The AI Renaissance: the price of cognition and the economics of the transition

Kennard R. Laviers, PhD (Computer Science), and Maria Siakola

Abstract

AI systems now contribute to the research that improves AI systems, which makes the rate of innovation endogenous. We build a small, fully specified simulation of the economic consequences from 2025 to 2050, in which algorithmic progress is produced by human and AI research labor and experiment compute with diminishing returns, capability is effective frontier compute, substitutability is anchored to the length of task an agent can complete, adoption lags capability, compute and power are built through lagged pipelines, households face a budget, labor markets have job finding, scarring and exit, and politics reacts to a composite measure of stress. Parameters are anchored to four evidence dossiers (about 150 primary sources), and the early path is fitted to the 2025-26 record: the METR agent horizon doubling every four months, most new code at frontier labs written by AI, 55,000 AI-cited job cuts in 2025 and 120,000 in the first nine months of 2026, and young workers in exposed occupations 19% below their peers through foregone hiring. That fit reproduces 25 of 25 anchors for the years it was fitted to; it is calibration, not out-of-sample validation, and the anchors are marked as measured, proxy or constructed. The 2019-2025 record is shown on every chart and used as a hindcast; three parameters on which it is informative (frontier concentration, the timing and horizon of capital commitments) are set to it, and the vendor-revenue anchor is an accounting bridge of what households and businesses actually paid (§4.7). Extrapolated through the model's own brakes, the 2025-26 pace cuts the compute cost of a cognitive task by 8.0 orders of magnitude by 2050 and the buyer's all-in cost, review labor included, by 114x (to 0.9% of its 2025 level); puts capability 5 years ahead of adoption rather than a decade; keeps capex growing through 2033 with power binding until 2034 and the cut landing in 2034; and brings the labor shock forward to the early 2030s: AI-attributed separations near 3.4 million a year, a displaced pool that peaks at 8.0% of the workforce in 2031 on the cognitive wave with the modeled political response and 11.4% in 2035 without it, and 72 million people who see more than half of their trade's tasks automated at the 2032 peak. Physical work absorbs that first wave, and robots, once general-purpose machines undercut the physical wage per delivered task-hour from 2029, take it back in the 2040s: physical employment falls from 60% to 53% of the workforce and the displaced pool is 9.9% and still rising in 2050. Across 1,000 joint draws of the uncertain parameters the timing of the cognitive wave is robust and the size of the robot wave is not, and the aggregate real-income gain rests on a structural prior rather than on the record. The gain is in cognition and in what cognition makes cheaper; it is not yet in essentials. Priced by housing, food, electricity and healthcare rather than by the model's single physical index, aggregate household purchasing power in 2050 is 1.38x its 2025 level rather than 2.8x, the middle quintile's disposable income is 1.31x and the bottom quintile's 1.21x, worse off than in 2025 (changes from 2025 on the baseline, not gains against a no-AI path; across the sector cost shares the bottom quintile reads 1.00x to 1.61x and the middle 1.08x to 1.75x; across the joint draws the bottom quintile is below its 2025 level in 15% of runs). The skill premium written down by 2050, 77% of one year's 2025 GDP, is 7 times the compute written down (10.3%). What becomes possible is on the same scale as the disruption: agents that complete tasks which take people a work-month at 2027 on the measured trend (an extrapolation of a benchmark that stops at sixteen hours), the compute cost of a cognitive task at 2.7 × 10⁻¹⁰ of the 2025 human wage (9.4 × 10⁻⁹ of its own 2025 level) and the buyer's all-in cost at 2.2 × 10⁻³ of the wage, the price of everything with thought in it falling, and a research engine in which AI agents are most of the labor by 2027. Who shares in it depends on five institutional outcomes rather than on capability: whether AI rents are competed away, who pays for the grid, how fast and how well displaced workers are re-employed, how the accountability floor is set, and whether entrants get a route in. We are in an AI renaissance. The model says it is real in the price of cognition and in output, says that on the baseline it has not reached the essentials of the median household by 2050, and says what reaching them is conditional on.

1. Introduction

The debate about AI and the economy is usually conducted in two currencies that do not convert: capability benchmarks on one side, jobs on the other. This paper uses a third. The thing that is actually happening is a change in the price of cognitive output of a given quality, which is falling toward the cost of electricity plus amortised capital. That has happened before, to light, to computation and to communication, and in each case the price fell by orders of magnitude, the people who sold the old good did not get rich, the good became ambient, and the economy reorganised around its abundance (§3). Jobs follow prices, with friction and lag, and the lag is where the politics lives.

Three claims from the exchange that began this work survive the model intact. First, AI-driven improvement of AI is already a present-day force and its economic consequences begin at that stage, not at superintelligence. Second, the binding constraints are physical and slow: fabrication, power, interconnection, land, and the human institutions of liability and licensing. Third, the dangerous period is not when AI becomes vastly more capable than people but when human cognitive labor loses economic value faster than society builds other ways for people to participate. Two claims did not survive. The energy constraint does not bite on capability the way we expected, because algorithmic progress dominates and efficiency gains absorb most of the load; it bites on timing and on who pays. And "the basket swings to physical goods" is an accounting error: cheap cognition is also an input to physical goods, and physical prices fall.

Our contribution is a model simple enough to read in an afternoon and specified enough to refute. Section 2 summarises what the 2023-2026 evidence already shows. Section 3 gives the historical base rates we anchor to. Section 4 is the model, block by block. Section 5 is calibration. Section 6 reports results as five mechanisms, each with the observation that would refute it, and a scenario grid. Section 7 maps the aggregates onto occupations, education and households, including the value of trades that are devalued. Section 8 states what becomes possible, in the model's own units. Section 9 states the conditions under which the transition is a renaissance rather than a transfer. Section 10 is limitations. The claims register in Appendix A lists every quantitative claim with its status.

2. What the evidence already shows, 2023-2026

This section summarises the jobs, education and lifestyle evidence dossiers, which carry the full citations (Appendix B).

Exposure is broad and usage is narrow. About 80% of US workers have at least a tenth of their tasks exposed to language models and about a fifth have half or more (Eloundou et al.), but observed usage reaches three-quarters of tasks in only about 4% of occupations and a third of workers show no coverage at all (Anthropic Economic Index, 2026). Firm adoption rose from under 4% of firms in 2023 to about 10% on the narrow Census definition and 17-20% on the broad one by 2026; workers using AI at work rose from 28% to 45%; hours actually assisted are 1% to 5%. Logistic fits to these series imply ten to fourteen years from 5% to 95% adoption.

Measured aggregate effects are small so far. Danish administrative data put the effect on earnings and hours within one to two percent of zero (Humlum and Vestergaard). The one robust displacement signal is at the entry margin: employment of 22-to-25-year-olds in the most exposed occupations is 13% (2025) to 19% (2026) below trend, through reduced hiring rather than separations (Brynjolfsson, Chandar and Chen; ADP). Entry-level postings in software fell from 43% to 28% of the total; large-firm new-graduate hiring fell by half; first-year law classes shrank by a sixth in a cycle. AI-attributed layoffs reached about 116,000 in the first eight months of 2026, a fifth of announced cuts and 0.07% of employment.

The 2025-26 record, summarised in the dossiers on AI coding and on layoffs, is faster than the series above. The METR 50% time horizon went from 2 h (o3, April 2025) to 5 h 20 (Opus 4.5, November 2025), about 12 h (Opus 4.6, February 2026) and at least 16 h (Mythos Preview, March 2026), a doubling every 4.3 months; METR's suite saturates at 16 h and the 80% horizon runs about five times behind. SWE-bench Verified went from 49% (October 2024) to 81% (November 2025) and 94% (April 2026). The share of new code written by AI at Google rose from 30% (April 2025) to 75% (April 2026); over 80% of Anthropic's merged production code was written by its models in May 2026; across 15,000 developers about 47% of code is fully agent-written and 90% use agents weekly (JetBrains, August 2026); one in three pull requests on GitHub involves an agent. Agentic coding went from nothing to several billion dollars of run-rate revenue in a year. Fixed-capability inference prices fell about forty-fold a year while frontier list prices stayed near $5 per million input tokens, so a resolved software task costs $0.50 to $25 against $500 to $800 of engineer time, and review became the bottleneck. METR's own randomised trial (July 2025) found experienced developers 19% slower with AI; METR declared it outdated in February 2026 and judges speedups likely but unmeasured.

Layoffs attributed to AI have become measurable. Challenger counted 54,836 AI-cited cuts in 2025 (4.5% of announced cuts, with another 20,000 coded "technological update") and 120,136 in January-September 2026 (21%, the top stated reason, leading every month from March to July); the series is stepped, with 31,000 in October 2025 (Amazon, Salesforce, Accenture) and 100,000 in March-July 2026 (Block, Oracle, Meta), then a fade to about 4,000 a month. Company-stated cases include Salesforce's support staff cut from 9,000 to 5,000, Accenture's 11,000 exits of staff who could not be reskilled, Oracle's 21,000 with AI named in its annual report, Block's 4,000 and Amazon's 30,000 across two rounds. Oxford Economics, Yale's Budget Lab and Forrester treat a large part of this as AI-washing of cost-cutting; Goldman's net estimate is 16,000 jobs a month (25,000 substituted, 9,000 created). Payrolls show sector erosion rather than economy-wide displacement: computer systems design down 5% from its 2023 peak, telephone call centres down 17% since January 2024, business support services down 13%, while professional services as a whole are flat to up. The measured effect is at entry: the 22-to-25 gap in exposed occupations widened from 13% to 19%, entry-level roles are 4.5% of software postings, recent-graduate unemployment is 5.6% against 4.2% overall and 7-8% for computing majors. Outside the US, India's top IT firms went from adding 50,000 a year to a net loss, UK graduate postings fell 46% in a year, and the Philippines cut its 2028 outsourcing headcount target by a fifth.

The model's early path is fitted to this record (section 5). The table compares observed and model values, with each anchor marked as measured (a published series), proxy (a related series stands in for the model quantity) or constructed (a judgement from several sources). It is a fit to these years, not a validation on years the model did not see:

Prices and capital are moving fast. The price of a fixed-capability inference token has fallen about forty-fold a year; frontier price per token two- to three-fold. Hyperscaler AI capital spending grew 55% to 65% a year in 2023-2025 to about $400 billion; installed AI compute roughly doubled each year; frontier training-run cost grew 2.4x a year. Power is the binding constraint at the point of interconnection: utility queues run two to five years, new generation three to seven, gas turbines are sold out to 2029, and data-centre load is already visible in capacity prices and in household rates in some regions, though the pass-through is contested.

The education pipeline is already responding, with a one-to-two-year lag. New computer science majors fell 13% in 2025, the first reversal since 2008; humanities are at a 34-year low; health, trades and apprenticeships are growing; law applications are up 27% in two years; coding bootcamps have largely closed. Recent-graduate unemployment now exceeds the all-worker rate. Licensing bodies have said explicitly that accountability stays with the human licensee, which is the institutional basis of what we call the accountability floor.

Households spend most of their income on things that stay scarce: housing a third, transport a sixth, food an eighth, health a twelfth. Over 2000-2024 hospital services rose 256% and tuition 187% against an 87% rise in the price level, while televisions fell 98%. Cross-sector substitution elasticities are low (about 0.26). Equity ownership is concentrated: the top decile holds 87% of equities, the bottom half 1%. Cash-transfer experiments cut hours by about one a week per $1,000 a month and leave participation nearly unchanged.

3. Historical base rates

This section summarises the history evidence dossier.

Producer-side general-purpose technologies that required reorganising production took twenty to forty years to reach most of the relevant capital stock: factory electrification went from under 5% of mechanical drive in 1899 to 78% in 1929; tractors took about forty years; enterprise cloud was at 45% of EU firms seventeen years after it launched. Consumer technologies on an existing platform diffuse in fifteen to twenty-five years. The capability-to-productivity lag was about forty years for electrification and a decade for information technology.

Price collapses of the kind we model are precedented: the true price of light fell about 3,460-fold between 1800 and 1992 and official indexes missed the decline by a factor of 900 to 1,600; computing fell by twelve orders of magnitude; genome sequencing by five orders in two decades. The occupations that sold the old good disappeared, their incumbents were worse off a decade later, and later cohorts re-routed.

No national displacement episode exceeded about one percent of the workforce a year: agriculture at most 0.65 points a year, manufacturing 0.4 to 0.5 percent a year in 2000-2010. Individual re-employment is fast, two-thirds within one to three years, with persistent earnings losses of a quarter; regional reabsorption after a concentrated shock is close to zero for ten to twenty years, with three-quarters of the adjustment into non-participation. Political responses fired at very different incidences: repression at localised collapses, redistribution at 25% national unemployment, electoral realignment at about 1% national or 4.5% local manufacturing loss, fifteen years after onset. Capital booms around general-purpose technologies overshot by a third to 97% of capacity and the assets were used within about a decade by different owners.

4. The model

One annual loop, 2025 to 2050, indices equal to one in 2025 unless stated. The equation companion (Appendix C) lists every equation and parameter; here we give the structure and the choices that matter.

