AI-Driven Growth: The Variables

November 19, 2026
blog image

What actually decides growth or shrinkage

Built on a library of 105 primary documents spanning the NBER, IMF, OECD, BIS, Federal Reserve System, World Bank and the major AI labs and forecasters — ENSI Foresight Division.


Ask an economist whether AI will grow or shrink the economy and the honest answer is a shrug dressed up in a range: Goldman Sachs says a 7% uplift to global GDP over ten years — $7 trillion, plus 1.5 percentage points a year of US productivity growth. McKinsey Global Institute says $17.1–25.6 trillion, on top of automation gains already banked, worth 1.5–3.4 additional points of annual GDP growth in advanced economies. Daron Acemoglu, using the same underlying task-exposure data as everyone else, says no more than 0.66% cumulative TFP growth over the entire decade — and possibly as little as 0.53% once you account for the fact that early productivity studies sample the easy tasks. The OECD’s own general-equilibrium model lands in between, at 0.25–0.6 percentage points of TFP growth per year. These are not four opinions about the future. They are four different machines built from the same parts, and the fact that they disagree by more than an order of magnitude is itself the finding: the growth-or-shrinkage question cannot be answered with a single number, because it is not a single question.

It is eight questions, stacked. Will AI substitute for the humans doing a task, or make them more productive at the part only they can still do — and in what ratio, across millions of tasks? How fast does the substitution happen relative to how fast the offsetting productivity gains actually diffuse through the economy? Can the physical capital — chips, data centres, electricity — be built fast enough to matter, or does the buildout become its own economic drag? When a worker is displaced, how long before they are re-employed, and does that duration determine whether the shock is absorbed or compounds? When productivity rises, does the gain reach wages and consumer prices, or does it lodge as margin in a small number of balance sheets? Does the resulting shift in who holds purchasing power feed back into demand for the very output AI is producing more of? Does any of this even register in the statistics we use to declare victory or defeat? And does the answer to all seven prior questions land the same way in Tokyo as in Lagos, given that neither has the same demographic profile, chip access, or fiscal room to cushion the transition?

This report treats those eight questions as eight load-bearing variables and argues something more specific than “it’s complicated”: the variables interact multiplicatively, not additively. A fast task-substitution shock (variable 1) moving through the economy at high diffusion speed (variable 2), captured narrowly by a handful of compute-owning firms (variable 5), in a country with thin fiscal capacity to cushion the transition (variable 8), is not a moderately worse version of the same shock landing broadly, slowly, and in a well-cushioned economy. It is a qualitatively different outcome — arguably a shrinkage scenario rather than a growth one — even though the underlying technology, variable 1, is identical in both cases. Multiply a bad number into an otherwise good chain and the whole product turns bad; a technology that is unambiguously productivity-enhancing at the task level can still net out to falling aggregate demand, rising unemployment, and contracting realized GDP if the variables downstream of it point the wrong way.

That framing has a sharp and useful corollary, and it is the spine of this report: variables 1 through 4 are close to technological or physical givens — the elasticity of substitution in a given task, the diffusion lag of a general-purpose technology, the gigawatts a grid can add per year, the months it takes a displaced worker to find a new job are not things a finance ministry can vote to change on Tuesday. But variables 5, 6 and 8 — market structure, the demand feedback loop, and fiscal capacity — are substantially policy-shapeable. Antitrust enforcement, wage-and-price pass-through, unemployment insurance generosity, retraining infrastructure, robot taxation, and AI dividends are all instruments a state actually holds. Variable 7, measurement, is a special case: it doesn’t change the underlying reality, but it changes which reality gets seen, funded, and acted on. This is ENSI’s reframe of the entire growth-vs-shrinkage debate: it is less a technological forecasting problem than an institutional design problem wearing a technological costume. Report 3 in this series takes up the agentic, action-oriented question of what a state, firm, or worker actually does about that. This report’s job is narrower and prior to it: to take the eight dials apart, show what each one does, and show why the multiplication — not the individual settings — is what decides whether the outcome is growth or shrinkage.

The eight variables, one line each:

  • 1. Task-level substitution elasticity — for any task AI touches, whether it displaces the human doing it or augments them; aggregated across millions of tasks, this is the single most load-bearing number in the whole system.

