration does not destroy the surplus; it re-addresses it. That asymmetry is the whole policy argument of the next page.

The policy-determined diagonal

Scenario matrices usually end with a shrug: four futures, pick your prior. This one does not, because the two axes are not symmetrical. The row — tool or actor — is set by model capability and adoption, which no mid-sized state controls and which the Analysis-11 instrument panel merely observes. The column — open or concentrated — is set by protocol rules, access rights, payment rails and competition enforcement: ordinary legislative material, decided jurisdiction by jurisdiction, in the next few years, while the equilibrium is still unformed. The audit's history supplies the precedent. The 2.2 per cent producer share that anchors this entire report was itself manufactured — by antitrust, patent expiry, open science and low entry barriers relentlessly handing surplus to consumers. The twentieth century chose its column, and the choice, not the technology, is what made capitalism the donation machine of Analysis 2.

Hence the diagonal that matters runs from the Quiet Toll to Kerala at Scale — from small gains privately gated to large gains publicly kept — and every step along it is a policy act, not a forecast: an open-agent-access rule, interoperable payment rails, agent identity and liability law, model-layer concentration watched the way central banks watch systemic banks. If capability stalls, those acts cost little and still bend the markup curve. If capability arrives, they are the difference between S3 and S4 — between the largest consumer-surplus event since the container ship and the largest toll booth ever built. Asymmetric stakes, symmetric cost: that is what "no-regret" means in an audit.

That sentence closes the audit. Movement I established what the machine pays, to whom, at what cost, against which alternatives, and under which futures. Movement II now opens the machine itself — the seven mechanisms that generate the payment, ranked by the value each creates.

The machine, exploded

Before the deep-dives, the schematic. Five working modules run on one information bus, and the whole assembly runs on one operating system. Rank order is by the social value each layer creates — and the output pipe leaks, by design, to everyone.

Each module has been measured separately, repeatedly, and by the best empirical economics of the past four decades. Innovation converts research effort into permanently higher living standards. The price system aggregates dispersed knowledge no planner could collect. Competition forces the gains through to consumers. Finance steers capital toward its best uses. Entry and exit run the experiment that discovers which uses those are. Trade extends the whole machine across borders. And institutions set the expected payoff to every action the other six reward.

The literature does not just say "markets work." It says which parts of the market machine do how much of the work — and that is what this movement extracts, one mechanism per section, in a fixed anatomy that makes the seven directly comparable.

Sources for this movement: NBER Jones-Summers · Nordhaus · Levine · Haltiwanger MIT AJR 2001 UChicago Hayek 1945 Stanford Hsieh-Klenow Fraser EFW 2024 PIIE Hufbauer-Lu

How the ranking works

A ranking exists to allocate scarce reform capacity. A state cannot defend everything at once; it needs to know what to protect first and what it can safely leave alone. Three criteria produce the order, applied in sequence.

One recurring finding across all seven sections — the closest thing this literature has to a law — deserves stating before the list: the value each mechanism creates does not stay with the people who operate it. Innovators keep ~2 percent; venture investors fund knowledge that spills to rivals; competitive firms hand efficiency gains to customers. The machine is a surplus-leaking engine, and the leak is the point.

The ranking at a glance

The whole movement in one strip. A Depth-2 reader can stop here; the deep-dives that follow supply the identification, the failure modes, and the agentic hooks.

The couplings, previewed

Ranking the mechanisms separately understates them jointly, because the strongest results in the library are about the interactions: import competition raised European firms' innovation and productivity (rank 3 feeds rank 1); venture capital exists to fund the high-variance young firms that create the disproportionate jobs (rank 4 feeds rank 5); prices let every module act at a distance; and the Hsieh-Klenow misallocation numbers show three mechanisms failing as one coupled system when institutions override prices. The machine's components are load-bearing for each other — the full treatment is the Compounding Stack, p. 93.

The build sequence. Each layer presupposes the ones to its left — which is why reform programs that start mid-chain (finance without courts, trade without prices) keep failing.