4.1 Physical layer

Compute is a vintage stock. Each year's additions form a vintage with its own quantity, book value and straight-line depreciation over five years; quantity wears at 10% a year; the carrying cost of the stock is depreciation plus operating cost plus electricity, so a 2026 cluster costs what it cost rather than today's chip price. Vintages are impaired when value in use and fair value fall below book, and the write-downs accumulate. Fabrication caps annual additions.

Power is firm capacity built through a commitment pipeline. Each year utilities and operators forecast need at three horizons, capped by the conservatism of integrated resource plans, and commit the gap through grid interconnection (five-year lag, 60% of the mix), behind-the-meter gas (two years, 30%) and retrofit (one year, 10%), at different unit costs; capacity arrives when the lag elapses, subject to a buildout cap. Effective compute is the smaller of installed compute and what can be powered; efficiency improves and saturates. Electricity carries a scarcity markup when headroom is short, grid capital is socialised into household rates, and power is drawn only by compute that is running. Climate is a switch, off in the baseline: a drift on physical unit cost and on the electricity price, an adaptation-investment term that avoids damage at a benefit-cost ratio of six and a regional preset, reported in §6.4. Commitments made before 2025 and arriving in 2026-27 are seeded from the record, since a pipeline that started empty could not power the 2026 additions the record shows being built; the advanced-packaging bound on compute additions grows at the CoWoS record, +65% in the first year decaying to +40% a year.

4.2 Progress and the recursion

Algorithmic efficiency $R$ grows at a rate produced by research effort with diminishing returns to the idea stock (the full equation set, with notation, is Appendix C):

$L = H_r + s_{\mathrm{AI}} L_{\mathrm{AI}}$ is research labor, the human research workforce $H_r$ (1 in 2025, falling to the frontier handful, §4.4) plus AI agents whose count scales with the compute allocated to research and with efficiency, and whose quality rises with capability; E combines labor with experiment compute through a CES with elasticity one half, so labor cannot substitute for experiments without limit; phi is the fishing-out exponent; kappa is calibrated so progress is 0.9 log units a year in 2025. Capability $A$ is effective frontier compute, with frontier training compute its own state growing 3.5-fold a year near term and decaying toward stock growth. The share of capability growth from algorithms is emergent and starts near 0.4.

The model carries three distinct recursive-improvement mechanisms, rather than parameter points of one production function, as a switch (research_regime). Diminishing returns is the equation above: more AI researchers produce progressively smaller gains through fishing-out. Sustained recursive growth lets AI-attributed progress raise subsequent research productivity, $P_{t+1} = P_t \exp(\gamma\,\text{recursion}_t)$ with $d\ln R / dt$ scaled by $P_t$, up to an effective-limits cap: the improvements compound into the research process itself, which is what AlphaEvolve first measured when AI-discovered kernels cut the training time of the model that discovered them. Breakthrough-driven progress draws occasional discontinuous gains from a hazard that rises with research effort; each gain is lognormal in size, lands after a two-year deployment lag and ramps over three years, and hits algorithms, hardware cost or energy efficiency, so that discovery, feasibility, deployment and diffusion are separate steps on separate timescales. It is stochastic, and the paper reports an ensemble of seeds (§6.7). The scenario grid's progress axis runs from the compute-bound parameterisation of diminishing returns to the sustained mechanism.

Effective compute is allocated by shares that sum to one: the frontier run's share grows with the race from 4% in 2025 to a cap of 30%, reached in 2029, and takes its share from inference, which falls to 40% of effective compute by 2030 and is why the baseline is capacity-short from 2028; research agents hold 10% and experiments 20%. Agents scale with the compute allocated to them, so a run with no research compute has no AI research labor. Both channels are measured in units of the 2025 compute at a reference allocation fixed across scenarios, with the research productivity constant calibrated there; normalising each channel by its own scenario's 2025 allocation instead would leave progress unchanged when a fixed budget moves between researchers and experiments. Moving the same 30% budget to 1% researchers and 29% experiments changes 2040 efficiency by +77%, and to 29% researchers and 1% experiments by -98%: agents are abundant and experiments bind, so the CES rewards the experiment-heavy allocation, while a run with no agents at all collapses. One efficiency index enters capability, inference cost and agent productivity through separate exponents, so the channels can be varied apart. The AI-attributable part of progress is a counterfactual, the difference between progress with and without the agents, rather than their share of research labor: 79% of algorithmic progress in 2030 and 99% in 2040.

4.3 Substitutability, cost and diffusion

The length of task an agent completes at 50% success is anchored to the METR series through March 2026 (2 h in April 2025, 5 h 20 in November, about 12 h in February 2026, 16 h or more in March; doubling every four months); the 80% horizon is a fifth of it and is what substitutability $\tau$ is anchored to, since it is the better proxy for work that runs unsupervised:

with $h_0 = 3.5$ h, $k_h = 1.9$ (about 27 h in 2026, 172 h in 2027, 1,026 h in 2028), $r_{80} = 0.2$, $h_{\mathrm{mid}} = 60$ h (half of cognitive tasks substitutable when agents complete 1.5 reliable work-weeks unsupervised), $w = 2$, fitted so that software's fully-automated share is near 10% in 2025 and 25% in 2026; $f$ is the accountability floor. METR's own caution applies (Kwa, January 2026): the horizon measures the human-work length of tasks completed, not autonomous operating duration. The model therefore carries three capabilities rather than one horizon: task competence, the sigmoid on the 50% horizon, what an agent can do with a person in the loop (reported; the natural driver of AI-services uptake, which stays on reliability so that the defaults reproduce the fitted baseline); autonomous reliability, the sigmoid on the 80% horizon times an explicit discount $\delta$ for unsupervised long-duration operation, what runs unattended, which is what substitutability follows; and research autonomy, the same unattended horizon against a research midpoint $\kappa_r$ times the economy's, which sets the research tasks AI can take and discounts agents' research labor in the recursion while it lags task autonomy (§4.2, §4.4). The baseline sets $\delta = 1$ (the 80% ratio is the whole discount) and $\kappa_r = 1$, so one horizon is recovered; both carry uncertainty priors ($\delta$ 0.3-1, $\kappa_r$ 1-10, Appendix D), the autonomy_discounted ($\delta = 0.5$) and research_long_horizon ($\kappa_r = 5$) presets show what they move, and the readings in section 8 are conditional on them. What these variables do not do is identify real-world reliability: with $\delta = 1$ and $\kappa_r = 1$ research autonomy coincides with the task mapping, and multiplying a benchmark horizon by a scalar is a scenario assumption, not a measurement of unattended operation. METR's series is built largely on software, machine-learning and cybersecurity tasks, does not imply that every job of the same length is automatable, marks measurements above sixteen hours as unreliable, and has not been maintained since September 2026. The 60-hour midpoint is therefore a central sensitivity, not a fact read off the benchmark: at a 30-hour midpoint AI performs half the 2025 stock in 2032 and at 240 hours in 2033, the cognitive-wave pool peaks at 8.6% and 8.2%, cognitive employment in 2035 is 31% and 32%, and the 2050 labor share 0.57 and 0.57 (the table below; the required success rate, the autonomy discount and the research ratio are varied beside it; domain heterogeneity is the occupation layer of §7.1, with its own midpoints per group). The dates move by a year or two and the levels by a few points because the record of 2025-26 pins the early path whatever the midpoint; what the midpoint governs is how far the substitutable share eventually reaches and when the floor binds.

On the baseline competence is 40% and autonomous reliability 23% in 2026, 81% and 65% in 2028 (research autonomy 65%):

With half the 80% horizon running unattended, half of cognitive tasks become substitutable in 2029 (baseline 2028), the pool peaks at 9.8% in 2050, cognitive employment in 2050 is 28% and physical-goods real income 2.78x. With research programmes five times the median task, agents' research labor is discounted while research autonomy lags task autonomy: the AI share of research labor is 95% in 2028 (baseline 97%) and 100% in 2030 (baseline 100%), and algorithmic efficiency is 0.90x of the baseline's in 2030 and 1.04x in 2040. The lag bites for a few years in the late 2020s and then closes, because the horizon doubles every four months and a five-fold research midpoint is crossed within a year; and the human research workforce follows adoption rather than autonomy, so the handful is reached in 2046 either way (baseline 2046). What would hold research with people longer is a slower horizon, not a longer research task. The vendor's unit cost of AI output is the carrying cost of running compute over effective compute, 0.028 of the human wage per task in 2025 across the two tiers (the vendor price of a tenth adds the gross margin and the frontier tier's rent, below). The review, rework and accountability time that stays with each AI task is not vendor cost but the buyer's own labor, $\rho_t$ human task-hours per AI task (0.15 in 2025, falling 15% a year with reliability) paid at the human wage, counted in human cognitive demand, employment and wages; booked as vendor revenue, it would be income with no recipient. The buyer's all-in cost per AI task is the vendor price plus $\rho_t w_h$, a quarter of the wage in 2025, which is what the 2025 anchor measures. Cost is not price. Vendors price at cost plus a gross margin, the pricing power that funds compute at scale: the labs reported gross margins of a third to a half in 2025 (AI revenue dossier), so the 2025 price of a tenth of the wage is compute cost of 0.028 (the mix of the two tiers below) plus the margin and the frontier tier's rent, and the margin erodes toward 27% by 2050 as open weights and competition catch up. Above that floor the price adds a slice of the gap to the human alternative net of review labor, taken by concentrated vendors as rent (which fades as open weights catch up) or by scarcity when last year's inference demand exceeded powered supply:

The scarcity share is revenue to compute owners and removes the cost pressure that drives diffusion while it lasts. The buyer's all-in cost has two parts: the vendor price, which falls with efficiency and hardware by orders of magnitude over the horizon, and the buyer's own review, rework and accountability labor on each delivered task, 0.15 task-hours per AI task in 2025 and falling 15% a year with reliability, so the all-in cost falls two orders of magnitude, to 3.0 × 10⁻² of the 2025 wage in 2035 and 2.2 × 10⁻³ in 2050, with review labor 3.0 × 10⁻² and 2.6 × 10⁻³ task-hours per AI task; that labor is 40% of human cognitive employment in 2035 and 56% in 2050 (§6.6). Delivered AI volume is defined once and used for inference, revenue, energy and capacity: it is the AI share of the delivered cognitive bundle, capped in the same year by powered inference capacity; a shortfall cuts induced volume first and then returns 2025-stock tasks to humans at the wage. Transactions reconcile inside the fixed point, so that no sales are booked for volume the cap removes and no wages or AI revenue are paid out of money that buys nothing: humans are paid the wage for their tasks; the planned budget buys the delivered AI quantity at the posted price times a clearing premium, bounded by the human wage because nobody pays a machine more than a person; whatever the budget could not buy at that cap is recorded as unspent and saved. Sales equal price times delivered volume in every year of every scenario (a validation identity, Appendix E), and the premium is the rationing signal: in the baseline it reaches 5.4x in 2029 and is gone by 2031, with demand at the posted price exceeding delivery by up to 8%. AI services (§4.7) are served from the same powered pool as tasks, and the two clear against it jointly. Capping tasks alone would let a services budget at a collapsing price ask for more than the pool holds, so that tasks would be delivered at zero while services were booked above capacity, most of all in the competitive-pricing case. Instead each demands a volume at its base price, and when the two exceed capacity one clearing premium rations both, each price bounded by the human alternative; at the bounds the shortfall is rationed pro rata and the budget that buys nothing is saved. Services revenue is booked on delivered volume only. Delivered tasks plus services never exceed capacity in any preset, grid case or draw (a validation identity, Appendix E); on the baseline the joint constraint is slack because its services never reach the pool. The review requirement also binds delivered AI output when reviewers cannot be staffed: AI output that nobody is there to review, rework and sign for is not delivered but lost (ai_unreviewed_lost). Review is staffed first out of the matched hours, so the bound never binds on the baseline (largest shortfall 7 × 10⁻¹⁵ of a volume unit); it binds in 2 of the paper's presets (ace_plus_capability, capital_markets_off), and in the overbuilt no-capital-markets case it is what keeps the labor share at 0.57 in 2050 rather than a quarter: the stock can produce far more than people can review. In the scarce-intelligence scenario demand at the posted price exceeds delivery by up to 41%. The number of frontier actors falls with the ratio of run cost to budgets at an elasticity of 0.2, set by the 2019-26 record (two labs to five-seven while run cost grew 2.4x a year), so the frontier consolidates from four actors in 2026 to two in 2035 rather than collapsing to one by 2030.