  • 2. Diffusion speed vs. displacement speed — displacement can happen in a fiscal quarter; the offsetting productivity gains from reorganised firms and new task categories have historically taken decades to arrive.

  • 3. Capital deepening and the compute/energy bottleneck — AI substitution requires chips, data centres and electricity to actually be built, and that build-out has its own boom-bust economics that can throttle every other variable.

  • 4. Labor market reallocation frictions — whether displaced workers are re-employed in a quarter or a decade, which determines whether displacement is absorbed or becomes structural.

  • 5. Market structure and rent capture — whether productivity gains flow through to wages and prices, or get captured as margin by a small number of compute-owning firms.

  • 6. The aggregate demand feedback loop — the mechanism that converts variables 1–5 into either realized growth or realized contraction, because someone has to be able to buy the extra output.

  • 7. Measurement — whether any of this shows up in official statistics at all, and whether the ruler currently in use overstates the hype or understates the welfare gain.

  • 8. Second-order modifiers: demographics, geopolitics, fiscal capacity — the same underlying shock lands as labor-saving relief in an aging economy and as a surplus-labor crisis in a young one, and a well-resourced state can cushion what a thin one cannot.


1. Task-level substitution elasticity: automation vs. augmentation

The whole architecture of this report rests on a framework built by Daron Acemoglu and Pascual Restrepo, and it starts from a deceptively simple observation: a technology does not act on “labor” in the abstract, it acts on tasks. Production requires a range of tasks; each can, in principle, be performed by capital or by labor; and any new technology either expands the set of tasks capital can do (automation) or creates new tasks in which labor retains a comparative advantage (reinstatement). Their founding 2016 paper, “The Race Between Machine and Man” (NBER Working Paper 22252), formalizes this as a race between a displacement effect — automation always reduces the labor share and may reduce wages, full stop, because it is by definition capital substituting into tasks labor used to hold — and a productivity effect, in which the resulting efficiency gain raises demand for labor in the tasks that remain. New tasks generate something categorically different: a reinstatement effect, which is the mirror image of displacement and always raises both the labor share and labor demand.

Several things follow from taking this framework seriously, and each is a concrete evidentiary point:

  • The historical base rate for reinstatement is not small. Using data on when new job titles appear within occupations, Acemoglu and Restrepo find that about half of total US employment growth between 1980 and 2007 (8.84 of 17.5 percentage points) is explained by occupations with more new job titles growing faster — in 2000, roughly 70% of computer software developers, an occupation employing a million people, held job titles that hadn’t existed a generation earlier.

  • Their 2019 companion paper, “Automation and New Tasks” (IZA DP 12293, later published in the Journal of Economic Perspectives), decomposes the slowdown in the US wage bill over the last three decades and finds it is not explained by weak productivity growth alone: it reflects an acceleration in the displacement effect, concentrated in manufacturing, combined with a weakening of the reinstatement effect — automation has been running ahead of new-task creation, not in balance with it, for the first time in the postwar data.

  • The distributional mechanics are the sharpest empirical result in the whole literature. In “Tasks, Automation, and the Rise in US Wage Inequality” (Econometrica, 2022; NBER working-paper predecessor in this library), Acemoglu and Restrepo show that between 50% and 70% of the entire change in the US wage structure from 1980–2016 is explained by direct task displacement — the share of a worker group’s tasks lost to automation. Automation alone explains 80% of the rise in the college wage premium and cut the real wage of high-school-dropout men by 8.8% and dropout women by 2.3%. And here is the load-bearing sentence for this whole report: that scale of redistribution accompanied only a cumulative 3.4% increase in TFP between 1980 and 2016. Massive winners-and-losers churn is fully compatible with modest aggregate productivity growth — the two are not the same phenomenon and one does not certify the other.