Two mechanisms carry live warning lights, flagged in their sections and read together in Movement I's debit ledger: competition (US markups up from 18% to 67% over cost, 1980–2014) and entrepreneurship (the high-growth skewness premium down from 16% to 4% since 2000). Both faults sit in the machine's enforcement and access layers — the wear surfaces — not in its core function. Nowhere in the 106-document library is there evidence that innovation stopped returning multiples, that prices stopped aggregating information, or that entry stopped creating jobs.

0 · Institutions — the operating system

Evidence chain: MIT AJR, AER 2001 NBER Rodrik-Subramanian-Trebbi Stanford/JEP North 1991 LSE Besley-Ghatak 2010 UT Austin Haggard-Tiede 2010 World Bank B-READY 2024 Fraser EFW 2024

The close: insure, don't invest. For a policymaker asking where does my next unit of reform effort earn most, institutions are usually the wrong answer in a functioning democracy — the 67% social return at rank 1 wins. For the same policymaker asking what must I never allow to break, institutions win and it is not close. Judicial independence, property security, constraint on expropriation: zero measured annual dividend, total counterfactual exposure. The premium on this insurance is vigilance, and it is the cheapest line in the national budget.

1 · Innovation & creative destruction

The engine: how creative destruction works

Growth in the Schumpeterian account is not accumulation but replacement: quality-improving innovations that render the old obsolete, formalised in the Aghion-Howitt model, where the engine of rising living standards is precisely the destruction of incumbent rents by better products. The market's contribution is to make this replacement routine — to give any firm with a better idea both the incentive (a temporary monopoly profit) and the right (free entry) to displace the installed base. The temporary rent is the carrot; imitation, patent expiry and competitive entry are the scheduled confiscation; and the confiscated surplus — lower prices, better products, freed ideas — is what society collects.

Hold the input in view alongside the return, because the ratio is the policy argument. The denominator of the Jones-Summers calculation — total US R&D across business, government, universities and non-profits — is only about 2.7 percent of GDP; venture capital, the sharpest allocation channel at rank 4, reached roughly $130 billion in 2018 but has often run below $30 billion a year. The entire innovative effort of the largest economy on earth is a rounding error against its own output, earning a return no pension fund will ever see.

Scale the ratio to the world and the magnitude shows: global R&D spending is roughly $2.9 trillion a year (2025). At $5–13.30 of social value per dollar, the world's research effort generates on the order of $14.5–38.6 trillion a year of social value — of which private actors capture a small fraction. The capitalist innovation system is society's highest-return asset class, and it is systematically under-supplied — which is exactly why the 67 percent social return persists instead of being arbitraged away.

The giveaway: why 98 percent escapes

Ideas are getting harder to find at exactly the moment a new class of researcher becomes available — the setup for this mechanism's agentic hook, overleaf.

2 · The price system

The users need never learn why. The price rises; every user economises; every producer of substitutes expands; the substitutes of the substitutes adjust; and the effect propagates through the whole economic system — without any participant knowing the cause and without any central instruction. The whole acts as one market not because anyone surveys the field, but because individual fields of vision overlap enough for the relevant information to reach all, through a number.

The marvel of unplanned coordination

Hayek's description of the machinery is literal, not poetic: the price system is "a system of telecommunications" in which the producer watches "the movement of a few pointers, as an engineer might watch the hands of a few dials," adjusting to changes of which he may never know more than is reflected in the price movement. In abbreviated form, by a kind of symbol, only the most essential information is passed on — and only to those concerned. The economy of knowledge with which the system operates is its most significant property: how little each participant needs to know in order to take the right action.

Hayek chose the word deliberately — "to shock the reader out of the complacency with which we often take the working of this mechanism for granted." Had the price system been the product of deliberate human design, he wrote, it "would have been acclaimed as one of the greatest triumphs of the human mind." Its misfortune is double: it is not the product of design, and the people guided by it usually do not know why they are made to do what they do. His companion essay completes the argument: competition is a discovery procedure — the facts the market coordinates (who can produce what, at what cost, for whom) are not data waiting to be collected but outputs of the market process itself, which is why no questionnaire can substitute for it.