The market has two tiers: frontier reasoning, the tasks only the frontier can do yet, $(\tau - \tau_{\mathrm{open}})/\tau$ of AI volume, served by the concentrated frontier at cost-plus on $k_f$ times the compute per task (reasoning models spend ten to a hundred times the tokens; $k_f = 4$, a prior of 1-10) with rent on concentration and its own scarcity on its reserved capacity; and commodity inference, the rest, competitive and cost-plus with its own scarcity on the rest of the powered pool. The posted price is the volume-weighted price of the tiers; the multiple splits the inference need calibrated to the 2025 record between the tiers (it does not scale the total, which follows delivered volume and efficiency) and sets their relative costs; and the frontier's capacity share adjusts to its need share with a lag. In 2025 the frontier tier is 56% of AI volume at 0.164 of the wage against 0.018 for commodity inference; it peaks at 58% in 2026 and is 0% by 2035 as open weights catch up, so the frontier price runs 177.5x the commodity price in 2028 and 5.9x in 2030, and the frontier tier's share of the tasks' compute need falls from 34% in 2025 to 12% in 2030. In scarce intelligence the frontier tier's scarcity in 2030 is 38% against 62% on the commodity tier, and the tiers are priced at 0.47 and 0.55 of the wage; with frontier reasoning at ten times the compute (frontier_heavy) AI does half of cognitive work by volume in 2030 (baseline 2030) and the pool peaks at 10.0%.

Diffusion of AI into the 2025 task stock, $D \le \tau$, is logistic toward the ceiling, gated by cost advantage and slowed by political friction $\varphi$:

with a lag of 4.4 years from 5% to 95% at full pressure (the 2025-26 pace of the margin that substitutes for labour; the extensive margin's 10-14 years is a scenario).

4.4 Labor

Human task demand is the non-automated share of the 2025 task stock times trend volume $T_{\mathrm{trend}}$ (served by humans one for one, as in an economy without AI) times the AI-induced multiplier $T / T_{\mathrm{trend}}$ to an effective pass-through power $\psi_{\mathrm{eff}}$; below trend both sides shrink in proportion (serving the whole shortfall from humans would send every adopted stock task back to them at a 6% shortfall below trend, so that adoption would not be production in the frozen-capability case, §6.8); the AI share by volume counts the induced volume:

with $\psi = 0.4$, $\kappa_c = 1$, $\eta_w = 0.35$ and $T_{\mathrm{trend}} = 1.02^{t}$. $\gamma$ is the capability gap already closed: the share of the automatable cognitive stock AI can do, which is the capability sigmoid itself. The pass-through is not a fixed complementarity. Induced tasks are drawn from the same capability distribution as the existing stock, so once AI does a task better than people the induced volume of that task goes to AI as well; the complementarity that lets humans keep a slice of every new task decays as the gap closes, and human task demand converges to the accountability floor of trend volume, the work institutions insist a person signs for. In the baseline $\gamma$ is 89% in 2030 and $\psi_{\mathrm{eff}}$ is 0.05; by 2035 it is 0.00. This replaces one assumption with another, and an uncertainty prior on the decay alone, which excludes zero, would make the bands conditional on substantial decay. The paper therefore carries three structural families of human comparative advantage, each with its own parameter uncertainty (complementarity_family): declining, the equation above (baseline); sustained, in which humans keep $\psi$ of every induced task however capable AI becomes (fixed_complementarity); and created, in which AI creates new human tasks as it covers the old ones, $\psi_{\mathrm{eff}} = \min(1, \psi(1 + \kappa_n \gamma))$ with $\kappa_n = 0.5$ (created_complementarity; Autor's reinstatement). The three are compared in §6.5 and §6.6, with the observations that would tell them apart by the early 2030s. The premise of the baseline, that future tasks inherit the automation characteristics of today's, is a structural choice and is reported as one. At $\psi = 1$ every unit of induced volume needs humans in proportion while the gap is open, the optimist's case; at $0$ induced volume is all AI from the start. The damping applies to AI-induced volume, not to trend growth: applied to trend growth as well, it would attribute a quarter of the 2050 volume of an economy with no AI at all to AI. The pass-through takes the place of a new-task rate, which is not identified, and of a task-substitution elasticity, which is degenerate once AI is nearly free.

Research and engineering are treated explicitly because they are where the capability argument bites first and hardest. The field is 9.5 million people in 2025 (computer and mathematical, architecture and engineering, life, physical and social science; 5.7% of the workforce). Its tasks follow the same capability sigmoid with the accountability floor replaced by the frontier handful $\eta_r$, the share of the 2025 field that stays whatever AI can do, the people who set direction and answer for what gets built:

$L_r$ is direct research and engineering work. One definition serves both the aggregate series and the occupation layer: research and engineering employment is direct research plus review labor on AI-produced research and engineering output (the field's share of the 2025 stock times total review employment, §4.3), and the three groups in §7.1 sum to it. $H_r$ is also the human research labor in the recursion (§4.2), so the AI share of research labor, $s_{\mathrm{AI}} L_{\mathrm{AI}} / (H_r + s_{\mathrm{AI}} L_{\mathrm{AI}})$, rather than the raw AI research-labor index, is the quantity the paper reports: the index runs to billions by 2050 and says nothing a reader can use, while the share answers the question asked. Labor is conserved across cognitive employment, physical employment, a two-compartment displaced pool and exits. Displacement is gross, not net: besides any fall in cognitive employment, a share $\theta$ of the task volume newly automated each year is shed as headcount while aggregate employment holds, the rest being absorbed as within-job productivity and output expansion; of the gross, a share $e$ falls on entrants as foregone hiring, capped by the annual entry cohort, and the remainder are incumbent separations, the series that layoff trackers count:

with $\theta = 0.3$, $e = 0.7$, $c_{\mathrm{entry}} = 0.03$. Both shares are fitted to 2025-26 (Challenger's counts, Goldman's net estimate, the Stanford/ADP young-worker gap).

The labor market is an explicit transition matrix over five states, cognitive employed, physical employed, short-term pool, long-term pool and out of the workforce, and the reported stocks are end-of-year and partition the workforce on every row. Each year: gross cognitive job destruction and physical separations move people into the short-term pool; vacancies are demand above the post-separation stock; hires come only from the pools, at the job-finding rate (short-term first, long-term at half that rate), never beyond vacancies and never beyond the people available, with unfilled vacancies reported; the short-term pool leaks into the long-term pool and the long-term pool into non-participation. The pool is tracked by origin sector so that each household archetype has its own employment probability. The short-term pool finds jobs at 0.4 a year, faster with active measures and slower when the pool is large; the long-term pool finds jobs at half that rate and leaks into non-participation; re-employed workers carry a 30% wage discount that closes at 12% a year.

Timing and the joint closure of employment, output and income. The timing convention within a year is explicit and the matching sits inside the output and income fixed point, so that an unfilled job neither produces output nor pays a wage: labor demand from the current iterate sets separations and vacancies; hiring from beginning-of-year stocks fills them at the job-finding rate; production and income are computed from the matched stocks. Hours per worker can stretch by up to 5% to cover unfilled vacancies (average weekly hours move a few per cent over the cycle), and the matched share of demanded cognitive hours, $\mathrm{fill} = \min(1,\, C\,(1+\phi_h)/C^d)$, scales the human tasks done; wages are paid for hours done, not for hours demanded, and the human tasks that could not be staffed are recorded (cognitive_unfilled_value) and not produced. The planned cognitive budget clears against the delivered volume as in §4.3, so sales equal price times delivered volume in every year. Physical capacity is the matched physical workforce, hours-stretched toward demand, plus robots. In the baseline hiring keeps up: hours stretch at most 0.2% (in 2025) and the fill stays at 100%, as it does in all 59 scenarios this paper reports (minimum 49%). The fill binds only in diagnostic presets in which cognitive demand outruns the workforce: across all 67 presets it falls to 49% (capital_markets_off, 2050), where the unstaffed human demand reaches 26.3% of GDP and is neither produced nor paid. The physical side binds more often: with capacity tied to matched workers rather than to everyone available, physical demand exceeds what can be produced by up to 2.1% of GDP in 2034 on the baseline, cleared by the price premium of §4.5. The reconciliation is tested as identities (Appendix E).

The entry margin is a decomposition under a stationary workforce. Each year a cohort equal to retirements enters; the cohort that would have replaced retiring cognitive workers is hired into their posts when demand holds, and when it does not the posts go unfilled and the entrant joins the pool. Foregone entrant hiring is therefore retirements not replaced: it reduces the employed stock by exactly that number, which is why the whole gross flow leaves employment while only the incumbent part is a layoff. At the 2030 peak of the wave none of the 2.0 million cohort is hired and incumbent separations are 63% of gross destruction; by 2040 845 thousand of a 1.5 million cohort are hired and 699 thousand posts go unfilled. The three series (entrant_cohort, entrants_hired, retirements_unreplaced) are reported and the stock law of motion is a validation identity.

4.5 Households and the demand closure

Households consume out of labor income, transfers and the payout of capital income with propensities 0.85, 0.6 and 0.35; budget shares between cognitive and physical goods add up, with a variety term so that a price collapse does not mechanically shrink the cognitive share; government purchases are real and track potential; non-AI investment is last year's saving. Physical output is demand-determined and rationed at potential with a price premium; cognitive gross output is final demand plus the cognitive-input share of physical costs. GDP equals consumption plus government plus investment by construction and is tested every year. The cognitive price, cognitive volume and the physical price are

with $a_c = 0.25$ the cognitive-input share of physical costs. Real income is reported three ways: nominal in the 2025 wage numeraire, deflated by a chained Fisher basket over three goods (cognitive tasks at $P_c$, AI services at their own clearing price $P^{\mathrm{s}}$, physical goods at $P_p$, each weighted by what households actually spent on it; services are not deflated at the task price, whose path differs because it carries human tasks and review labor), and deflated by physical prices only. The chained real-output index counts the same three goods. Under rationing the household's cognitive spending is the task budget's final-demand share less its pro rata part of the budget that could not be bought; the intermediate part of the task budget is not household consumption and carries the rest of the shortfall.

4.6 Political economy

A composite stress measure combines the uncompensated pool, its rate of increase, the long-term pool, the three-year gap between non-labor and labor income growth, and local concentration of displacement above the regional trigger the China-shock literature identifies; entrants hit by hiring freezes count half:

Friction on diffusion rises with stress, escalates above a second threshold, and is blunted by frontier concentration. Policy adjusts with a lag toward targets: transfers add a stress-linked replacement rate under a fiscal cap, active labor-market effort raises job finding, the accountability floor rises with friction and is eroded by lobbying only in the part politics added, antitrust cuts rent capture when concentration and revenue are high. A rising floor can roll adoption back. Above a 15% pool the model flags itself as outside the historical record.

4.7 Capital and concentration

Capital commitments land two years after the decision and are judged on four years of extrapolated revenue growth against a hurdle that includes obsolescence, both set by the 2019-26 record: hyperscaler AI capital spending grew 55-65% a year in 2023-25 and about 45% in 2026 at single-digit current returns, which a same-year, one-year-horizon rule cannot reproduce.

Investment responds to the return on the replacement value of the stock against a hurdle equal to a base rate plus physical wear plus obsolescence, where obsolescence is the hardware price decline, about 0.33 a year, so the hurdle is near 0.5:

with $n_h = 4$ years of extrapolated revenue growth and the multiplier decided two years before it lands. Only the frontier run's share of compute is built regardless of returns, and only to the extent lab budgets hold (AI revenue against its peak to date); research and experiment compute follow the same investment response as inference. With review labor booked to buyers (§4.3), the task block's vendor revenue, AI sold against the 2025 cognitive task stock at the modelled price, is 0.07% of GDP in 2025. The revenue record it is compared with is an accounting bridge (AI revenue dossier, Appendix B); global run rates taken without the conversions that make them comparable would put the record at 0.25-0.35% of GDP and leave a residual that reads as household-funded. The bridge takes filings and press figures, run rates converted to revenue in the year, Google Cloud at an AI share, the AI lines no filing separates as a band, the compute the labs buy from the hyperscalers eliminated, the US share attributed, and the result split by payer. On it, US-attributable net AI vendor revenue was 0.10% of GDP in 2025 (band 0.06% to 0.17%) and 0.27% in 2026 (0.17% to 0.50%), of which households paid 36% in 2025 (0.035% of GDP) and businesses most of the rest as an intermediate input. The difference between the task block and the record is AI sold outside that stock: consumer assistants, search and advertising, cloud AI services bought as final products. The model carries that revenue itself rather than anchoring the capital block to the record by a scale factor. Households spend a share of consumption on AI services, 0.03% of GDP in 2025, rising with the capability gap already closed toward a ceiling of three per cent of consumption (of the order of digital advertising plus consumer software today), priced at the vendor price times a premium of 0.3x implied by the 2025 split of inference compute between services and tasks, and the vendors receive it as income inside GDP; the services use inference compute and energy like any other AI volume. Total vendor revenue is 0.10% of GDP in 2025 and 0.39% in 2026: 1.1x and 1.5x the bridge's central, inside its bands. The services block's 2025 level is the bridge's household-paid AI (0.05% of consumption, 0.03% of GDP, or 1x what households actually paid for AI in 2025), business-paid AI is the task block, consumer usage is bounded by the share of powered inference it can run on (0.6 in 2025, so the budget sets the price per unit of usage rather than a budget over a collapsing price setting an unbounded volume), and the vendor-revenue anchor carries the bridge's band and is enforced; the ceiling remains a prior. The 2025-26 capex record is reproduced on that revenue because the 2025-27 commitments are the record's; the expectations rule takes over from 2028. Total vendor revenue is 5.7% in 2030 and 2.0% in 2050, of which services are 1.8% and tasks 0.12%: a revenue base that scales with consumption rather than with the collapsing task price. The vendor price itself floors at cost plus the gross margin (§4.3). Together these set what funds the compute stock at scale, and they set the capex cycle: commitments keep accelerating through 2033 on the record, the cut lands in 2034 as the task price collapses into saturating demand, and the build of the early 2030s, which services revenue of 1.5% of GDP in 2030 funds against hardware a thousandth of its 2025 cost, peaks at 14% of GDP in 2035, gated by power: the compute stock reaches 260x of 2025 by 2040 and 311x by 2050, with investment at its floor from the late 2030s and no second build by 2050; the power capital serving it is 16% of 2025 GDP in 2050, paid for through electricity rates (§4.5). Sized instead to total vendor revenue, 0.29% of GDP in 2025 (services_total_revenue_sizing), the services block gives a 2050 compute stock of 238x and power capital of 11% of 2025 GDP (second compute build: none by 2050). How much of the 2030s build and whether anything follows it rests on the services ceiling: at a ceiling of one per cent of consumption the 2050 compute stock is 128x rather than 311x and power capital in 2050 is 8% of 2025 GDP rather than 16%; at two per cent the stock is 267x:

With no services revenue (no_services) the investment rule cuts capex from 2026, the stock reaches 361x by 2050 and power capital peaks at 31% of 2025 GDP; with the gross margin competed down to a tenth (thin_margins) the 2050 stock is 311x, because by then services carry the revenue. Were the whole lab share built unconditionally, the compute stock would grow without limit at negative returns; it is bounded by returns and power. Frontier concentration follows run cost against budgets with a Pareto tail, $n = n_0\,[(K_f/K_{f,0})/(c_{\mathrm{hw}} C)]^{-\xi}$ with $\xi = 0.2$; deployment concentration is the share of substitutable tasks only the frontier can do, which fades as the horizon saturates.