  • Applying the same logic directly to generative AI, Acemoglu’s 2024 paper “The Simple Macroeconomics of AI” (NBER WP 32487) does the arithmetic explicitly. Roughly 20% of US labor tasks are exposed to AI; of those, about 23% can be profitably automated (the rest cost more to automate than they save); average task-level cost savings run about 27%, which nets to roughly 14.4% in overall cost terms once industry labor shares are applied. Run through Hulten’s theorem, that yields a TFP effect of no more than 0.66% over ten years — about 0.064% per year — falling to as low as 0.53% once you account for the fact that current productivity estimates are drawn disproportionately from easy-to-learn tasks, and the harder, more context-dependent tasks that remain will be slower and costlier to automate.

  • Acemoglu’s paper adds a second, sharper warning: AI’s task complementarity can raise inequality even when it is unambiguously helping low-skill workers, because raising a worker’s productivity at a task they still perform alongside a machine is a different economic event from creating a wholly new task in which they hold comparative advantage. Task complementarity, in other words, is not automatically reinstatement — and Acemoglu finds no evidence that AI will narrow labor-income inequality even though its micro-level exposure is more evenly spread across demographic groups than earlier waves of robotics. Instead, he predicts it widens the gap between capital and labor income specifically.

The reframe: this is not a debate that resolves once “the data comes in,” because the task-level elasticity is not one number — it is a distribution across millions of tasks, re-drawn every time a model gets better at one more of them. What the framework buys you is a diagnostic: any claim about AI’s growth effect that doesn’t specify whether it is describing the displacement channel or the reinstatement channel is not yet a claim about growth at all — it’s a claim about redistribution wearing growth’s clothes, and the 3.4%-TFP-for-70%-wage-restructuring result above is the proof that the two can come apart entirely.

2. Diffusion speed vs. displacement speed

If variable 1 sets the direction of the shock, variable 2 sets its timing — and the timing mismatch, more than the long-run endpoint, is what determines whether a transition is merely disruptive or actually contractionary. The historical evidence on general-purpose technologies is unambiguous on one point: displacement is fast and diffusion is slow, and the gap between them can run to decades.

  • Paul David’s 1990 paper “The Dynamo and the Computer” (American Economic Review) is the founding document of this argument. Electricity was commercially available from 1881. Yet in 1899, electric lighting reached just 3% of US residences, and installed electric motor capacity was under 5% of total factory mechanical drive. It took roughly two more decades to reach 50% diffusion, and factory electrification did not “reach full fruition” — in productivity terms — until the early 1920s: forty years after the first central generating station opened. The delay wasn’t ignorance; farsighted engineers understood the potential from the start. It was that profitable adoption required tearing down and rebuilding factories around a fundamentally different power-transmission architecture (the “unit drive” system replacing line-shaft-and-belt), and that rebuild only paid for itself once existing capital had physically depreciated. David’s statistic for the eventual payoff: roughly half of the five-percentage-point acceleration in US manufacturing TFP growth during the 1920s (versus the 1910s) is explained purely by growth in installed secondary electric motor capacity that decade — the gain was real, it was just twenty years late.

  • Nicholas Crafts’ growth-accounting study “Steam as a General Purpose Technology” (LSE Working Paper 75/03) finds an even longer lag for the prior GPT: James Watt’s improved steam engine was patented in 1769, and steam power did not reach parity with water as Britain’s dominant power source until 1830 — sixty years, and by some readings of the diffusion data, closer to a century before its full productivity contribution registered in the growth-accounting record.

  • The mechanism that explains why the lag exists — not just that it does — is Brynjolfsson, Rock and Syverson’s “Productivity J-Curve” (NBER WP 25148). GPTs require large complementary investments — new business processes, managerial know-how, retrained workers, redesigned organizational charts — that are intangible and go almost entirely unmeasured in the national accounts, even though they are genuine, costly capital formation. The model this produces is a J-curve: measured productivity growth is underestimated in the early years, while resources are diverted into building unmeasured intangible capital stock, and overestimated later, once that stock starts generating flows. Applied empirically to R&D, software and hardware investment, their adjusted TFP measure runs 11.3% above the official figure by the end of 2004 and 15.9% above it by the end of 2017 — meaning the “productivity paradox” of the 2000s and 2010s was not evidence that IT didn’t matter, but evidence that the accounting framework was systematically blind to the investment phase.