The negative proof: what broken prices cost

Because the price system's value is a precondition, the cleanest measurements are of its absence. The comparative literature on state versus private ownership finds private, price-disciplined operation systematically outperforming state operation across industries and countries, for information and incentive reasons rather than ideology (Shleifer 1998; Djankov et al. 2003).

The marginal estimate comes from within market economies: Hsieh and Klenow measured marginal products of capital and labour across manufacturing plants within narrow industries in China and India — where subsidised credit, administered inputs and connections override prices — and found dispersion a working price system would arbitrage away:

Misallocation is not a rounding error; it is a large fraction of why poor countries are poor — and it is the price system's value read in the negative. China's own within-sample improvement — reallocation gains of roughly 2 percent of TFP per year, 1998–2004 — tracks the market reforms Movement I credited.

3 Competition — the enforcement mechanism

Evidence base for this mechanism:  NBER Borenstein & Rose 2007 LSE CEP Bloom & Van Reenen 2006 LSE CEP Bloom, Draca & Van Reenen 2011 CEA 2016 CMA 2024 OECD 2014 NBER De Loecker & Eeckhout 2017

The engineering read: this is the best-understood fault in the machine, with a two-century-old toolkit — antitrust enforcement, merger control, entry-barrier removal, occupational-licensing reform. The markup series is best understood as a maintenance backlog, not a design flaw. A machine with a worn brake pad needs a brake job, not a referendum on braking.

4 Capital allocation — the steering system

5 Entrepreneurship — the entry-exit experiment

6 Trade — the machine, extended across borders

Σ The compounding stack

Ranking the mechanisms separately, as this movement just has, understates them jointly — because the strongest results in the library are about the couplings.

The compounding is the deep reason referendum-style politics mishandles capitalism: you cannot keep the innovation dividend while suspending competition, or keep entrepreneurship while letting finance be politically allocated. The correct policy posture is the maintenance of a machine, not the adjudication of an ideology.

How the stack assembles — and why the whole exceeds the parts

The whole exceeds the parts because each mechanism's measured dividend was estimated with the others running. Jones and Summers' $13.30 assumes competition passes gains to users and finance funds the labs; Levine's +1pp assumes prices carry honest signals; Haltiwanger's job engine assumes entrants can borrow and incumbents can be displaced. The stack's arithmetic is multiplicative, and the leak — 98 per cent of the surplus escaping to society — is the emergent property of all seven layers operating at once.

III.1 The Amplification Thesis

Look again at the seven mechanisms of Movement II through one lens: every one of them is an information-processing or transaction-cost machine. That is not a metaphor — it is Hayek’s and Coase’s original arguments, and it is why agentic AI is not one more productivity tool.

Hayek’s point about the price system was never that markets are virtuous but that they are the only computer big enough: a distributed computation over knowledge no participant possesses in full. Competition works only where consumers can actually compare offers — search costs and obfuscation are its friction. Innovation is the production of ideas, whose main input is cognitive labour. Entrepreneurship is bounded by how much coordination one founder can afford — Coase’s transaction costs are what makes firms necessary and small teams weak. The BIS calls capital allocation the brain of the economy: “the processing and aggregation of vast amounts of information into price signals that coordinate participants in the economy.” Trade is limited less by tariffs than by the costs of search, contracting, language and coordination across borders. And institutions — North’s humanly devised constraints — are priced in the apparatus of verifying, enforcing and complying with rules.

If that reading is right, a technology that collapses the cost of information-processing and transacting does not add to the seven mechanisms the way a new industry adds output. It is an input into the machinery that produces all the other gains — it multiplies through every mechanism at once. That is the amplification thesis, and it is why the Goldman–Acemoglu gap of Analysis 11 (+7% of global GDP versus ≤0.71% of TFP) is not a disagreement about coefficients: it is a bet on whether AI stays a tool applied to fixed tasks or becomes an actor inside the mechanisms themselves.