4.8 Physical automation

Robots perform a share $D_p$ of 2025 US physical task-hours (70 million physical jobs at 1,800 hours), measured from the US fleet against the US workforce; a robot is capital, not a displaced worker, and physical employment is demand-determined rather than a residual. $D_p$ is the sum of four classes whose economics and evidence differ, each converted from its US fleet to delivered task-hours (units x hours a year x uptime x task-hour equivalence x reliability):

Industrial arms and warehouse robots are the whole of today's 3.4%; autonomous vehicles are a few thousand units and general manipulation delivers nothing yet. Each class carries its own cost, reliability, ceiling and lag:

The substitutable physical share has a pre-AI automation base and rises with the same 80% horizon as cognitive work, with a much later midpoint and a higher floor. The mature classes have fixed addressable ceilings; general manipulation takes whatever the capability ceiling leaves, so it is the class that moves with the horizon. Reliability is a state rather than a discount folded into the 2025 cost: it rises logistically toward a ceiling, and the effective cost per delivered task-hour is the all-in cost per robot-hour divided by it. Diffusion is logistic within each class, gated by that effective cost against the physical wage, slowed by friction and capped by how fast a fleet can be built:

with $b_p = 0.08$, $f_p = 0.40$, $h^p_{\mathrm{mid}} = 3000$ h, $w_p = 3$ (Appendix C; §9 of the equation companion). For general manipulation the effective cost starts at 3.3 times the physical wage ($c = 1.0$, $\rho = 0.30$), cost falls 15% a year and reliability reaches 0.8 in 2034; it undercuts the wage in 2029 and performs 0.2% of physical task-hours in 2035, 0.9% in 2040 and 21% in 2050, against 28% for the three mature classes at their ceilings. Physical labor demand is

with $\psi_p = 0.6$, $\eta_w^p = 0.5$ and hiring inertia; people available for physical work beyond it join the pool, and demand above the people available is filled from the pool at the job-finding rate. Robots enter the physical unit cost at the fleet-average effective cost and their share of value added goes to their owners. Switching the block off recovers a residual treatment of physical employment; robots_frozen holds every class at its 2025 share.

Robot capital. The robot fleet is a stock inside the capital block, so that a cost advantage becomes a robot task share only as the fleet is financed and competes for saving, energy and manufacturing capacity. Capacity is the task share times the physical volume index; one unit of capacity carries a capital value equal to the capital share of the all-in robot-hour cost (0.7, the capital charge of an industrial cell against its all-in cost in the robotics dossier) at the hurdle rate plus depreciation over an eight-year life, so it falls with the cost per task-hour (4x lower by 2050). Each year the economy buys the additions the cost advantage asks for plus replacement, the purchases land in the next year's investment demand and are funded from the same private saving as AI compute and power (after those), the book value depreciates and is written down to replacement cost as robots get cheaper, and the fleet's electricity counts in the power pipeline's demand and in the compute build's power gate. Purchases are capped at a share of the saving left after AI and power capex (0.5; the robot_finance_loose and robot_finance_tight presets set it to 1 and 0.2); when the cap binds the year's additions are scaled back across the classes. Deployment capacity is the per-class fleet growth cap already in the block. On the baseline robot purchases peak at 4.9% of GDP in 2050, the fleet's book value is 19% of GDP in 2050 after cumulative write-downs of 2.6% of 2025 GDP, the fleet draws 6% of the compute fleet's electricity in 2050, and the financing cap binds never on the baseline (first year it binds; AI capex in 2050 is 0% of GDP, the compute build long over). Robots and a late compute build compete for the same saving: with the services block sized to total vendor revenue (services_total_revenue_sizing), the first year the cap binds is never and the 2050 robot share is 48%. With the cap at one the 2050 robot share is 48% against 48% on the baseline and 48% at 0.2, so the dates of the second labor wave in §6.1 are conditional on how the fleet is financed and on whether a compute build competes for the saving, and should be read as a band, not a year.

5. Calibration and fitting

Anchors come from the four evidence dossiers and the external audit reports (Appendix B); the evidence synthesis maps each anchor to its sources with a confidence (Appendix B). The parameters that calibration moved furthest from their starting values, and why:

What is fitted: kappa (to 2025 progress), the 2025 normalisations (GDP, basket, demand index, cognitive consumption share, employment), the autonomous-demand residual, and the 2025-26 early path (horizon, substitution midpoint and width, diffusion lag and level, the separation and entry shares, the sector leads) to the anchors in section 2, which are re-checked on every run (Appendix B). What is set from evidence: everything else in the table. What is a prior: the political reaction parameters, the pipeline mix, the CES weight and curvature, the pass-through power psi. Sensitivity (one at a time, ±30%): the economic outputs respond most to the demand elasticity, maximum friction, job finding and absorption, psi, and the diffusion lag; the progress parameters move capability but not the economics, because capability saturates far above the substitution midpoint.

The timeline begins in 2019. The six years before the 2025 anchor are the record, not simulation: where an observed annual series exists (the METR horizon, frontier training compute, inference cost and list price, AI task share, capital spending, real GDP, the number of frontier actors) the rows carry it, and the rest is reconstructed from the fitting targets below. Every chart shades those years and indices read "2025 anchor = 1", because 2025 is the one year every quantity is fitted to. Figure 0 shows the run-up: compute at 0.7% of its 2025 level in 2019, the agent task horizon at two seconds in 2019 (GPT-2), four minutes in 2023 (GPT-4), twenty to forty minutes in 2024 and three and a half hours in 2025. Fitting targets for 2020-2025 (frontier compute 4-5x/yr, efficiency 2x per 8 months, capex +55-65%/yr, run cost 2.4x/yr, inference price 40x/yr, horizon doubling every 4-7 months, firm adoption +2 points/yr) are listed in the external AI-progress audit (Appendix B); the model reproduces the horizon doubling (four months since 2024), the run-cost growth (2.5x), the progress rate and the adoption level by construction, and the capex path only in direction.

The record is also used as a hindcast and a continuity check (the tables below). The hindcast runs the backcast formulas against the 2019-2024 observations and shows where they fail: the horizon mapping with the post-2024 exponent puts 2019-2023 five to twenty times below METR, because the horizon-per-capability slope steepened with reasoning and agentic training in late 2024, and the constant-rate price formula misses the flat 2020-22 list prices; in both cases the charts show the record. The continuity check compares observed growth to 2025 with the forward model's 2025-28 rates and flags a kink; algorithmic efficiency, the AI task share, the horizon, hardware cost and real output are continuous, frontier compute and capital spending slow with the capex cycle by design. Three forward parameters are set by the record: the concentration elasticity (at 0.6 the frontier collapses to one actor by 2030, while the record went from two labs to five-seven), a two-year commitment lag on capital spending, and a four-year investment horizon (a one-year horizon cuts capex in 2025-26, while the record grew it 45% in 2026). The power pipeline carries the commitments made before 2025: started empty, its power gate would leave 31% of the 2026 stock unpowered and put 2026 capex at 0.9x of 2025, while the record built 1.45x. With the firm capacity that BNEF counts arriving in 2026-27 (8.5 GW realised in 2025, 13.6 GW scheduled for 2026 on a 62 GW base, AI-focused load up 50% in 2025) seeded as a parameter, and the advanced-packaging bound growing at the CoWoS record (+60-70% in 2026, then +40% a year), 2026 capex reads 1.25x and the anchor is enforced, with a firm-capacity anchor for 2026 alongside it.

6. Results

6.1 The baseline by decade

Figures 1-7, 11-13. Separations and foregone hiring are in people, at a 170 million workforce. Two rows use different denominators and the gap between them is a result, not an inconsistency: the substitutable share is a share of the 2025 task stock and saturates at one minus the accountability floor (0.85), while the AI share of all cognitive tasks counts the volume that cheap cognition induces. Human task demand is H = (1 - D) T_trend max(1, T / T_trend)^psi_eff, so the human share of the cognitive bundle is h = H / T and the AI share is 1 - h. By the mid-2030s psi_eff is zero (the capability gap is closed, §4.4) and humans hold the non-automated share of trend volume only: in 2050 they still do 16% of the 2025 task stock, but that stock is a small fraction of the 2050 volume, so the AI share by volume is 100%. The task horizon is capped in this table at a work-year because substitutability saturates long before the mapping does; a horizon of ten thousand years is a number the formula produces, not a capability the model uses. The task-accounting table below gives the identity in numbers for the baseline run. The labor-demand equation uses $D$, the share of the 2025 task stock performed by AI, which saturates at the substitutable share; the volume-based share does not belong in it.

Critical transitions. The timing of qualitative changes is a primary output of the model: the table below gives the first year each threshold holds, for the baseline and the scenarios that move it. Two rows deserve the reader's scepticism, as section 4.3 says of the horizon mapping: the research rows come from the recursion's AI-research-labor index crossing the human workforce in 2027, which is the model's own capability index applied to research, not an observed quantity.

The years of capability, which are now. On the fitted horizon, agents that reliably complete month-long tasks at 50% success arrive in 2027 (at 80% success, 2028), and the substitutable share of cognitive tasks passes one half in

1. Adoption follows three to four years behind rather than a decade: AI performs

half of cognitive work by volume in 2030 and 90% in 2034, and half of the 2025 task stock in 2033. The capex cycle follows the record for as long as the record runs: capital spending grows on commitments made in 2025-26 and peaks in 2033, power binds from 2026 (utilization bottoms at 0.70) and clears in 2036, and the cut decided as prices collapse into saturating demand holds investment at the floor through the 2030s. Impairments reach 7.5% of 2025 GDP by 2035 and 10.3% by 2050.

The years of labor, which begin in 2028. AI-attributed separations run near 79 thousand in 2025 and 211 thousand in 2026, which is the record, then 2.1 million in 2028 and 3.4 million to 1.7 million a year from 2030 to 2035, on top of a foregone-hiring channel that keeps one to two million entrants a year out of cognitive work. Human cognitive task demand falls to 0.69 by 2035; cognitive employment falls from 45% to 30% of the workforce by 2040. On the cognitive wave the displaced pool peaks at 8.0% in 2031, between 2009 and the early 1980s, because physical demand, rising with real income, hires: physical employment grows from 55% 60% of the workforce by 2044. The composite stress peaks in 2029 at 0.35 and friction at 0.75 in 2030, which is the brake that keeps the cognitive-wave pool below the 11.4% it reaches in 2035 with no political response; transfers replace 56% of lost cognitive wages by 2035 and the accountability floor rises from 15% to 27.8%. The peak displacement rate is 6.3% of the workforce a year in 2029, several times any national precedent, and is the number in this paper that most needs the reader's scepticism: it is the 2025-26 separation share applied to years in which AI's task share rises ten points a year.