  • This is precisely why displacement and diffusion can move at such different speeds within the same firm, let alone the same economy. A company can license a model and reduce headcount in a single fiscal quarter — the displacement effect requires no organizational rebuild, just a subscription and a decision. The offsetting productivity effect — the reorganized workflow, the new task categories, the retrained managers who know what to do with an AI-augmented team — is exactly the kind of intangible, slow-diffusing complementary investment David and Brynjolfsson et al. describe, and historically it has taken between two and six decades to fully arrive.

The reframe: “AI hasn’t shown up in the productivity statistics yet” is not evidence against the technology — David’s and Crafts’ data says that’s the default pattern for every GPT in history. But it is precisely why the transition period is where the real policy stakes sit. A ten-to-forty-year gap between fast, cheap displacement and slow, capital-intensive reinstatement is not a rounding error to be waved away by “it evens out eventually” — it is a multi-decade window in which a whole generation of displaced workers lives through the costs before any of the benefits diffuse widely enough to reach them, which is exactly the terrain variables 4, 6 and 8 have to operate on.

3. Capital deepening and the compute/energy bottleneck

AI’s substitution for labor is not a pure software event — it is gated by how fast the physical capital stack (chips, data centres, electricity) can actually be built, and that build-out has its own economics, separate from the underlying technology’s potential, that can bottleneck every other variable in this report regardless of how favorable the task-substitution math looks.

The physical constraint, quantified by the IEA’s 2025 special report “Energy and AI,” is concrete rather than speculative:

  • Data centres accounted for 1.5% of world electricity consumption in 2024 — 415 terawatt-hours — with the United States holding 45% of that global total, China 25%, and Europe 15%. Consumption has grown roughly 12% a year since 2017, more than four times the growth rate of total electricity consumption.

  • The scale asymmetry within the category is stark: a typical AI-focused data centre draws as much power as 100,000 households; the largest facilities now under construction will draw twenty times that.

  • The IEA’s central projection has data-centre electricity demand more than doubling to around 945 TWh by 2030 — slightly more than Japan’s entire current electricity consumption — and rising toward roughly 1,200 TWh by 2035. In the United States specifically, data centres account for nearly half of all electricity demand growth through 2030, and by decade’s end the US is projected to consume more electricity for data centres than for aluminium, steel, cement, chemicals and every other energy-intensive industrial category combined.

  • The bottleneck is not generation capacity alone but the grid connecting it: the IEA estimates around 20% of planned data-centre projects are at risk of delay because new transmission lines take four to eight years to build in advanced economies and wait times for transformers and cables have doubled in three years. In advanced economies specifically, data centres already account for more than 20% of all electricity demand growth, against a backdrop of decades of essentially flat electricity demand before AI arrived — meaning the grid, in the countries where AI adoption is theoretically most advanced, is the least prepared for the load.

The financial architecture of the build-out compounds the physical constraint rather than easing it. The BIS’s July 2026 working paper “The AI Investment Race” (Rungcharoenkitkul, WP 1367) models the buildout explicitly as a winner-take-most contest: hyperscaler capital expenditure is set to exceed $700 billion in 2026 alone, and — measured against its own pre-boom trough — the current AI buildout is on track to outgrow every prior technology-investment mania in US history (the 1830s canal mania, the 1840s British railway mania, the 1920s, the dotcom boom) only three years in. The paper’s contest-theory model finds this is not accidental: because most of the reward in a winner-take-most race goes to a handful of contest winners rather than being shared proportionately, each coalition of hyperscaler-plus-lab has an individually rational incentive to over-invest relative to the socially efficient level — calibrated to actual balance-sheet and deal data, the model finds over-investment of roughly 1.5 times the efficient level, rising to as much as three times where demand is less elastic. The financing structure — heavy reliance on debt plus “circular financing,” in which a hyperscaler takes an equity stake in an AI lab in exchange for the lab’s commitment to purchase future compute — creates network fragility: a shock to one firm’s balance sheet can cascade through the web of equity stakes and purchase commitments to its partners. The paper’s most important number for this report: once cumulative build-out passes roughly $3 trillion, expected net economic surplus from the AI investment turns negative — and industry capex projections of $3–4 trillion over the coming years place the sector squarely in that downside-risk zone.