BIS WP 1194 · Aldasoro et al. 2024 Hayek AER 1945 North JEP 1991 Goldman 2023 vs Acemoglu 2024

What the field experiments already show

The base of the argument is no longer speculative: real deployments already show double-digit gains — and every number below was produced by the weakest possible form of agency: a suggestion tool.

Usage is fact, and it is now measured

Exposure is potential. The Anthropic Economic Index turns usage into telemetry: over four million Claude.ai conversations mapped onto the US Department of Labor’s O*NET task taxonomy.

The telemetry shows an economy in the transition zone. Software development and writing account for nearly half of all usage, but diffusion is already broad: a third of occupations use AI on a quarter of their tasks within three years of the technology existing. Usage peaks in the upper-middle of the wage distribution and falls at both extremes. AI is performing economic tasks at scale — mostly alongside humans, increasingly instead of them.

Read as a dashboard rather than a debate, the evidence floor gives the amplification thesis its empirical tripwires. If the automation share stalls, if the profitably-automatable fraction of exposed tasks stays near Svanberg’s 23 per cent, and if measured task portfolios never contain tasks absent from O*NET, then Acemoglu’s bound holds and this movement overclaims. If the automation share climbs and new task categories appear, the actor scenario is arriving through measurable channels.

Anthropic EI 4M+ conversations · 2025 O*NET task taxonomy Acemoglu MIT 2024 · tripwires

III.2 Amplification Map I: Prices, Competition, Innovation

Prices — the continuous Kerala

Competition — the end of the search-cost moat

A large share of modern market power lives not in patents or scale but in friction: consumers who cannot compare, switching costs deliberately inflated, drip pricing, obfuscated fee schedules, loyalty exploited because re-searching the market costs an evening. The counterfactual of unpoliced friction is already on the books — average US markups rose from 18 per cent above marginal cost in 1980 to 67 per cent in 2014 (De Loecker & Eeckhout).

Microsoft Research’s analysis of the agentic economy identifies exactly this as the deep effect: the profound impact of generative AI lies in reducing “communication frictions between consumers and businesses”, not in speeding up existing workflows. Their canonical example is mundane and devastating: a consumer stays with a worse tax preparer because explaining her financial situation again is too costly. An assistant agent that carries her context, interrogates every rival’s service agent and re-runs the comparison monthly cuts that cost to roughly zero. Obfuscation pricing does not survive a counterparty that reads every footnote; search-cost moats do not survive a shopper with infinite patience. Every dark pattern in the e-commerce playbook is a bet on human attention limits, and agents void the bet.

But the same paper carries the conditional on which this whole movement turns. Unscripted interaction (agents can talk to anything) is not unrestricted interaction (agents are allowed to transact across firm boundaries). The architecture fork:

Who captures the surplus is therefore a policy outcome, not a technical one. Interoperability mandates, agent-access rights to published offers, data-portability instruments extended to agent-readable form: protocol policy is the competition policy of the 2030s.

Innovation — attacking the 18×-researchers problem

The research dividend, priced — and labelled

The report’s rule: back-of-envelope arithmetic is shown in full, every input sourced, every estimate labelled. This is the largest single term in the agentic dividend, so it gets the fullest audit.

The honest statement, in full: nobody has yet measured agentic R&D’s effect on research productivity at the economy level. The paper that does will be the most important growth-economics result of the decade. Until it exists, this channel is a sized hypothesis with strong priors — the field-experiment floor on cognitive work, the 5:1 social-return ratio, and visible agentic pipelines in materials, proteins and mathematics — not a fact.