The years of robots, which begin around 2040. Industrial arms and warehouse robots are below the physical wage already; general-purpose robots cross it per delivered task-hour in 2029 once reliability is counted, and the all-class share of physical task-hours moves from 20% in 2040 to a quarter in 2044 and 48% in 2050, 21% of it general manipulation (the fleet is bought, financed and powered inside the capital block, §4.8, and the late-2040s share is financing-bound wherever a second compute build takes the saving, so these dates are a band of a few years, not a year); physical employment falls from its 60% peak to 53%, and this time nothing hires: the displaced pool rises from 4.5% in 2040 to 9.9% in 2050 and is still rising at the end of the horizon, outside the historical record, with the labor share at 0.57 and 9.1% of the 2025 workforce gone from it. With every robot class held at its 2025 share and the same demand-determined physical labor (the robots_frozen preset, the deployment counterfactual used throughout this paper because it keeps the closure) the pool is 0.2% in 2050 after a 6.5% peak in 2030. The no_robotics preset is a different closure, not a counterfactual: it switches physical automation off and restores residual absorption of displaced workers into physical work, and is reported here once for continuity (pool 0.4% in 2050 after a 5.8% peak in 2029). With fast robotics the pool is 16.8%; with slow robotics robots reach only 22% of physical tasks by 2050. The robot wave is the part of the model that most depends on a forecast rather than a record: its cost path, its midpoint and its fleet cap are 2026 estimates, and the wave is unfinished in 2050.

The years of prices. Nominal GDP in the wage numeraire falls 44% by 2040 while real income rises: 2.8x deflated by physical prices by 2050, 9.7x by the chained all-goods basket. The difference is the price of cognition, whose compute cost falls by 8.0 orders of magnitude (the buyer's all-in cost by 114x) while cognition is a quarter of the basket. Physical prices fall 84% through the cognitive-input channel and the robot share of physical unit cost. Household electricity rates rise 26% by 2050 from socialised grid capital.

6.2 Seven mechanisms, each with its refutation

Mechanism 1. The price of cognition falls by orders of magnitude. From 2025 to 2050 at fixed capability the compute cost of a task falls 8.0 orders of magnitude, the vendor price 3.9 orders once rents and scarcity are counted, and the buyer's all-in cost, review labor included, 114x, under the baseline assumptions of sustained algorithmic progress, continued hardware efficiency gains and rents competed away as the frontier consolidates slowly; this is a conditional model result, not a forecast. Precedents: computing twelve, light three and a half, inference prices forty-fold a year today. Refuted if fixed-capability inference prices stop falling faster than twofold a year for three consecutive years.

Mechanism 2. Stress is timing times institutions, not capability or energy. Across the 81-run grid, peak stress ranges from 0.25 to 0.67 and the 2035 pool from 3.7% to 9.0%; the worst corner is always fast diffusion with weak institutions, in any progress regime and at any energy buildout. With robots on, the end-of-horizon pool runs from 8.9% to 14.8% and 0 of 81 runs leave the historical record before 2050, almost all on the robot wave of the 2040s rather than on the cognitive wave. The energy axis moves utilization, capability and electricity prices but moves peak stress by under a tenth. Refuted if a fast-adopting economy with strong active labor-market institutions shows persistent displacement above a slow-adopting peer.

Mechanism 3. The capex cycle is driven by power and obsolescence, and the record sets its timing. Capital spending grows through 2033 on commitments made while returns were single-digit, utilization runs at 70% to 87% from 2026 to 2034 because power arrives through a 5-year pipeline, the cut is decided 2 years before it lands in 2034 as the task price collapses into saturating demand, and investment runs at 0.20 of trend by 2040 (floor 0.2); cumulative impairments 7.5% of 2025 GDP by 2035 and 10.3% by 2050. Whether the cycle has a second leg turns on AI services revenue, near 2.0% of GDP against hardware a billionth of its 2025 cost (first year of a second build: never); gated by power, the compute stock reaches 311x by 2050 and the power capital serving it 16% of 2025 GDP, paid through electricity rates; without services revenue the stock is 361x (§4.7). Refuted if hyperscaler capex is cut before 2034 with utilization above 85%, or grows past it with no write-downs.

Mechanism 4. Who buys the output depends on who owns what, who pays for the grid, and whether intelligence is scarce. Three pricing regimes bracket the baseline. Competitive intelligence (no rents, no scarcity) prices AI at cost and barely differs from the baseline, whose rents fade as the frontier consolidates slowly. Scarce intelligence (power buildout at 15% a year, efficiency gains spent on quality) has inference demand outrun powered supply from 2028: scarcity takes up to half of the gap to the human wage, AI is priced at 0.54 of the wage in 2030 against 0.04, AI revenue in 2030 is 1.1x the baseline's, the labor share ends at 0.38 instead of 0.57, cognitive volume stays depressed, and the displaced pool passes 10% of the workforce in 2035. That is the regime in which technological efficiency does not become consumer savings. With AI rents competed away, the physical-goods real income of a physical worker in 2050 is 5.3x, a cognitive worker 4.5x, a capital owner 2.5x; narrow ownership with no transfers reads 2.69x against the baseline's 2.80x for the average household; household electricity rates rise 26% by 2050. Refuted if frontier margins stay above half while open weights trail by under a year, or if retail rates fall in data-centre states.

Mechanism 5. Trades are stranded the way compute is, and at 7 times the scale. The trades pass half automation in sequence: business and financial operations, legal, arts and media, office and administrative support in 2029; management, architecture and engineering, life, physical and social science, educational instruction and library in 2030; computer and mathematical, sales and related in 2031; community and social service in 2032; healthcare practitioners and technical in 2034; licensed care is capped at its floor. At the 2032 peak, 72 million people, 42% of the 2025 workforce holding 187 million person-years of trade-specific training, see AI doing more than half of their trade's tasks. The skill premium lost peaks near $1.5 trillion a year in 2033 and cumulates to $22.3 trillion by 2050, 77% of one year's 2025 GDP, against compute impairments of 10.3%. Most of these people remain employed; what they lose is the value of what they trained for. The soft-landing preset defers the peak to 2038 and lowers the loss to 51% of GDP; the stress preset brings it to 2031 and raises it to 80%. Refuted if occupational wage premia in the exposed groups hold through 2035 while AI task shares pass one half.

Mechanism 6. Remove the favorable assumptions and the aggregate index still rises; wages and participation do not. What remains of the renaissance if displaced workers do not find work, AI rents are not competed away, cognitive gains do not flow into physical prices, and ownership is narrow? The regime_change preset removes all of them at once, with no new human tasks: job finding 0.15, rent capture 0.8, scarce intelligence, a cognitive-input share of physical costs of 0.05 and a fifth of capital income paid out. The all-goods real-income index still reads 3.0x in 2050 and physical-goods household income 1.7x, because the index counts near-free cognition and transfers and capital payouts reach households. Everything that runs through work goes the other way: the labor share falls to 0.49, the cognitive wage in physical goods ends at 0.75 of 2025 against the baseline's 5.13 (tasks a scarce AI cannot deliver revert to humans, which is why it does not collapse), the physical price index ends at 1.02 instead of 0.16, the displaced pool peaks at 13.7% in 2038 and 13.4% of the 2025 workforce has left it by 2050, with transfers replacing 66% of lost wages. The no-new-tasks case (psi = 0) does not separate from the baseline by 2050: because the complementarity decays with capability (§4.4) the baseline itself serves induced volume with AI from the mid-2030s, and the two end with the same labor share (0.57). The index-number problem of that scenario therefore belongs to the baseline. Cognitive volume reaches 131 times 2025 at a price near zero; nominal household income falls to 0.39 of 2025 while the chained basket price falls to 0.047, so the chained all-goods index reads 9.7x while the physical-goods lens reads 2.8x (AI services enter the chained index as a third good at their own price; deflating them at the task price would understate the fall in the basket price and lower this reading). A chained Fisher index is the right deflator for a consumer who values the marginal unit of cognition at its price, and the wrong one for a reader who wants to know whether people are better off, because nobody consumes a hundred times today's legal advice. The paper therefore adds a satiation-bounded lens: cognitive consumption counted at its 2025 value up to three times its 2025 volume while cognition is the cheaper way to buy value, with the rest of income buying physical goods, which reads 3.3x in 2050 (1.8x in 2035). The headline real-income claims of this paper are the physical-goods and satiation-bounded lenses; the chained index is reported because national accounts will report something like it. Under fixed complementarity (the sustained family of §4.4) the physical-goods lens reads 2.4x while the chained index reads 5.0x instead of 9.7x: the physical-goods lens moves far less with who does the work than the chained index, which moves through the weight of near-free cognition in the basket, one more reason not to stand on it. So the answer is split: recursive improvement and cheap cognition produce the aggregate abundance on their own; whether it is a renaissance for workers depends entirely on the four assumptions named, which is the point of section 9. Refuted if, by 2035, exposed-occupation wages hold while rents, scarcity and narrow ownership are all observed.

Mechanism 7. Physical work absorbs the first wave and robots take it back. With physical employment demand-determined, the 2030s hire: real income rising 36% in physical goods lifts physical demand, and physical employment grows from 55% 60% of the workforce by 2044 while cognitive employment falls to 30%. The 2040s reverse it as robots pass the physical wage: physical employment falls to 53% and the pool reaches 9.9% by 2050. With every robot class frozen at its 2025 share (the same closure) the pool ends at 0.2%, the fast path at 16.8%, the slow path (general-manipulation cost falling 8% a year, half-speed diffusion) leaves robots at 22% of physical tasks. Refuted if humanoid-class robots are not below the loaded physical wage per delivered task-hour by 2035, or if physical employment keeps rising through the 2040s while they are.

6.3 The scenario grid

Axes: progress regime (compute-bound diminishing returns, baseline, sustained recursion; §4.2), diffusion speed (a diffusion lag of 12 years, the fitted 4.4, and 3 years), institutional adaptation (job finding 0.25, 0.4, 0.6 with policy speed), energy buildout (0.15, 0.4, 0.8 a year). 81 runs, robots on; all converge; 0 take the displaced pool above the 15% of the workforce that the historical record covers, almost all on the robot wave after 2040.

The labor share moves across the grid from 0.416 to 0.589; the stress variable moves fivefold and the 2035 pool more than threefold.

6.4 Counterfactual switches

The political reaction function is the most consequential prior in the model, so it is reported as regimes rather than as one baseline. Without politics (no friction, no reaction) the pool is 11.4% in 2035 rather than 7.6%, and the share of the 2025 task stock done by AI in 2040 is 0.85 rather than 0.76: the political response cuts the cognitive-wave shock by well over half, at the cost of a decade's delay in adoption. With friction but no policy reaction the 2035 pool is 8.6%.

Climate. Climate enters as a drift on physical unit cost and on the electricity price, off in the baseline (every number in this paper is without it unless it says otherwise), with an adaptation-investment term and a regional preset. On the central drift (Swiss Re's 2.6 °C path, US GDP seven per cent smaller by 2050) net damage on physical unit cost is 7.8% by 2050, the physical price ends at 0.172 instead of 0.164 and physical-goods real income at 2.78x instead of 2.80x, a 1.0% loss on the lens the paper stands on; adaptation at half a per cent of physical value added a year cuts net damage to 4.8% and leaves income at 2.79x; an exposed region at twice the drift reads 16.1% and 2.75x. The cheap-cognition gain is an order of magnitude larger than the climate loss on this horizon, which is why the drift is reported as a lens and not folded into the baseline; the sector-price lens (§7.6) is where it would attach by sector.

Technology against politics. Because the baseline rests on one reaction function, two decompositions answer how much of the labor outcome is technology and how much politics. First, the institutional regimes crossed with the technology: each cell below lays an institutional preset over a technology preset and reports the peak displaced pool on the cognitive wave and cognitive employment in 2035.

On the baseline technology the pool peaks at 18.7% with no political response and 8.0% with a responsive one: politics' reach is 11 points, and the modeled reaction removes 57% of the no-politics peak. Among the reaction functions that do respond, from responsive to captured, the spread is 2.4 points, while technology's reach from slow to fast diffusion is 3.7 points under the baseline institutions and 6.4 with none. So the cognitive-wave peak is a political quantity in one sense and not in another: whether a reaction exists is worth more than a doubling of the technology's pace, and which reaction it is, within the plausible range, is worth about as much as the pace. Cognitive employment in 2035 is the reverse: the technology rows differ by ten points and the institutional columns by a few. Second, a variance decomposition over the 81-run grid (progress regime, diffusion speed, institutions, energy; every run with a reaction function). Four caveats go with it: the percentages depend on the ranges chosen for each axis and weight the levels equally; the grid has no no-response case, so the reach of politics as a whole (the first table) is outside it; the institutions axis bundles the reaction function with job-finding and scarring, which are labor-market parameters; and the interaction terms are large, so the main-effect shares are a description of this design, not a causal allocation between technology and politics:

Within the grid the institutions axis carries 1% of the variance of the peak pool against 90% for the technology axes together (9% is interactions), 0% of cognitive employment in 2035 (diffusion speed alone carries 99%), and at most 6% of any outcome (the labor share in 2050 and the exits, where transfers and the accountability floor act). Read together: technology sets the size and timing of the cognitive wave and what people do afterwards; politics decides whether the displaced are carried through it, and that decision is binary in its effect, present or absent, more than it is graded. The robot wave of the 2040s is largely indifferent to the regime. Without the demand closure the 2035 pool is 11.4% and all-goods real income in 2050 is 3.7x rather than 9.7x. Without capital markets the stock is overbuilt and capability higher. Wage insurance at 60% raises real income and the replacement rate and leaves peak stress within a few per cent (0.35 against 0.35: higher replacement lowers stress per displaced worker while the faster diffusion it permits raises the pool at the peak), but it does not lower displacement (2035 pool 7.6%), because compensated workers resist less and diffusion runs faster: compensation enables disruption, which is the compensation hypothesis of the trade literature appearing as a model result; the instrument buys lower stress, not fewer displaced.