The reframe: the question “can AI substitute for labor at scale” implicitly assumes the compute to run that substitution simply appears. It doesn’t — it has to be financed, sited, connected to a grid that takes years to expand, and built by firms whose own contest dynamics push them toward exactly the kind of overbuild-then-retrench cycle that historically ends in a bust. That means the pace at which variables 1 and 2 can actually operate — how fast displacement, augmentation and reinstatement propagate through the real economy — is capped not by model capability but by transformer lead times and financing sustainability, and a bust in the AI capex cycle would compress the diffusion timeline of variable 2 involuntarily, and probably violently, regardless of what the underlying task economics say is possible.

4. Labor market reallocation frictions

Even where new tasks genuinely exist in aggregate — even where variable 1 nets out favorably at the level of the whole economy — workers do not teleport into the jobs that replace the ones they lost. How long that reallocation takes is itself a determinant of whether displacement is a manageable friction or a structural, demand-suppressing wound, and the best available real-world analogue for a slow-moving displacement shock is not automation at all — it’s trade.

  • Autor, Dorn and Hanson’s “The China Shock” (NBER WP 21906) remains the sharpest empirical demonstration that labor-market adjustment to a large exogenous shock is not fast, frictionless, or evenly absorbed, contrary to the pre-2000 economic consensus the paper explicitly overturns. In local labor markets most exposed to Chinese import competition, wages and labor-force participation remained depressed and unemployment remained elevated for at least a full decade after the shock began. At the national level, the offsetting employment gains the standard trade model predicts in unexposed industries had not materialized even years later. This is the load-bearing empirical fact for this whole variable: the textbook assumption that displaced workers “readily reallocate to other regions” turned out to be false for one of the largest, best-studied labor demand shocks in modern economic history.

  • The mechanism behind why reallocation is slow, rather than just the fact that it is, comes from Fallick, Haltiwanger, McEntarfer and Staiger’s “Job Displacement and Job Mobility” (NBER WP 29187), using matched US administrative data. The paper’s central finding reframes the entire displaced-worker literature: earnings losses are not driven by displacement itself, but by the duration of the jobless spell between jobs. Job changers who find new work within a quarter suffer minimal earnings losses; those who experience prolonged joblessness fall down the “job ladder” to lower-paying firms and suffer earnings losses that persist for years. The policy-relevant lever this identifies is precise: it is not displacement that needs to be prevented so much as the length of the gap that needs to be shortened — which is a reallocation-speed problem, squarely inside the domain of unemployment insurance design, active labor-market policy and retraining infrastructure (variable 8), not a technology problem.

  • A companion piece of evidence in this library, though catalogued for a different purpose, sharpens where reallocation friction actually concentrates. The Federal Reserve/Census “Dual U.S. Labor Market” analysis (Ahn, Hobijn and Şahin, NBER WP 31241) statistically identifies three labor-market segments: a primary tier — 55% of the population — that is almost always employed and rarely unemployed; a secondary tier — just 14% of the population — that absorbs most short-run labor-market fluctuations, with six times the turnover rate of the primary tier and ten times the likelihood of unemployment; and a loosely attached tertiary tier. The secondary tier, despite comprising under 12% of total employment, accounts for 61% of all US unemployment. This matters directly for AI displacement risk: reallocation friction is not a diffuse tax spread evenly across the workforce — it concentrates overwhelmingly in a specific, already-precarious labor-market segment, meaning the same aggregate displacement shock lands as a minor inconvenience for most workers and a structural crisis for a much smaller, identifiable group.

  • The scale of exposure this friction has to absorb is itself sizeable and unevenly distributed. The ILO’s 2025 refined global exposure index (”Generative AI and Jobs,” Working Paper 140) finds that one in four workers globally hold an occupation with some generative-AI exposure, with 3.3% of global employment in the highest exposure category — split unevenly by gender (4.7% of female employment versus 2.4% of male employment in the top category) and sharply by income level (11% of employment exposed in low-income countries versus 34% in high-income countries). Because most occupations combine automatable and non-automatable tasks, the ILO’s own framing is that transformation, not wholesale replacement, is the modal outcome — which is precisely why reallocation speed, rather than net job counts, is the variable that determines the human cost.