III.3 Amplification Map II: Entrepreneurship, Capital, Trade, Institutions

Entrepreneurship — the one-person agentic firm

NBER Haltiwanger et al. · age not size NBER Decker et al. 2015 World Bank B-READY 2024

Capital — shrinking the misallocation tax

Finance is information processing — the BIS’s framing, not ours — and every step-change in information technology, from double-entry book-keeping onward, has restructured it. The paper works through the four functions, and in each the agentic gain has the same shape. Intermediation: credit decisions on richer data at lower cost extend lending to borrowers currently priced out by assessment costs — the small firm whose loan appraisal costs more than the margin it would generate. Insurance: underwriting and claims collapse from weeks to minutes. Asset management: private-bank quality at retail cost. Payments: the machine-native rails of page 107.

Agent underwriting is the Hsieh–Klenow correction run through the credit channel: capital flowing to marginal-product signals rather than to collateral and relationships. Even in advanced economies, the gap between well-managed and badly-financed firms is the same tax at smaller scale. The state’s lever is not to run credit agents; it is the data architecture that determines what agents can see — open banking, e-invoicing, machine-consumable registries. A compact economy with mandatory e-invoicing could make itself the cheapest place in Europe to underwrite a small firm, a comparative advantage worth more than most subsidy programmes.

NBER Levine · depth quartile ≈ +1pp growth NBER Kortum & Lerner · VC 3–10× patent potency BIS WP 1194 · 2024

Trade — the export department, rented by the hour

The binding constraint on who trades has never been tariffs alone; it is the fixed costs of exporting — finding counterparties, language, contract law across jurisdictions, customs paperwork, certification, payment risk. Those fixed costs are why exporting is dominated by large firms with compliance departments. Every one of them is an information or transaction cost, which is to say: agent-soluble. An export agent that identifies demand in nineteen markets, drafts compliant contracts in the counterparty’s language, files the customs declarations and monitors payment is a trade department rented by the hour. The distributional consequence matters most for small open economies: agentic trade intermediation is small-firm-biased, because it commoditises exactly the fixed costs only large firms could previously amortise. The prize is priced by the Frankel–Romer coefficient — every percentage point of trade share bought by lower fixed costs returns a multiple in income; the US baseline from trade expansion runs $2.1T a year, about $18,131 per household.

Institutions — ambient compliance, and the gap

Institutions create value by making promises enforceable, and their price is the apparatus of verification, enforcement and compliance — text, procedure and checking: the home turf of language-model agents. The amplification runs in three layers. Compliance: the fixed cost of knowing and following the rules, which falls disproportionately on small firms, becomes an agent service — regulation consumed by machines rather than interpreted by lawyers. Enforcement: agents that monitor performance and flag breaches early shrink the lag that Haggard and Tiede identify as the growth-relevant core of the rule of law. Reach: societies enforce only the rules they can afford to enforce; when verification costs collapse, the enforceable set expands — better institutions become cheaper to have.

The upside of the same asymmetry: the first legal systems to make agent commerce safe — identifiable, insurable, adjudicable — capture the institutional franchise, as Delaware did for the corporation and English law did for global contracting.

PIIE Hufbauer & Lu 2017 NBER Frankel & Romer 1999 JEP North 1991 Haggard & Tiede 2011

III.4 The Agent-Native Economy

Everything so far amplifies mechanisms that already exist. The final layer is new in kind: agents transacting with each other, at machine speed, in volumes and granularities no human economy could sustain — what Google DeepMind names the sandbox economy, a layer where agents transact and coordinate “at scales and speeds beyond direct human oversight.”

DeepMind’s central warning is a coordination-failure diagnosis, and it is worth quoting: “our current trajectory points toward a spontaneous emergence of a vast and highly permeable AI agent economy.” Permeability, the paper stresses, is a collective property no single actor controls — lowering it requires solving a collective-action problem, which is why drift, not design, is the default. Their agenda follows: auction mechanisms for fair resource allocation among agents, “mission economies” that point agent swarms at collective goals, and the trust infrastructure of verifiable credentials.