6.5 Joint uncertainty

The joint analysis traces the 2050 outcomes to the parameters that carry them. 40 parameters are sampled jointly from ranges the dossiers bracket (marked evidence) or that are structural assumptions without a direct measurement (marked prior), by Latin hypercube, 1,000 draws, seeded (the priors are Appendix D and the command is in Appendix B). Each draw is a full run; runs that leave the historical record are counted, not dropped. A year converges when the fixed point's step on the plan (the income index, physical and cognitive sales, the matched fill, services spending) falls below 1e-8 within 600 iterations; a run with any year that does not is reported and excluded from the distributions rather than collected as an observation, and among the excluded a residual below 1e-5 is called near-converged (a limit cycle at the fourth decimal, excluded all the same) and one at or above it failed. For the run reported here, all 1,000 draws converged. First-order sensitivity indices are variance ratios over equal-count bins, less what an independent variable would show by chance; interactions are in the unexplained remainder.

Three things the distributions say that the baseline does not. First, the timing of the cognitive wave is robust and the size of the robot wave is not: half of cognitive work by volume falls in 2030-2033 in the middle half of draws, while the 2050 pool runs from 0.7% to 20.3% and 20% of draws leave the historical record before 2050, almost all on the robot wave. Second, the aggregate real-income figure rests on a prior. The all-goods index in 2050 runs from 4.3x to 14.6x, and its variance is carried by ai_review_overhead_decline, a structural assumption, with evidence-ranged parameters explaining 0.23 of it against 0.58 from priors; the physical-goods lens is tighter (1.9x to 3.1x). Third, the labor outcomes are carried about equally by evidence and priors: the pool by political capture, the physical floor and the robot cost decline; exits from the workforce almost entirely by the job-finding rate, which is measured. The honest statement of the paper's central numbers is therefore the interquartile range, not the baseline: a displaced pool of 4.7% to 8.6% in 2035 and 4.5% to 13.2% in 2050, physical employment of 44% to 58% of the workforce in 2050, a labor share of 0.56 to 0.58, and real income in physical goods of 2.2x to 2.6x.

Structural families. The joint run above is conditional on the declining family of §4.4. Run under each family with the same priors apart from the family's own parameter (300 draws each), the labor outcomes separate by structure more than by anything inside a structure:

The families are not yet distinguishable on the 2025-26 record, since the capability gap is small; they separate by the early 2030s in whether total cognitive employment and the human cognitive wage keep falling while AI volume explodes (declining) or hold (sustained, created), and in the displaced pool of 2033. The computer and mathematical occupations are not the place to look: they are bound to the research series (§4.4), whose direct work falls to the handful in every family and whose review labor is largest where AI volume is largest, so their employment is lower, not higher, under the families that keep more humans at work.

6.6 What the people do

This section asks where the people go. At some point AI does research and engineering better than all but a handful of people, and then it does the rest of cognitive work better too. Those jobs are gone. A model in which humans keep a share $\psi$ of every task that cheap cognition induces, however capable AI becomes, holds cognitive employment near a third of the workforce in 2050 even while AI performs most of the task stock; that complementarity is an assumption, not a result, and the baseline does not make it (§4.4, where it is kept as the sustained family). The table is generated from the baseline run.

The displaced pool is a state of this model, the workers separated from cognitive or physical work who have not been re-employed, retrained out of the workforce or retired, and it is not the measured unemployment rate: it counts people the statistics would class as unemployed, marginally attached, in training or out of the labor force, and it does not count frictional unemployment that exists without AI. The labor states should be read together rather than through the pool alone, so the table below reports participation, employment by sector, the pool and its long-term part, hours per cognitive worker, the human cognitive wage, labor income in physical goods and household purchasing power in essentials on the same rows: participation ends at 91% of the 2025 workforce, the long-term pool at 5.9%, hours per cognitive worker fall to 1.00 at their lowest, and the wage, job-finding, review and hours-flexibility responses behind these rows remain assumptions with priors rather than estimated behavior (Appendix D).

Three things happen, in order. Research and engineering go first. AI is the majority of research labor in 2027 and 90% of it in 2028; the human research and engineering workforce is 52% of its 2025 size in 2030, 8% in 2035 and 2.9% in 2040, from 9.5 million people to 793 thousand and then to the handful, reached within a factor of two in 2046 (100 thousand on the 1% parameter; a tenth of the field staying gives 972 thousand). That is direct research. Counting the review labor on AI-produced research and engineering output, the three occupations employ 3.5 million in 2035 and 3.4 million in 2050, 3.3 million of them checking what the machines designed: the field does not empty, it changes what it does. The engine of §4.2 keeps running on AI research labor; the people who built it are not in it. Direct cognitive work halves; what remains is largely reviewing AI output. Cognitive employment falls from 45% of the workforce to 37% in 2030, 31% in 2035 (a quarter gone in 2033) and 28% in 2050: 29.6 million cognitive jobs gone since 2025, 17% of the 2025 workforce, not relocated within cognitive work. Of the cognitive employment that remains in 2050, 56% is review labor on AI output (26.2 million people, from 21.4 million in 2035), and 12% of the workforce does direct cognitive work: the accountability floor, the share of trend volume that institutions insist a person signs for, and the hours spent checking what the machines produced. Politics sets the floor; reliability sets the review time, and the uncertainty study carries its decline as a prior. The people go to physical work, then out of the workforce. Physical employment rises from 55% 58% in 2035 and 53% in 2050, because a workforce swollen by cognitive refugees and the stress friction they produce slow the robot wave (robots do 48% of physical task-hours in 2050, general manipulation 21%); the long-term pool holds 5.9% of the workforce in 2050 and 9.1% have left it. Wages carry 73% of household income in 2025 and 67% in 2050; transfers carry 10% and paid-out capital income 24%. Labor income stops covering household consumption in 2029.

Under fixed complementarity (fixed_complementarity, the sustained family) cognitive employment in 2050 is 36% instead of 28%, 5.7% have left the workforce instead of 9.1%, and the labor share is 0.57 instead of 0.57. That difference is one assumption: whether a person keeps a share of work AI does better. The paper's position is that they do not, outside the floor, and that the floor is the policy variable; the created family of §4.4, in which AI makes new human work, gives 39% cognitive employment and a labor share of 0.57 in 2050. The physical-goods real-income index moves much less across the families (2.8x baseline against 2.4x sustained) than the chained all-goods index (9.7x against 5.0x and 3.9x), which counts the near-free cognition that the declining family lets AI serve. What changes most is who receives the income, which is why §8 and §9 are about distribution and not about output. The handful-of-researchers result is conditional on its retention rule: the handful parameter (1% of the 2025 field) and the rule that research tasks follow the capability sigmoid with that handful as the floor; a tenth of the field staying is reported alongside it.

6.7 Recursive-improvement regimes

The three mechanisms of §4.2 are run side by side, with two further parameterisations of diminishing returns (compute-bound and algorithm-bound) for comparison, and the breakthrough regime over 60 seeds.

The regimes differ by orders of magnitude in what they say about algorithms, by nothing in the threshold dates and the terminal labor share, and by a great deal in the transition path between them. Under sustained recursion algorithmic efficiency in 2040 is 2.8 × 10⁹ against 8.6 × 10⁴ in the baseline, with research productivity reaching its effective-limits cap in 2032; under breakthrough-driven progress the 2040 figure runs from 8.6 × 10⁴ to 5.1 × 10⁵ across seeds, with 2 breakthroughs by 2035 and 7 by 2050 at the median. Yet AI performs half of the 2025 task stock in 2032 under sustained recursion and 2033 in the baseline, in every breakthrough seed the year lies in 2032 to 2033, the pool peaks at 10.0% against 9.9%, and the 2050 labor share is 0.58 in both (breakthrough seeds: 0.57 to 0.58, the low end from hardware-cost breakthroughs that strand compute capital). The dates and the end state are shared; the pool peak is not, and a transition twice as deep is a different decade for the people in it. Three things sit between algorithms and the shared dates and end state, and they are properties of this model's architecture, not evidence that labor markets are insensitive to recursive progress. The substitutability sigmoid saturates at the accountability floor within a few hundred hours of horizon, which every regime reaches before 2030, so further capability has nothing left to substitute. Diffusion, friction and the political reaction gate adoption at the same pace whatever the capability. And the buyer's all-in cost of cognition by 2035 is mostly their own review labor, not the compute: the AI unit cost reads the same to two figures in every regime because the compute component is already a few percent of it. The regime question is therefore a question about compute, energy and the price of AI output after 2035, where the regimes diverge, and not about the labor transition of the 2030s, which the record of 2025-26 has already set in motion. It would become a labor question only if a breakthrough raised the accountability floor's complement, that is, if institutions let AI sign for the work; that is the political block's variable, not the research block's.

6.8 Against the task-based benchmark: Acemoglu (2024)

We treat Acemoglu's "simple macroeconomics of AI" as a competing model rather than a citation. His decomposition is an identity: the productivity gain from AI over a decade is the GDP-weighted share of tasks that are exposed and profitably automatable at today's capability and cost (4.6%, Eloundou's exposure times Svanberg's profitability), times the average cost saving on those tasks (14.4% of task cost, 27% of labor cost), giving 0.66% of TFP over ten years and about one per cent of GDP with capital deepening. The same arithmetic applied to this model's quantities year by year (the share of the 2025 cognitive task stock performed by AI, the cognitive share of GDP, the buyer's all-in cost per AI task against the wage) is reported beside the model's own output gain against a no-AI counterfactual.

The estimand of that comparison needs stating. A limit that switches off cognitive substitution only still carries AI services, AI research agents supplying 100.0% of effective research labor in 2034, robots on 54.8% of physical tasks by 2050 and a trace of cognitive diffusion (0.00%), so that "output gain versus no AI" would be measured against a world with most of the AI in it. The counterfactual here is no new AI development from 2025: nothing substitutable and nothing diffused, no AI services, no AI research agents, robots held at their 2025 share, and only the compute committed in 2025 landing (no frontier race, no investment floor), so that 0 of the five AI channels are on, compute peaks at 5.6 times 2025 in 2030 and depreciates, and real output reaches 1.10 by 2050 on physical productivity alone. The cognitive-only limit is kept as no_cognitive_ai, and the table below reads each case's active channels from its run rather than from its name; the baseline's gain against any row but the first is a channel ablation, not a gain from AI:

Against the estimand the baseline's real output is 68% higher in 2034 and 429% in 2050; against the cognitive-only limit it is 64% and 341%, the difference being the robots and services that limit keeps. Cognitive substitution alone (cognitive_only) accounts for 44% over the estimand by 2034 and 186% by 2050; the rest is services, research agents and, after 2040, robots. These are changes in the model's real-output index, which prices cognitive output at its falling price; the essentials lens of §7.6 is the other reading. Acemoglu's arithmetic beside the model's gain against the estimand, with four objects kept apart: potential exposure (the substitutable share at the year's capability and floor), adopted capacity (the diffusion state $D$), desired AI production, and delivered automation (the share of the 2025 stock AI actually performed, stock_automation_delivered). The arithmetic runs on delivered automation, and the saving is converted to his basis: his 27% of labor cost is 14.4% of task cost, so labor is 53% of task cost in his calibration and the model's labor-cost saving is multiplied by the same share:

The acemoglu_2024 row is a frozen-capability sensitivity case, not a reproduction of his calibration: it holds capability at 2024-25 (no horizon growth), holds the compute cost (no hardware or efficiency decline), sets a ten-year diffusion lag, no new tasks and no review-overhead decline, but it does not impose his exposure or cost-saving inputs, the demand closure and the capital block run, and the all-in cost still moves with the wage, the margin and scarcity, from 0.25 of the wage in 2025 to 0.22 in 2034, so the price is not fixed. On that case the substitutable share in 2034 is 3.6% of GDP, adoption 3.6% and delivered automation 3.2% against his 4.6%; the saving is 78% of labor cost, 42% of task cost against his 14.4% (the model's 2025 all-in cost is a quarter of the wage, his is a third less than the wage); and the task arithmetic on delivered automation at his basis gives 1.31% of TFP against his 0.66%. The frozen case is where the below-trend rule of §4.4 matters: nothing induces volume, so volume falls a few per cent below trend, and a human-share rule that served the shortfall from humans would send every adopted stock task back to them, reporting adoption with no delivered AI task volume. With both sides shrinking in proportion below trend, adoption is production. The model's own output gain on the frozen case is 2.5%, within a per cent of zero either way, because the task arithmetic books the cost saving as productivity while the demand closure also books the wage loss of the displaced tasks and the compute built for them.