The reframe: “will there be enough jobs” is close to the wrong question — McKinsey’s own workforce-transition modeling concludes there is likely to be enough work in aggregate through 2030 under most scenarios, while adding, in the same breath, that the transition will be “very challenging, matching or exceeding the scale of shifts out of agriculture and manufacturing” the economy has previously absorbed. The right question is how long the gap lasts for the specific, structurally exposed segment of the workforce that reallocation friction concentrates on — and that duration is what converts an aggregate productivity gain into a persistent regional and demographic wound, which is the mechanism variable 6 needs to translate into demand effects.

5. Market structure and rent capture

Identical productivity gains produce entirely different macro outcomes depending on who captures them. If AI-driven cost savings flow through to broad-based wage growth and lower consumer prices, the shock is demand-sustaining even as it displaces specific workers. If those same savings are captured as margin by a small number of compute-owning firms, the economy can be more productive on paper while purchasing power stagnates — and this is not a hypothetical risk grafted onto the AI story from outside; it is the continuation of a market-structure trend that predates AI by four decades and that AI’s own supply chain (chips, cloud, foundation models) is unusually well-positioned to intensify.

  • The foundational evidence is Jan De Loecker and Jan Eeckhout’s “The Rise of Market Power and the Macroeconomic Implications” (NBER WP 23687), which measures firm-level markups across the US economy from 1950 using Compustat data covering roughly a third of US employment and 41% of sales. Markups were essentially flat, even slightly declining, from 1950 to 1980. Then, starting in 1980, average US markups rose from 18% above marginal cost to 67% above marginal cost by 2014 — an increase by a factor of 3.6 over 35 years, with the rise concentrated almost entirely in the top of the distribution: the 90th-percentile markup rose from 1.4 to 2.6, while markups in the lower percentiles stayed flat or fell. The paper shows this single trend is statistically sufficient to account for seven separate secular patterns simultaneously: the falling labor share, the rising capital share, falling low-skill wages, falling labor-force participation, falling job-to-job mobility, falling interstate migration rates, and the broader slowdown in aggregate output growth.

  • The complementary mechanism — why markups concentrate the way they do — is Autor, Dorn, Katz, Patterson and Van Reenen’s “superstar firms” model (”The Fall of the Labor Share and the Rise of Superstar Firms,” NBER WP 23396). Using US Census micro-data across manufacturing, retail, wholesale, services, utilities/transportation and finance, they show the labor-share decline is driven overwhelmingly by reallocation of sales toward a small number of highly efficient, low-labor-share firms within each industry — not by a broad-based decline inside incumbent firms — and that the pattern holds not just in the US but across other OECD economies using comparable firm-level data. This is the structural template AI market concentration would plug directly into: a technology whose core inputs (frontier compute, proprietary training data, model scale) exhibit exactly the winner-take-most economics that produce superstar-firm dynamics.

  • That this is not merely an academic pattern-match is precisely why competition regulators are treating it as a live, present-tense concern rather than a future hypothetical. The UK Competition and Markets Authority’s “AI Foundation Models Technical Update Report” assesses concentration risk directly in compute access, training data and model distribution. The US FTC’s 6(b) study of cloud-AI partnerships investigates exactly the kind of tie-up — cloud providers taking financial and infrastructural positions inside the AI labs they host — that BIS’s circular-financing analysis (variable 3) independently flags as a source of both concentration and systemic financial fragility: the same handful of firms and the same web of equity stakes and compute-purchase commitments show up in a competition-policy report and a financial-stability report, from two different institutions, investigating two different risks.

  • The IMF’s own cross-country analysis of rising corporate market power (Staff Discussion Note SDN/21/01) links the De Loecker-Eeckhout markup trend directly to weaker aggregate investment and wage growth internationally, not just in the US — meaning this is not an American peculiarity that AI happens to be built inside of, but a general-equilibrium pattern that AI’s specific supply-chain economics (a handful of foundation-model labs, an even smaller number of frontier chip and cloud providers) is structurally suited to reproduce and potentially accelerate.