Google DeepMind Tomašev et al. 2025 · verified verbatim

Machine-native money, and value beyond human bandwidth

The plumbing for the agent layer is being built now. The x402 protocol — initiated by Coinbase — resurrects the long-reserved HTTP 402 “Payment Required” status code as an open standard that “enables AI agents and web services to autonomously pay for API access, data, and digital services”: stablecoin settlement in ~200 milliseconds, no API keys, no subscriptions, no human checkout. The economics hiding in that engineering are large. Legacy rails carry per-transaction costs and chargeback risk that make anything below roughly a dollar uneconomic — which is why the internet’s business models collapsed into advertising and subscriptions. Machine-native micropayments reopen the design space Microsoft’s authors call unbundling and rebundling: paying per article, per API call, per second of compute, per answer. A million transactions of a tenth of a cent is not a smaller version of a thousand transactions of a dollar; it is a different economy, with different market structures, operating below the threshold of human attention.

What lives in that economy? Transactions that were always worth making and never worth a human’s time:

• Procurement: an agent collects and compares 10,000 supplier quotes before lunch — a task no purchasing department ever priced, because no purchasing department could do it.
• Data: per-query licensing of a dataset — a market that cannot exist at subscription granularity, priced per question at machine speed.
• Insurance: per-hour cover on a single shipment, underwritten and settled in minutes — the BIS’s weeks-to-minutes collapse applied to a contract too small to exist today.
• The dormant surplus: every unswitched tariff, unclaimed refund, unexploited price difference and unread contract clause — executed instead of forgone.

A human consumer executes perhaps a few thousand economic decisions a year; almost everything else is surplus left on the table because attention is the scarcest input in the consumer economy. Agents convert that dormant decision-space into executed transactions. The aggregate is not a productivity statistic; it is the gap between the demand curve as economists draw it and the demand curve as exhausted humans actually walk it. None of it appears in a Hulten-theorem calculation — and none of it is governed yet, which is why the quadrant chart opposite comes first.

x402 whitepaper Reppel et al. 2025 · verified verbatim Microsoft Research unbundling · 2025 BIS WP 1194

The conditional dividend: $10–20 trillion a year, assembled in the open

The headline number of this movement is built from four channels, each shown with its arithmetic and its label. It is a sized conditional, not a forecast — and it deliberately excludes the agent-to-agent layer, which is real but unpriceable today.

III.5 The Nordhaus Question for AI

Will agentic producers keep 2 per cent of the surplus — or 40? The whole report turns on Nordhaus’s 2.2 per cent, so the honest question is whether the agentic century repeats it. The answer starts with an uncomfortable fact: the 2.2 per cent was never a law of nature.

The guardrail conditions — under which the multiplier keeps its sign

None of these is exotic. Each is ordinary policy of the kind that produced the original 2.2 per cent — applied to the agent layer while its equilibrium is still unformed.

1. Open-agent access. Any offer published to humans must be lawfully accessible to their agents on non-discriminatory terms. One clause, aimed squarely at the walled-garden equilibrium — the difference between the open web of agents and the app store, written into law.
2. Model-layer concentration monitoring. Treat the supply of frontier agents the way supervisors treat systemically important banks: measured market shares, dependency mapping, published concentration telemetry — before the chokepoint forms, not after.
3. Interoperable rails. Machine-native payments (x402-class and their regulated-currency equivalents), agent-readable public data, and portability of context and contracts — so machine commerce is not forced through proprietary gardens by default.
4. Agent identity and liability law. Verifiable credentials and a clear human-accountability chain for every transacting agent — the missing layer DeepMind and Microsoft both name, and the institutional franchise awaiting its first mover.

The stewardship framing closes the movement. The machine of Movement II hands society ninety-eight per cent of what its winners create — not out of benevolence, but because someone maintained the competition that forces the donation. The multiplier of Movement III obeys the same rule at higher stakes: it multiplies whatever structure it is poured into. Poured into open, competitive, well-governed markets, it multiplies consumer surplus — the Kerala result at civilisational scale. Poured into concentrated, closed, under-governed ones, it multiplies rents and correlated fragility. The technology does not choose; the institutions do.