From his benchmark to the baseline the inputs are then switched one at a time:

Capability growth is the input that matters: with the horizon doubling and everything else frozen, delivered automation in 2034 is 14.2% of GDP, the task arithmetic 6.1% and the model's gain 23%. The price path alone takes the all-in cost to 0.15 of the wage by 2034 and moves output by 3.0%; the shorter lag, new tasks and the review decline move it by about as much; removing the demand closure from the frozen case gives 8.9%, the gap being the wage loss the closure books. The baseline has all of them: delivered automation 27.9% of GDP in 2034 rather than 4.6%, a saving of 96% of labor cost (51% of task cost) rather than 14.4%, a diffusion lag fitted to the 2025-26 record (§4.3) rather than ten years, and declining complementarity sending induced volume to AI, where his headline exercise has no new tasks (his accepted manuscript discusses them; the created_complementarity family of §4.4 is the Autor reading). The task arithmetic on the baseline gives 14.2% of TFP by 2034; the model's own gain is 68% of real output (20% by 2030) and 24% of physical-goods real income, the difference being induced volume, the cognitive-input channel into physical prices and the capital built. Where the two diverge is therefore not the accounting but two priors this paper states and he does not need: that the horizon keeps doubling, and that the price of AI output keeps falling.

7. Extrapolations

7.1 Occupations

BLS major groups with 2025 employment, constructed AI task-exposure scores and accountability floors, and a sector lead and a demand elasticity per group fitted to 2025-26. Computing follows its own adoption curve rather than a lead on the economy-wide one: its substitutable share comes from the capability horizon with a midpoint of eight hours (the 80% horizon at which half of its tasks are substitutable) times an adoption fraction that is never below a half, fitted to its 2025-26 record (fully-automated share 10% in 2025, 20% in 2026, 38% in 2028, 47% in 2030, at its 90% ceiling by 2036), and the other groups are renormalised to the rest of the economy-wide share; its headcount barely falls, because in 2025-26 output per engineer rose severalfold while payrolls fell only 1-2% a year. Administrative work runs two and a half years ahead with little output expansion. Baseline 2050, in 2025-workforce units: office and administrative support 18.8 million to 4.6 million; computer and mathematical 5.5 to 2.3 million in 2035 and 2.2 million in 2050 (bound to the research series of §4.4: direct research to the handful plus review labor); business and finance 11.4 to 5.8 million; management 13.7 to 7.5 million; sales 14.5 to 9.1 million; healthcare practitioners 10.2 to 10.5 million; education 9.9 to 9.0 million. Physical groups follow the robot wave rather than a residual: production 8.8 to 4.4 million, transportation and material moving 14.1 to 8.0 million, construction 7.5 to 8.0 million, healthcare support 8.4 to 14.0 million; 35.7 million displaced and not re-employed; 5.9 million out of the workforce. Within physical work each group's headcount falls with its robot task share, damped by its demand elasticity, care groups track ageing, and the pattern is renormalised to the model's physical employment, so the split across physical groups is that allocation, not a forecast.

7.2 Education

Hiring in exposed fields falls with employment; replacement demand in licensed and physical fields rises with their growth. The pipeline is already moving with a one-to-two year lag. The binding constraint is not what to study but retraining capacity: the model's absorbed flow peaks at 4.7 million a year in 2035, nearly four times a system that completes about 1.2 million substantive credentials. AI tutoring helps when scaffolded and harms when unguarded; it is a tool for the transition, not a substitute for the institutions.

7.3 Households

Three archetypes built from the model's income types, each in nominal, all-goods real and physical-goods real terms, with an employment probability by origin sector (the pool is tracked by the sector it came from). In 2035, on the cognitive wave, the physical worker is best off (2.0x in physical goods, employed with probability 100%) and the cognitive worker worst (1.5x, employed 80%), with the capital owner at 1.2x. By 2050, on the robot wave, the physical worker's employment probability has fallen to 90%, and the three read: cognitive worker 4.5x, physical worker 5.3x, capital owner 2.5x. Time use moves little: a century of productivity bought four to five extra leisure hours a week, and cash transfers cut hours by an hour a week per $1,000 a month. The scarce layer is care: the over-65 population grows about a third by 2050, direct-care openings run near nine million this decade, and humanoid robots do not reach mass adoption before the late 2030s.

7.4 Stranded human capital

Mechanism 5 above. The measure values a trade as the present value of its skill premium over the remaining careers of its 2025 holders, weighted by training years. It overstates management, whose premium is not all trade-specific, and understates administrative work, whose trade carried little premium so that the loss is the job rather than the premium. We present it with job loss as a pair of bounds.

7.5 Distribution: who receives the income and what it buys

The distribution block separates production (who or what produces), ownership (who receives the income), distribution (how it reaches households) and consumption (what households can afford), and reports median disposable income, bottom-quintile purchasing power, capital-income concentration and the share of households able to afford a fixed basket of necessities rather than a single labor-share statistic. It allocates the model's three income flows each year, wages by sector, transfers and capital payouts, to five household quintiles with 2025 shares from the record: labor income by quintile from CBO's distribution of household income, split between cognitive and physical wage bills by occupation wage levels; capital income with the top quintile holding 83% (Fed Distributional Financial Accounts: the top 10% hold 87% of equities, the bottom half 1%). It adds nothing to aggregate income; the quintiles sum to the model's household income in every year (a validation identity). Wage losses inside a sector fall proportionally on the quintiles that earned those wages; transfers go to households in proportion to their lost cognitive wages (wage insurance) or, by parameter, equally; payouts follow ownership unless a share is broadened to all households (a citizen dividend). The basic-needs basket costs 76% of the 2025 middle-quintile mean income, which reproduces the 58% of households above the ALICE basic-needs threshold (United For ALICE, 2024; the Fed's "doing okay" share is not used because it measures subjective well-being, not needs) as the 2025 affordability share, and is repriced each year at the physical price with an electricity component; affordability inside a quintile uses a lognormal spread. The block is a static allocation: the ownership and transfer levers change who receives the model's income, not the macro path, and the 2025 comparison base and the basket are the same in every policy scenario (levers off). "Q3" is the middle quintile's mean, not a computed median.

Three results. In physical goods, every quintile gains and the gains are nearly equal. The middle quintile's mean real disposable income is 2.66x in 2050 and the bottom quintile 2.45x (3.16x and 2.91x on the satiation-bounded lens), because the bottom quintiles earn physical wages, physical employment holds up through 2050 (§6.6), and the physical price falls; the top quintile share of disposable income moves from 50% to 52% and the Gini between quintiles from 0.418 to 0.442. Concentration is in the capital income, and it is inherited from 2025. Capital payouts are 24% of household income in 2050 and 83% of them go to the top quintile, which is the 2025 ownership distribution carried forward; the model has no mechanism by which ownership broadens on its own, and the case in which it narrows is the narrow_ownership preset. Affordability is flat, and the bottom quintile is not reached. The share of households able to afford the basket is 59% in 2025 and 68% in 2050 although the basket's price falls to 0.33 of 2025: the bottom quintile's nominal income falls with the wage bill and transfers replace only 5% of its income, so 1% of bottom-quintile households afford the basket in 2050 against 47% of the second quintile. Cheap cognition does not reach the household that cannot pay for housing.

The scenario table shows how purchasing power reaches households. Under regime_change (scarce intelligence, narrow ownership, weak re-employment, no pass-through) the middle quintile is at 1.75x and only 14% of households can afford the basket in 2050, because physical prices stop falling while wages do not stop falling; the aggregate index still rises. A citizen dividend on half of capital payouts (citizen_dividend) cuts the top quintile share to 45% and the Gini to 0.360 and lifts affordability to 77%; paying the same transfers as a flat amount rather than as wage insurance (universal_transfers) puts the bottom quintile at 3.55x and affordability at 74%, because the model's transfers are large and wage insurance pays them to the people who had cognitive wages to lose, who are not the bottom. The distribution of the renaissance is set by the ownership and transfer rules, not by the technology, which is the finding of §6.4 stated in households rather than in shares.

7.6 Sector prices and the essentials basket: a price sensitivity, not a sector block

The single physical price hides housing, land, regulation and climate. This section is a sector-price sensitivity analysis: five sector price indices are computed after the fact from the single physical price the fixed point solves, with the model's Cobb-Douglas aggregation and sector-specific cost shares (judgement values from the expenditure and input-output structure, not fits): housing, food, electricity, healthcare and manufactured goods, each with its share of human-and-robot labor (at a sector robot exposure), cognitive inputs at the cognitive price, scarce inputs that do not get cheaper with automation (land under housing, generation and grid under electricity) held at their 2025 price in wage units, electricity at the household rate, and the climate drift of §6.4 with a sector exposure. Sector prices, household budgets, sector demand, employment and production do not clear together, so the sector prices are not rescaled to the model's index, their weighted mean is reported beside it, and the gap is what the single index leaves out on this selection of shares. A fixed nominal land price is a price assumption, not a land-quantity constraint.

Two things follow. First, the renaissance is uneven across what people buy: by 2050 manufactured goods fall to 0.093 of their 2025 price and healthcare to 0.090 (its largest input is cognition), while housing falls only to 0.415 because a third of its cost is land, and electricity rises to 1.32 as grid capital is paid through rates (§4.7). The essentials basket, housing, food, electricity and healthcare, ends at 0.333, against 0.164 for the model's single physical index: 2.0x the price. Healthcare's fall is the most assumption-driven number in the table: it follows from a cognitive-input share of 0.40, and at a share of 0.20 (care rather than administration) the essentials basket and the bottom quintile read differently (the sensitivity table below). Second, on these shares the single index sits below the sector-weighted mean once scarce inputs are counted: the weighted mean of the sectors in 2050 is 0.270, 1.6x the model's index, so the physical-goods real-income lens the paper stands on (2.80x in 2050) reads 1.38x when deflated by the essentials basket instead. This is a statement about this selection of shares, not a general bound. The distribution block of §7.5 prices its necessities by the essentials basket rather than by the uniform physical price with an electricity weight (which would put the basket at 0.28 in 2050): the middle quintile's 2050 income in essentials is 1.31x and the bottom quintile's 1.21x, and the share of households able to afford the basket is the 68% reported above. The climate drift attaches by sector: on the central drift the essentials basket in 2050 is 0.390, with electricity the most exposed (+28% on its 2050 price):

The shares are judgement values, so the essentials reading is given across them (10 variants: scarce-input and cognitive-input shares halved and raised by half, the land share of housing, the cognitive share of healthcare, and no scarce inputs at all). The essentials basket in 2050 runs from 0.25 to 0.40, household purchasing power in essentials from 1.14x to 1.84x, the middle quintile from 1.08x to 1.75x and the bottom quintile from 1.00x to 1.61x, below its 2025 level in 1 of the variants; with no scarce inputs at all the basket is 0.25 and the bottom quintile 1.61x. Across the joint draws of Appendix D the essentials lens reads 1.01x to 1.83x (median 1.39x), the middle quintile 0.97x to 1.75x and the bottom quintile 0.87x to 1.70x, below its 2025 level in 15% of runs; the policy comparison of §7.5 carries the essentials columns:

8. What becomes possible

The disruption and the renaissance are the same event seen from two sides, and a paper that only counted the costs would misreport the model. This section states what the baseline implies becomes possible, each item tied to a quantity the model produces or an anchor in the evidence, and says plainly what the model cannot speak to.

The length of task a machine can complete. METR's horizon is the length, in human-expert working time, of tasks an agent completes at a stated success rate; it is not an autonomous operating duration, and every reading in this section that speaks of work running unattended is conditional on the autonomy discount of §4.3 ($\delta = 1$ here; at $\delta = 0.5$ each date below moves later by the time the horizon takes to double, about four months on the fitted trend), and METR marks measurements above sixteen hours as unreliable and the series as no longer actively updated. On the four-month doubling fitted to the record through March 2026, the 50% horizon is 3.5 hours in 2025, 27 hours in 2026, 172 hours in 2027 and a work-month (160 hours) in 2027; the 80% horizon, the model's proxy for work that runs without a person checking, reaches a work-month in 2028 and a work-year in 2030 (an extrapolation beyond the measured range; every figure past 2026 is the trend, not an observation). The model uses the horizon only through the substitutability sigmoid, which saturates at the accountability floor long before these lengths matter, so the labor results do not depend on the extrapolation holding past a few hundred hours. What the trend would mean if it held is the difference between an assistant and a colleague, and then between a colleague and a team: tasks that take a person months, delegated whole. Whether agents can run such work without supervision is not a quantity the horizon measures and the paper does not claim it; what the model can say is when the measured task length crosses each mark on the fitted trend, and that adoption lags capability by 5 years.

The price of thought. The unit cost of cognitive output falls to 3.5 × 10⁻⁷ of the 2025 human wage by 2035 and 2.7 × 10⁻¹⁰ by 2050; the price charged, which carries rents and scarcity while they last, is 4.0 × 10⁻³ and 1.2 × 10⁻⁵. Light fell 3,460-fold and became something nobody budgets for (§3). The model's cognitive volume rises 131-fold by 2050 with 99.8% of it done by machines: legal advice, diagnosis, tutoring, design, translation, analysis and planning become ambient, like light, for anyone with a connection. How much of that to count as income is an index-number choice, and the paper reports three lenses rather than one number: the chained all-goods index reads 9.7x in 2050 because it values the volume at its collapsing price; the satiation-bounded lens, which counts cognitive consumption at its 2025 value up to three times the 2025 volume and the rest of income in physical goods, reads 3.3x; the physical-goods lens, which ignores cognition, reads 2.8x and is the floor. The first is not how it will feel; the other two are the range the paper stands on (§6.4).

A second research workforce, and the first one gone. AI agents are 83% of research labor in 2027 and 99.9% in 2030, and algorithmic efficiency rises by orders of magnitude on their work; the human research and engineering workforce falls to the frontier handful (§6.6). The second workforce is not an addition to the first, it replaces it. The engine is specific to AI research in the model, but nothing in it is specific to AI: it applies wherever experiments are cheap relative to ideas, which is most of mathematics, software, materials simulation, drug design and the data-rich parts of biology. The CES structure is the warning: where experiments are the bottleneck (clinical trials, fusion, field ecology), the second workforce waits on the physical world. AlphaFold and AlphaEvolve are the template, and the model's own progress rate, one doubling every eight months, is the measured rate in the field that has it.

Everything with thought in it gets cheaper. A quarter of physical costs are cognitive inputs: design, logistics planning, legal and financial overhead, control software. The model's physical price index falls 66% by 2050 through that channel alone, before any robotics; the size of that fall is set by the input-share parameter. Energy is the exception that proves the mechanism: power is the one physical input whose price rises in the model, because AI itself is the new demand on it. The renaissance's visible frontier will be the physical trades that AI plans and humans execute: construction, installation, repair and care, which grow in every scenario (an allocation within physical employment, §7.1).

Medicine and care. The population over 65 grows by a third by 2050 with nine million direct-care openings this decade and no robot that fills them before the late 2030s. What cheap cognition does is move the diagnostic, administrative and coordination half of medicine toward zero marginal cost, which in the model is the healthcare-practitioner group holding its employment behind a 50% accountability floor while its tools transform. The licensed human remains the point of liability and the model treats that as institutional, not technical.

Education. Tutoring at zero marginal cost is the clearest early wonder and the evidence is already in: scaffolded AI tutoring doubles learning gains in randomised trials and harms it when unguarded. The model's implication for what to learn is not a list of safe majors; it is that trade-specific training with a long licensure or physical component holds its value, that the premium on routine cognitive credentials erodes, and that retraining capacity, not curriculum, is the constraint.

The arts, craft and the uses of time. The model does not value these and should not pretend to. What it provides is the budget: physical-goods real income at 2.8x its 2025 level by 2050 at hours that history says move little, and a basket in which the cognitive share rises only modestly even as its price collapses, which means people spend the gain on the physical and the positional: homes, places, care, each other, things made by hand. The lifestyle dossier's time-use record says a century of productivity bought four to five hours a week; the renaissance's leisure will arrive the same way, slowly, and through prices.

Law, governance and the accountability floor. The floor is the model's most political parameter and it is also where the upside is largest: when the cost of drafting, reviewing, and auditing falls by orders of magnitude, access to law and to competent administration stops being rationed by fees. The model only says the human licensee stays accountable; whether that becomes a guild barrier or a quality guarantee is a choice.

What the model cannot say. It has no account of discovery itself, no new goods, no creativity, no geopolitics, no catastrophe. It cannot distinguish a cure from a product. It measures the renaissance in the only units it has, prices, volumes and time, and finds them of the same order as the costs it also counts. The point of putting both in one model is that the conditions in the next section are the difference between them.

9. Conditions for a renaissance

The gains in section 8 are real in the model: cognition nearly free, real income up, output up, a labor share that holds, a second research workforce. They reach most people only under five conditions, each of which the model can switch off.

1. AI rents are competed away. Open weights within about a year of the frontier and no durable deployment concentration. The capital-owner outcome flips from best to worst on this alone.
2. The grid bill is not socialised onto households while the load is private, or is offset. Default sharing lifts rates 8% in 2030 and 26% by 2050.
3. Job finding stays fast and the wage discount is compensated. Two-thirds re-employed within one to three years is the current record; the stress variable is uncompensated loss, and compensation lets diffusion run faster.
4. The accountability floor is set deliberately. It is the modal political response in the model and it caps both the harm and the gain; it should not be set by lobbying in either direction.
5. Entrants get a route in. The 2023-2026 data are an entry-margin story; a decade of hiring freezes in exposed fields is the quiet way the transition fails, and it does not show up in the unemployment rate.

10. Limitations

One region, where the two real AI economies have opposite binding constraints. Climate is a switch and a drift with a sector exposure, not a damage function (§6.4, §7.6). The physical price is one index in the fixed point; the sector-price lens (§7.6) shows that on its judgement shares scarce land and generation would hold it about twice as high by 2050, so the physical-goods real-income lens reads high on that account. No monetary anchor: nominal GDP falls to 0.56 of its 2025 level by 2040 and 0.45 by 2050 in the wage numeraire, and a central bank would not let it. Exposure scores and floors by occupation are constructed. The peak displacement rate exceeds every national precedent; software needs no capex, but the number should be read as a stress test. Late-horizon ratios with vanishing denominators (return on a stock whose dollar value collapses with chip prices) are reported but not interpreted. The political reaction function is a set of priors. The stranded-skill measure needs the job-loss bound beside it. Everything is annual; the capex cycle in particular would look different at quarterly resolution, and the 2025-26 layoff record is stepped where the model is smooth. The robot wave rests on 2024-26 costs, shipments and a handful of deployments and on a substitutability path that is itself a forecast; robot capital spending and energy are in the capital block (§4.8), but robot manufacturing capacity enters only as a per-class fleet growth cap; and the wave is unfinished in 2050, so every 2050 labor figure is a point on a rising curve, not an endpoint. The early path is fitted to two years of data and then extrapolated: the separation share, the entry share, software's demand elasticity and the four-month horizon doubling are all 2025-26 estimates carried forward, and the years from 2028 inherit whatever is wrong with them. The slow-diffusion and soft-landing scenarios are the reading in which the 2025-26 pace does not persist.

The horizon is a benchmark, not a capability. METR's time horizon measures the human-expert duration of tasks completed at a success rate, on a suite that saturates at about sixteen hours and is no longer updated; it does not measure how long an agent operates without a person. Every horizon figure after 2026 in this paper is the fitted four-month doubling carried forward, labelled as such, and the model's labor results depend on it only up to the few hundred hours at which the substitutability sigmoid saturates. Statements about agents running research programmes or institutions on their own are not results of this model and the paper makes none; the capability the model needs is that agents complete, at the 80% level, tasks of the length the sigmoid's midpoint names, 1.5 reliable work-weeks.

Vendor revenue and the compute build. AI services outside the 2025 task stock are carried as a share of household consumption with a ceiling set by a prior (three per cent, about digital advertising plus consumer software today) and priced at a fixed premium over the task price, and they clear against the same powered pool as tasks (§4.3); the gross margin that floors the vendor price is a 2025 figure decaying to a long-run prior; and the block's 2025 level is the accounting bridge's estimate of what households paid for AI (§4.7), with its path to the ceiling a prior. On those assumptions services revenue of 1.5% of GDP in 2030 funds the build of the early 2030s (capex 14% of GDP at its 2035 peak), there is no second build by 2050, and the power capital is 16% of 2025 GDP in 2050 paid through rates; sized to total vendor revenue instead (services_total_revenue_sizing), the block puts power capital at 11% in 2050, and any late compute build competes with the robot fleet for the same saving. Whether consumers and advertisers will pay the ceiling, and whether the grid can be financed on rates, is not settled by the record; both are priors in Appendix D (§4.7).

11. Conclusion

Treat the AI transition as a price change and the usual questions reorder themselves. On the baseline's assumptions the price of cognition falls by orders of magnitude (the compute cost of a task by 8.0, the buyer's all-in cost by 114x); what stays scarce decides who gains; adoption runs 5 years behind capability, not a decade; the capex cycle has started and is about power and obsolescence; on the pace of 2025-26 the labor shock arrives in the early 2030s at a scale the political response cuts by 57% but does not prevent, and the trades it devalues are worth 7 times the compute it strands. Strip out re-employment, competition, pass-through and broad ownership and the aggregate index still rises while wages, the labor share and participation fall: the abundance is technological, the renaissance is institutional. On the other side of the same ledger: agents completing work-month tasks on the measured trend in 2027, the compute cost of a thought at 2.7 × 10⁻¹⁰ of the 2025 wage and the buyer's all-in cost at 2.2 × 10⁻³, a research workforce of AI agents that has replaced all but a handful of the people who built it, and everything with thinking in it getting cheaper. That is a renaissance by any historical standard, and the model finds it on the same scale as the costs. It is not, on the baseline, a renaissance in what people must buy: priced by housing, food, electricity and healthcare, household purchasing power in 2050 is 1.38x its 2025 level, the middle quintile is at 1.31x and the bottom quintile at 1.21x, because land, generation and care do not get cheaper with cognition and the wage bill that pays for them falls; that reading moves between 1.00x to 1.61x for the bottom quintile across the sector cost shares, which are judgement values (§7.6), and the policy levers of §7.5 move it more than the technology does. Its distribution is set by five institutional outcomes, not by how capable the systems become. The renaissance is here in cognition. Whether it reaches the essentials, and whether it is shared, is the decision in front of us.

Acknowledgments

This work was prepared with the assistance of various AI technologies.

Appendix A. Claims register

Appendix B. Reproduction

The model, the paper and the evidence are in one repository, and the model needs nothing beyond Python 3.10. python3 -m transition_econ run --all --out out writes every series and run_meta.json (model version, commit, date, scenarios and overrides); python3 -m transition_econ grid --out out the 81-run grid; python3 -m transition_econ extrapolate -s baseline --out out the occupation, education, household and human-capital tables; python3 -m transition_econ figures the figures in docs/paper/figures/; python3 -m transition_econ serve the interactive lab, which also renders the equation companion at /equations. The same commands run as transition-econ <command>; transition-econ list and transition-econ params list the named scenarios and every parameter with its default.

The tables of the text each have a command. transition-econ calibration prints the observed-versus-model table of §2 and the hindcast and continuity check of §5 (with --markdown, the task-accounting table of §6.1), and the test suite enforces the anchors; transition-econ transitions gives the critical transitions of §6.1, transition-econ politics the technology-against-politics decompositions of §6.4, transition-econ uncertainty --n 1000 --seed 1 the joint uncertainty analysis of §6.5, transition-econ regimes the recursive-improvement regimes of §6.7, transition-econ acemoglu the comparison of §6.8, transition-econ distribution the household quintiles of §7.5 (transition_econ/distribution.py), transition-econ sectors the sector prices of §7.6, and transition-econ validate the validation suite of Appendix E (transition_econ/validation.py), which exits with an error on any failed identity.

The paper source is docs/paper/DRAFT.md and the equation companion docs/paper/EQUATIONS.md; transition-econ paper renders both to HTML with the figures embedded and a manifest of the commit and parameter hash. The evidence dossiers are in docs/evidence/ (jobs, education, lifestyle, history, AI coding, layoffs, AI revenue, robotics and climate), with the mapping of anchors to sources and confidence in docs/evidence/SYNTHESIS.md. The accounting bridge of AI vendor revenue (§4.7) is transition_econ/data/ai_revenue.py, documented in docs/evidence/ai_revenue.md, and the 2019-25 record behind the hindcast is transition_econ/data/history.py.

Provenance. The model and the paper were audited by five independent external reviews in October 2026. Each finding was reproduced against the code before any change was made; the reports and the reproduction tables are in docs/audit/, and the log of each change is in the README. Every number in the text and tables is generated from the model (transition-econ paper --regen) and checked against it by transition-econ paper --check.

Appendix C. Equations

The complete equation set, with notation, is the equation companion published with this paper; its source is listed in Appendix B.

Appendix D. Priors for the joint uncertainty analysis

Appendix E. Economic validation suite

The validation suite is separate from the implementation tests: a unit test can prove that code reproduces an equation without establishing that the equation means what the paper says. The suite states each accounting and unit-consistency identity in the paper's notation, evaluates both sides on every simulated year of an actual run, and passes only within tolerance. It covers seven relationships: task stock against task flow, AI research labor against algorithmic progress, AI prices against compute costs, nominal GDP against purchasing power, conservation of workers across labor states, capital spending against computing capacity, and transfers against the resources that finance them. The worked example is 2040 on the baseline; 52 of 52 identities hold on the baseline and 59 of the 59 scenarios in this paper pass (one identity is not applicable in the supply-determined mode and is reported as such). Two of the identities shape how the model is stated: the satiation-bounded income lens of §6.4 counts cognition at full value only while cognition is the cheaper way to buy value, since otherwise it would fall below the physical-goods floor in the scarce-intelligence regimes; and the expenditure identity is stated net of the demand the physical sector cannot serve at potential, which the model records as excess demand rather than silently clearing.